Gradient analysis of landscape variation in Norway

A multitude of landscape characterisation and mapping methods exist, but few methods take into account that landscapes properties vary in a gradual, continuous manner along multiple directions of variation. In this study, we used gradient analytic methods, rooted in ecological continuum theory, to analyse landscape variation throughout Norway. The aim is to explain differences in landscape properties in the simplest possible way, by identifying ‘complex landscape gradients’ (CLGs), i.e. composite gradients of co-occurring landscape elements and properties. We collected data by stratified sampling of 100 test areas (20×20 km), in which we delineated a total of 3966 observation units (landscape polygons 4–30 km²) based on geomorphological criteria. For each observation unit, 85 landscape variables were recorded. We identified patterns of variation in landscape element composition by parallel use of two multivariate statistical methods, detrended correspondence analysis (DCA) and global nonmetric multidimensional scaling (GNMDS). The analyses revealed that the most important properties explaining differences in total landscape elements composition was location of the landscape relative to the coastline and coarse-scale landform variation. Most landscape elements had distinct optima within specific segments along broad-scale complex-gradients in landscape properties. A tentative landscape-type hierarchy was built by an iterative procedure by which the amount of compositional turnover in landscape-element composition between adjacent types was standardised. Six ‘major landscape types’ were identified based on geomorphological criteria. Within each major type, we identified a unique set of 2–5 important CLGs, representing geo-ecological, bio-ecological, and land use-related landscape variation. Minor landscape types were obtained by combining segments along two or more CLGs. The study shows that geological diversity, biological diversity and human land-use are tightly intertwined at the landscape level of ecological complexity, and that predominantly abiotic processes control and constrain both biotic processes and human land use.


Gradient analysis of landscape variation in Norway
Trond Simensen, a) Geo-ecological research group, Department of Research and Collections,6 Natural History Museum, University of Oslo, P.O. Box 1172, Blindern, NO-0318 Oslo, 7 Norway, b) Norwegian Environment Agency,P.O. Box 5672 Torgarden,8 Norway 9 10 Rune Halvorsen, Geo-ecological research group, Department of Research and Collections,11 Natural History Museum, University of Oslo, P.O. Box 1172, Blindern, NO-0318 Oslo, 12 Norway. 13 14 Lars Erikstad, a) Geo-ecological research group, Department of Research and Collections,15 Natural History Museum, University of Oslo, P.O. Box 1172, Blindern, NO-0318 Oslo, 16 Norway, c) Norwegian Institute for Nature Research (NINA), Gaustadalléen 21, Oslo, Norway  18  19  20  21  SUMMARY  22  23 A multitude of landscape characterisation and mapping methods exist, but few methods take 24 into account that landscapes properties vary in a gradual, continuous manner along multiple 25 directions of variation. In this study, we used gradient analytic methods, rooted in ecological 26 continuum theory, to analyse landscape variation throughout Norway. The aim is to explain 27 differences in landscape properties in the simplest possible way, by identifying 'complex 28 landscape gradients' (CLGs), i.e. composite gradients of co-occurring landscape elements and 29 properties. 30 We collected data by stratified sampling of 100 test areas (20×20 km), in which we 31 delineated a total of 3966 observation units (landscape polygons 4-30 km²) based on 32 geomorphological criteria. For each observation unit, 85 landscape variables were recorded. 33 We identified patterns of variation in landscape element composition by parallel use of two 34 multivariate statistical methods, detrended correspondence analysis (DCA) and global 35 nonmetric multidimensional scaling (GNMDS). 36 The analyses revealed that the most important properties explaining differences in 37 total landscape elements composition was location of the landscape relative to the coastline 38 and coarse-scale landform variation. Most landscape elements had distinct optima within 39 specific segments along broad-scale complex-gradients in landscape properties. A tentative 40 landscape-type hierarchy was built by an iterative procedure by which the amount of 41 compositional turnover in landscape-element composition between adjacent types was 42 standardised. Six 'major landscape types' were identified based on geomorphological criteria. 43 Within each major type, we identified a unique set of 2-5 important CLGs, representing geo-44 ecological, bio-ecological, and land use-related landscape variation. Minor landscape types 45 were obtained by combining segments along two or more CLGs. 46 The study shows that geological diversity, biological diversity and human land-use are 47 tightly intertwined at the landscape level of ecological complexity, and that predominantly 48 abiotic processes control and constrain both biotic processes and human land use. 49 1 2 Keywords: Gradient analysis; Landscape variation; Landscape ecology, Landscape analysis, 3 Landscape characterisation, Ordination, Spatial ecology 4 5 6 Abbreviations: A = land-use variable; AIV = area-independent variable; AA = area, measured 7 as proportion of observation unit; Abygg_a = large buildings; AlpA = proportion of area 8 above the forest line; Araaf_a = open areas; Arbar_a = coniferous forest; Arbla_a = mixed 9 boreal forest; Arfull_a = cultivated land; Arlov_a = deciduous forest; Arover_a = surface 10 cultivated land; Asp_n_a = north facing terrain; Asp_s_a = south facing terrain; BA = amount 11 of boreal/alpine landscape; BE = bio-ecological gradient; Bavstn_a = sedimentary rock; BGF 12 = basic geomorphological forms; Bohei_a = boreal heaths; Bomd_a = metamorphic rock; BP 13 = glacier; Bplu_a = plutonic rock; Bpoor_a = lime-poor bedrock geology; BrI = mean BrI in 14 the polygon; Brich_a = lime-rich bedrock; Build_a = built-up area; Bvul_a = volcanic rock; 15 Ccom_m = coastal complexity; C_ek_a = simple coastline; C_kk_a = complex coastline; 16 City_a = town/city area; CLG = complex landscape gradient; COU = count variables; Cr3_r_a 17 = smooth/flat coast; Cr3_u_a = rugged coast; Cre_b_a = steep coast; Cre_f_a = flat coast; 18 Crug3_m = coastal ruggedness, VRM3, mean; Crug9_m = coastal ruggedness, coarse scale, 19 VRM9, mean; Cul_a_s = number of archaeological heritage sites; Cul_b_s = number of 20 ancient rock art sites; Cul_k_s = number of church ruins; Cul_m_s = number of marine 21 cultural heritage sites; Cul_t_s = number of technical heritage sites; Cul_u_s = number of 22 cultural heritage sites outdoors; DCA = detrended correspondence analysis; Discoast = mean 23 distance to coast; Dislake = mean distance to lake; Dismire = mean distance to mire; DN = 24 valley form; DNI = relationship between valley depth and valley width; EDU = ecodiversity 25 distance units; EDU-L = ecodiversity distance units in landscapes; Ekspbe_a = slightly 26 protected coast; Ekspmo_a = slightly exposed coast; Ekspve_a = exposed coast; Er_m = 27 hypsographic index, ER, mean; FlA = amount of flat land; Flat_a = flat terrain; G = basic geo-28 ecological variables; GE = geo-ecological gradient; Gab_a = built-up; Gab_fi = fisheries-29 related buildings; Gab_nae = commercial buildings; Glac_a = glacier; GNMDS = global 30 nonmetric multidimensional scaling; Guro_t_a = rugged terrain; H.C. units = half-change 31 units; I = inland landscapes; IA = inland hills and mountains; INA = internal association; ID 32 = inland valleys; IF = fine sediment plains; IfI = mean infrastructure index in the polygon; not 33 used inter alia because of error in terrestrial area correction; IfIu = same as IfI, but calculated 34 by the formula for IfIutv in Erikstad et al. (2014); II = mean lake index in the polygon; 35 Innoy_s = freshwater lake islands; Inns_s = number of lakes; IP = freshwater lake properties; 36 IX = inland plains without dominance of fine sediments; JI = mean agricultural index in the 37 polygon; JK1 = primary soil/ sediment class; JK1-B = exposed bedrock (> 50% of area 38 covered by exposed bedrock); JK1-E = glaciofluvial deposits; JK1-H = marine deposits; 39 JK1-M = thick layer of till; JK1-0 = no soil/sediment class (> 50% not assigned to specific 40 sediment/soil type); JK2 = secondary soil/ sediment class; JK2-B = exposed bedrock (> 25% 41 of area covered by exposed bedrock); JK2-E = glaciofluvial deposits; JK2-EH = marine 42 deposits; JK2-EM = thick layer of till; JK2-E0 = no soil/sediment class (> 25% not assigned 43 to specific sediment/soil type); JP = amount of agriculture; K = coastal landscapes; Kbf_a = 44 exposed bedrock; Kelv_a = glaciofluvial deposits; KF = fjords and other coastal landscape; 45 KL = coast line; Klac_a = lacustrine deposits; Kmar_a = marine deposits; KS = coastal plains; 46 Kskred_a = landslide soil; Ktkmo_a = thick layer of till; KV = groups of characterising 47 variables; collective term for primary key and analysis variables; Lake_a = freshwater lake; 48 Land_a = terrestrial area; LEP = landscape property profiles; LGL = total landscape gradient 49 length; LG-variable = landscape gradient variable; LEC = local complex environmental 1 variable; Lled_a = power lines; LNr = serial number; LT = landscape type; M = marine 2 landscapes; Maro_s = marine islands; Meant_a = moderate slope; MI = mean mire index in 3 the polygon; Mire_a = mire; MP = abundance of mire; NiN = Nature in Norway; OI = amount 4 of infrastructure; ONV = orthogonal key variable; OU = Observation unit; Oyst_i = inverse 5 island size; PCA = principal component analysis; PD = proportional dissimilarity; PKV-group 6 = parsimonious KV-group, see KV; PLC = primary landscape major-type candidate; PNV = 7 primary key variable; pONV's = the primary orthogonal key variable (see ONV) values 8 ranged on a 0-1 scale; PRA = proportion of area (pixel number variable); R_net_a = river; 9 Rd_anl_a = reindeer husbandry facilities; RDA = redundancy analysis; RE = relief; River_a = 10 large river; rONV = rescaled ONV (see ONV); RR1 = mean RR1 in the polygon given in 11 meters; RR1_m = altitudinal range; rrONV = ranged, rescaled ONV (see ONV); Rug3_m = 12 terrain ruggedness VRM3, mean; SD = standard deviation;  1  2  3  4  THE LANDSCAPE LEVEL OF ECOLOGICAL DIVERSITY  5  6  7 Since the 1980s, ecologists have become increasingly aware that the relative importance of 8 factors controlling biotic and abiotic patterns and processes vary with the scale of observation, 9 often being non-linearly related to the gradient from finer towards broader spatial and 10 temporal scales (Swanson 1988, Wiens 1989, McGill 2010b, Zarnetske et al. 2019 Accordingly, any ecological phenomenon (e.g. the diversity of species, ecosystems and 12 landscapes), should be addressed at the spatial and temporal scale at which the phenomenon 13 appears (Estes et al. 2018). Hierarchy theory (Allen & Starr 1982, O'Neil et al. 1986, Urban et 14 al. 1987, King 2005 suggests that ecological systems are implicitly hierarchical, in the sense 15 that the interacting components are organised in 'levels of organisation' within a 16 hierarchically organised system. A well-known example of a biotic hierarchy is the series cell 17 organismpopulationcommunity (Turner & Gardner 2015). Complex entities at any 18 particular level in such a hierarchy can be explained and studied in terms of entities only one 19 level down in the hierarchy; entities which themselves are likely to be complex enough to 20 need further reduction to their own component parts (King 2005, Allen & Starr 2017). 21 Dawkins (1986) termed the study of such entities 'hierarchical reductionism', and pointed out 22 that the kinds of explanations suitable at high levels in the hierarchy may be qualitatively 23 different from the kinds of explanations suitable at lower levels. 24 Halvorsen et al. (2020) defined 'ecodiversity' as 'the diversity of units defined by 25 biotic as well as abiotic components and their interactions, and the processes that give rise to 26 variation in the structure and composition of these components'. 'Landscapes' are often 27 recognised as a separate level of ecological diversity, simultaneously addressing biotic and 28 abiotic variation in heterogeneous areas of kilometres-wide extent (Noss 1990, Allen & 29 Hoekstra 1990, Bailey 2009, Halvorsen et al. 2020. The domain of spatial scales typically 30 applied in landscape characterisation and mapping addresses Nature's diversity at a scale less 31 detailed than that of, e.g. an ecosystem, and more detailed than that of ecoregions (e.g. Bailey 32 2014). In this context, landscapes contain biotic and abiotic subsystems such as landforms, 33 ecosystems, meta-ecosystem complexes and other natural and human-induced landscape 34 elements at spatial scales from 10 6 to 10 10 m² (often referred to as meso-scale) responding to 35 abiotic and biotic processes occurring over timespans from 10 1 to 10 4 years (Delcourt et al. 36 1982, Dikau 1989. 37 We define 'landscape' as a more or less uniform area characterised by its content of 38 observable, natural and human-induced landscape elements, i.e. natural or human-induced 39 objects or characteristics, including spatial units assigned to types at an ecodiversity level 40 lower than the landscape level, which can be identified and observed on a spatial scale 41 relevant for the landscape level of ecodiversity (Halvorsen et al. 2020). Furthermore, we 42 define 'landscape element' as a natural or human-induced object or characteristic, including 43 spatial units assigned to types at an ecodiversity level lower than the landscape level, which 44 can be identified and observed on a spatial scale relevant for the landscape level of 45 ecodiversity (Halvorsen et al. 2020). 'Landscape types' are defined as more or less uniform 46 areas characterised by their content of observable, natural and human-induced landscape 47 elements (Halvorsen et al. 2020). 48 Knowledge about variation at the landscape level of ecodiversity is a prerequisite for 49 knowledge-based spatial planning and management (Marsh 2005). Such knowledge is also 50 considered as essential to fulfil obligations set by international conventions such as the 1 European Landscape Convention (Anonymous 2000) and legal frameworks such as the 2 Norwegian Nature Diversity Act (Anonymous 2009). The primary purpose of the latter is, e.g. 3 to protect 'biological, geological and landscape diversity' and promote conservation and 4 sustainable use of the 'full range of variation of habitats and landscape types'. While the 5 concept of landscape diversity presupposes knowledge about the relative frequency or 6 abundance of discrete entities (landscape types), defining such entities is challenging (Skånes 7 1997, Simensen et al. 2018. By and large, the composition, structure and functions of 8 landscapes vary in a gradual, continuous manner along multiple 'directions of gradual 9 variation ' (Halvorsen et al. 2016). For such objects, that are not naturally classified, all type 10 systems are artificial in the sense that they have to be constructed by some set of rules and 11 that many such rule sets may, in principle, be applied (Økland & Bendiksen 1985). 12 Accordingly, a multitude of different methods exist for the characterisation and mapping of 13 ecological diversity at the landscape level (Simensen et al. 2018 Einevoll 1965). Although the concept of 'landscape types' has been used informally in 24 geographical texts in Norway for a long time (Sømme 1938, Holt-Jensen 2009 comprehensive and evidence-based type systems for the landscape level of ecological 26 diversity, which simultaneously address biotic and abiotic variation, have still not been 27 established. Regional and national landscape characterisation efforts have mostly followed the 28 regional geographic tradition by which the individual character of singular areas such as 29 landscape units or regions is described rather than addressing geographical phenomena in 30 general (i.e. by a type-system; see e.g. Nordisk ministerråd 1984, Nordisk Ministerråd 1987 Puschmann 2005). 32 Moreover, the few examples of landscape characterisation methods applied at broader 33 scales in Norway have relied strongly on an intuitive, qualitative, expert-based and holistic 34 assessment of the total composition of landscape elements and properties (see, e.g. Nordisk 35 ministerråd 1987ministerråd , Puschmann 2005. While qualitative descriptions of spatially delineated 36 landscape units are intuitively appealing for communication purposes, the lack of 37 transparency, repeatability and scientific rigour behind expert-based methods will often 38 constrain their usefulness for a broader range of applied and scientific purposes (Simensen et 39 al. 2018). Expert-based methods are prone to a multitude of biases and uncertainties, such as 40 confirmation bias (seeking information that affirms pre-existing beliefs), coverage bias (the 41 extent to which different issues and are reported and considered) and concision bias 42 (selectively focusing on information, losing nuance), possibly leading to inconsistencies and 43 lack of correspondence to observable reality (see e.g. Gelman & Henning 2017). Quantitative 44 analysis, on the other hand, has the great merit of revealing relationships which are inherent in 45 the study material, but otherwise unrecognisable (Whittaker 1966). A high degree of observer 46 independence is a mandatory prerequisite for addressing several specific research questions 47 within landscape ecology and physical geography, such as the study of spatial distribution and 48 abundance of landscape elements, and studies of patterns, structure and processes in the 49 landscape. Alahuhta et al. (2019) and Schrodt et al. (2019) thus maintained that progress in 50 making connections between non-living and living nature requires systematic data collection 1 designed to improve our knowledge of the linkages between biodiversity and geodiversity at 2 several scales and levels. Several Norwegian scientists (e.g. Moen 1999, Strand 2011, Krøgli 3 et al. 2015, and Erikstad et al. 2015 have called for more systematic, observer-independent 4 and repeatable frameworks for the study of the total composition of landscape elements 5 throughout Norway. 6 7 8 9 METHODS FOR LANDSCAPE CHARACTERISATION AND MAPPING 10 11 12 Internationally, the recent trend has been towards increasing observer-independence in bio-13 physical landscape characterisation, due to improved availability of area-covering landscape 14 data derived by remote sensing and powerful novel statistical and geographical analysis 15 methods (Simensen et al. 2018, Zarnetske et al. 2019, Yang 2020 composition of landscape elements and properties are most often conducted by stepwise, 20 criteria-based GIS overlay techniques, multispectral segmentation, or GIS analysis in 21 combination with multivariate statistical analysis (Simensen et al. 2018, Yang et al. 2020. 22 The latter groups of methods typically include recording of a broad selection of physical 23 landscape attributes within fixed spatial units (e.g. 1×1 km raster cells) followed by 24 multivariate statistical analyses and supervised or unsupervised clustering techniques, to 25 classify or group landscapes with related characteristics (Bailey 2009). A map is produced by 26 drawing lines around cells classified to the same or related classes. Commonly applied 27 clustering methods are Two Way Indicator Species Analysis (TWINSPAN; Hill 1979) and K-28 means clustering (MacQueen 1967). In Norway, Krøgli et al. (2015) analysed landscape data 29 in geographical grids of 5×5 km and 1×1 km squares, respectively, and identified ten 30 landscape categories with a high degree of internal similarity. 31 Although classification by numerical clustering procedures have been successful in 32 identifying 'groups of similar landscapes', clustering methods are in general not well suited 33 for the analysis of observation units that vary more or less continuously along gradients, as 34 most often is the case with landscape properties. Much information is lost when 35 multidimensional networks are assigned to clusters within a unidimensional hierarchy because 36 the relationships between the classes at the same level in the hierarchy are hidden 37 (Tuomikoski 1942, Økland & Bendiksen 1985. Since clustering methods do not allow for 38 close examination of the abundance and distribution of landscape elements along multiple 39 gradients and dimensions in the landscape space (cf. Whittaker 1967), the results of cluster 40 analyses of landscape variation are challenging to interpret ecologically (cf. Austin 2002). An 41 additional disadvantage of clustering methods is that the number of clusters, and hence 42 landscape types, must be defined in advance rather than follow from the analysis of data. We 43 argue that alternative approaches are needed to obtain a better understanding of the 44 relationships between geodiversity and biodiversity at the landscape level (cf. Alahuhta et al. An understanding of natural variation based upon knowledge about environmental gradients 3 and species' responses to these gradientsa gradient perspective on species-environment 4 relationships (Halvorsen 2012)is supported by evidence from ecosystems all over the world 5 and has served as theoretical foundation for reserach in plant and community ecology for 6 more than 50 years (Gleason 1939, Curtis 1959, Whittaker 1967, Austin 2005. Application of 7 a wide range of methods of gradient analysis (including multivariate gradient analysis, i.e. 8 ordination) has been crucial for disentangling the underlying set of rules that regulate the 9 composition, structure and functions of ecosystems (Økland 1990(Økland , McGill 2010b including 10 microbiological systems (van Elsas et al. 2006). At the landscape level, however, the 11 continuum concept in general, and ordination methods in particular, are rarely applied to 12 understand and describe patterns of variation in landscape element composition. At this level, 13 gradient analysis is usually limited to the use of single continuous variables as an alternative 14 to 'landscape metrics' based on categorical landscape data (Cushman et al. 2010, Lausch et al. 15 2015. Continuous variables at the landscape level can in principle be obtained by two 16 approaches (Lausch 2015): (i) by direct measurements on a continuous scale (e.g. a digital 17 elevation model); or (ii) by deriving continuous variables from categorical land cover data by 18 a moving window approach (McGarigal & Cushman 2005). 19 A key point in the gradient perspective on species-environment relationships is the 20 concept of the environmental complex-gradient (Whittaker 1956), i.e. that a set of correlated 21 environmental variables act on the species in concert rather than one by one. Numerous 22 studies of species-environment all over the world have underpinned the notion that a few 23 major environmental complex-gradients typically account for a significant fraction of the total 24 variation in species composition that can be explained by variation in the environment 25 (Halvorsen 2013). Moreover, species generally are found within a restricted interval along 26 each major environmental complex-gradient. We hypothesise that the complex-gradient 27 concept, commonly used to understand and describe species' relationships to the environment, 28 can be extended to the landscape level of ecological diversity. Furthermore, we hypothesise 29 that ecosystem and other landscape elements, just like species, may have distinct optima 30 (maximum probability of occurrence) within specific segments of broad-scale, complex 31 gradients in landscape properties. If this hypothesis is correct, evidence-based and testable 32 type systems can be obtained by combining segments along two or more CLGs based on the 33 amount of compositional turnover (Halvorsen et al. 2020). While the principle of turning a 34 multidimensional space into types by combining gradient intervals has a long history in 35 vegetation ecology (e.g. Tuomikoski 1942, Økland & Eilertsen 1993, similar approaches to 36 the establishment of landscape typologies are uncommon. 37 In 2016, Halvorsen et al. launched the conceptual framework of 'Nature in Norway' 38 (NiN) for a description of landscape variation throughout Norway. NiN was later expanded 39 and developed to a universal system for systematisation of ecological diversity at several 40 scales and levels (EcoSyst; Halvorsen et al. 2020). EcoSyst consists of a set of general 41 principles and methods for systematisation of Nature's diversity that simultaneously addresses 42 biotic and abiotic variation across different levels of organisation. The EcoSyst framework 43 aims to provide the basis for an evidence-based systematics for all observable aspects of 44 ecodiversity, at spatial scales from microhabitats to landscapes. 45 A first, pilot, version of a gradient-based landscape type-system was developed within 46 the NiN framework for Nordland county (Erikstad et al. 2015), where 173 landscape variables 47 recorded in 258 observation units were subjected to multivariate analyses to identify 48 'landscape gradients'. The study showed that ordination methods allow flexible analyses of 49 landscape variables derived from a wide variety of data sources: (i) count data (e.g. 50 buildings); (ii) areal coverage of various land-cover types; and (iii) continuous variables 1 derived from direct measurements (e.g. remote sensing) or by neighbourhood calculations 2 (e.g. morphometric variables). The study also demonstrated that composite landscape 3 gradients could be obtained from multivariate analyses of landscape element compositional 4 data undertaken to reduce the dimensionality of an n-dimensional landscape-level hyperspace. 5 Nevertheless, the pilot study also revealed a clear need for a standardised procedure to 6 identify parsimonious sets of complex landscape gradients, as well as the need for a 7 standardised method to divide such gradients into segments by transparent and repeatable 8 criteria. 9 In this study, we expand on the experience gained from the Nordland pilot project and 10 analyse landscape variation by use of a sample of observation units from the entire Norwegian 11 mainland, including coastal areas. We use, for the first time, a novel, gradient-based iterative 12 procedure which involves calculation of 'ecodiversity distance', i.e., the extent to which the 13 landscape element composition differs between adjacent candidate types. The current study 14 was conducted as a part of the methodological development of the NiN framework (building 15 on Halvorsen et al. 2016), and has been part of the empirical basis for NiN and,subsequently,16 the theoretical framework of EcoSyst (Halvorsen et al. 2020 The aims of this study are fivefold: 23 (1) To identify major trends and patterns in landscape variation throughout Norway by 24 applying gradient analysis methods to landscape element compositional data. 25 (2) To study linkages between biodiversity and geodiversity at the landscape level of 26 ecological diversity by analysing the relative importance of co-occurring landscape elements 27 and properties within various functional categories, i.e. geo-ecological, bio-ecological and 28 land-use related landscape variation. 29 (3) To examine if, and in case to what extent, landscape elements have distinct optima 30 along broad-scale, complex gradients in landscape property composition, that is, intervals in 31 which they reach maximum occurrence probability. 32 (4) To explain variation in landscape element composition and properties in the 33 simplest possible way, by identifying a parsimonious set of gradients of co-ordinated 34 variation. 35 (5) To construct the first version of an evidence-based landscape-type hierarchy based 36 on the EcoSyst principles (Halvorsen 2020), based on standardisation of the amount of 37 landscape element compositional turnover between adjacent types .  38  39  40  41  STRUCTURE OF THE ARTICLE  42  43  44 Our study was conducted in several stages or phases, the interpreted results from each phase 45 forming the platform on which the next phase was built. The structure of this paper reflects 46 the structure of the study, by presenting the analytic results and their interpretation together 47 for each consecutive phase in a combined 'Results and interpretation' chapter. The final 48 chapter contains a more general discussion of the overall results and methodological issues. 49 1 2 3 4  5  6  7 STUDY AREA 8 9 10

MATERIAL AND METHODS
This study area spans latitudes from 57°57'N to 71°11'N and longitudes from 4°29'E to 11 31°10'E and comprises the entire mainland of Norway including the coastal zone, but 12 excluding the Svalbard archipelago, Jan Mayen and Bear Island. The range of variation in 13 natural conditions found in Norway includes most of the variation that can be found in the 14 circumboreal zone (Bryn et al. 2018), including both terrestrial, marine, limnic and snow and 15 ice ecosystems (Halvorsen et al. 2016). The study area is characterized by a wide range of 16 climatic variation; all seven temperature-related vegetation zones commonly recognised in 17 northern Europe, from boreo-nemoral to high alpine, occur in Norway (Bakkestuen et al. 18 2008). Norway has a high mineral and bedrock diversity (Ramberg et al. 2008), and a high 19 diversity of landforms (Gjessing 1978). 20 In addition to natural variation, the diversity of ecosystems in Norway is enhanced by 21 variation in human land use over time, that has affected the distribution and structure of 22 ecosystems throughout the country. Domestic animals and agriculture spread from the Middle 23 East to Norway around 3700 BC (Emanuelsson 2009), and by 1000 BC agriculture had 24 consolidated in the coastal regions of northern Norway as well. From the sixteenth century, 25 the Sami people started to herd semi-domesticated reindeer (Hansen & Olsen 2004). From 26 this time onwards, most Norwegian ecosystems have, to some extent, been affected by land-27 use activities such as domestic grazing, outfield fodder collection, heath burning, reindeer 28 husbandry, forestry, and industrial, urban and recreational development (Almås et al. 2004). 29 During the twentieth century, extensive land use based on collection of outfield resources 30 gradually declined. Recent landscape changes in Norway are among others related to 31 urbanisation (with associated changes in land use), intensified land use in central or highly 32 productive areas, abandonment of marginal agricultural areas, rapid forest encroachment in 33 previously open areas, reduction of areas with permanent snow cover and translocation of the 34 forest line into higher altitudes (Bryn et al. 2013, Fjellstad & Dramstad 1999. 35 Note that references to counties and municipalities in this study follow the 36 administrative division as it was in 2018. 37 38 39 40 ESTABLISHMENT OF OBSERVATION UNITS 41 42 43 A total of 100 sample areas, each 20×20 km, were selected using a stratified sampling 44 strategy. A subjective component was included in the sampling procedure to ensure that as 45 much as possible of the known variation in landscape properties in Norway was represented 46 in the data material (Fig. 1). Within each sample area, polygons/observation units (hereafter: 47 OUs) were delineated in terms of spatial landscape units using the criteria from the 'pilot 48 Nordland project' (see Erikstad et al. 2015;Halvorsen et al. 2016 and Supplementary material 49 Appendix S1). All 4166 OUs with more than a pre-defined fraction of their total area within 50 the sample area were included in the study (also the parts of the OUs outside the border of the 1 sample area limit were included, as indicated by the rugged outline of the OU clusters in Figs  2 2-3). All OUs were subjected to initial control, which formed the basis for removing 199 OUs 3 which failed to meet the following two criteria: (1) containing terrestrial land (193 marine 4 OUs); or (2) obvious error in the delimitation (digitisation error, etc.; 6 OUs). 5 The data set consisting of the remaining 3967 OUs was subjected to variable 6 transformation, etc. After initial analyses, polygon ID879, which should have been split into 7 two polygons according to the 'pilot Nordland criteria', was also removed. The total data set 8 used for the analyses therefore consisted of 3966 OUs. 9 The OUs were hierarchically nested within sample areas, and cannot be regarded as 10 strictly statistically independent observations. Based on the knowledge that spatial 11 autocorrelation in the data will tend to inflate the rate of Type I errors (rejection of a true null 12 hypothesis), we interpreted all statistical test conservatively, using test results for indication 13 rather than for hypothesis testing in the strict statistical sense (cf. For the purpose of interpretation of subsequent analyses, all OUs were characterised a priori 21 (i.e. before analyses) by assigning to each OU values for up to four types of variables. 22 (1) Landscape-type variables (LT variables), which affiliate the OUs with tentative 23 major-type group (Table 1) and major type at the landscape-type level in NiN version 2.0, i.e. 24 based upon 'pilot Nordland criteria' as implemented in the 'pilot Nordland project' 25 (Supplementary material, Appendix S1). 26 (2) Landscape-gradient variables (LG variables), which are ordered factor variables, 27 one for each landscape gradient identified in the 'pilot Nordland Project', upon which the 28 division into minor types in NiN version 2.0 was based. The LG variables (Table 2) provide a 29 more detailed characterisation of the OUs than the LT variables. An LG variable was used to 30 characterise an OU when the OU was part of a data subset in which the variable was relevant 31 to all OU members. 32 (3) Three categorical (unordered) landscape-factor variables (LF variables), which 33 addressed sediment/soil properties, were added to the LT and LG variables for descriptive 34 purposes (Table 3). The three variables were: (i) Sediment category (SK1), a hierarchically 35 nested categorical variable with classes defined on the basis of soil variables and additional 36 information on the proportion of sorted sediments. (ii) Primary soil/sediment class (JK1), an 37 unordered factor variable with up to 5 classes for dominant soil type. An OU was assigned to 38 a JK1 class when more than 50% of its area was covered by the sediment/soil type in 39 question. (iii) Secondary soil/sediment class (JK2), an unordered factor variable with up to 5 40 classes for co-dominant soil/sediment class. An OU was assigned to a JK2 class when more 41 than 25% of its area was covered by the sediment/soil type in question. 42 (4) Key variables, defined in NiN 2.0 as 'observable landscape properties or variables 43 derived from such properties, which are used for segmentation of a (complex) landscape 44 gradient', and used in the 'pilot Nordland project ' (Erikstad et al. 2015) to operationalise the 45 division of complex landscape gradients into segments. The raw key variables, here termed 46 'primary key variables' ( properties for the 81 100×100 m pixels with more than half of their area contained within a 49 neighbourhood circle of radius 500 m, centered on the focal 100×100 m pixel (Fig. 4) For all of Norway, information about properties that were observable on a landscape-relevant 9 scale were collected from available maps and databases. This information was organised into 10 111 primary descriptive variables, each represented by a binary (0/1) value for each pixel 11 (cell) in a raster with 100 m resolution, covering all of Norway. The binary variables thus 12 indicated presence or absence of the property in each 100×100 m pixel. These primary 13 variables are referred to as 'ecological basic maps'. Based on these, secondary, or analytic, 14 variables were derived to characterise the OUs of this study. Analytic variables may represent 15 mean values for primary variables, proportions of OU area with value of the primary variable 16 above a given threshold, etc. This is further explained below. 17 Many of the 111 secondary variables had strong pair-wise correlations. In order to 18 avoid a disproportionate emphasis on properties represented in the data by many, strongly 19 correlated, variables, secondary variables were successively removed until we obtained a set 20 of variables in which no variable pair had a Kendall's rank correlation coefficient of τ > 0.7. 21 This pre-selection reduced the number of variables to 85 (Table 5). This set of variables, 22 which will be referred to as 'the total analytic variables set', contains variables that can be 23 sorted into three statistical variable categories based on their relationship to the primary 24 variables: 25 (1) Pixel-number variables (PRA) were obtained by counting the number of 100×100 26 m pixels (grid cells) in which the primary variable in question had a value of 1, i.e. the 27 landscape property in question was recorded as present ('presence pixels'). Pixel-number 28 variables y were expressed as area fractions calculated as: 29 30 yP = (number of presence pixels)*10000/a (1)  31  32 where a is the total area of the OU, given in m 2 . 33 34 (2) Count, or density, variables (COU) were primarily obtained as counts of the actual 35 number of observable objects of a given category in an OU, e.g. the actual number of 36 buildings in the OU (and not the number of 100×100 m pixels containing at least one such 37 object = ln(ln(25 000 000)) -ln(ln(10000 * )) (3) 1 2 where x is the size of the largest island within the OU, given in km². This transformation  3  reversed the direction of the original variable x, providing low values for the mainland and  4 high values for the OUs in which the largest islands were small skerries and islets. The value 5 x = 2500 was used to represent the mainland as the largest island along the coast of the 6 Norwegian mainland, Hinnøya, comprises 2204 km². 7 The 85 analytic variables were also, regardless of statistical variable category, sorted 8 into three functional variable categories with subcategories, based on the type of landscape 9 property they describe: 10 (1) Basic geo-ecological variables (category G) describe basic natural conditions that 11 typically result in easily observable landscape properties (large-scale geomorphological 12 features such as lakes and rivers, glaciers, and other specific landforms incur massive loss of information. As a compromise between the individually desirable but 2 incompatible goals of skewness reduction and preservation of quantitative information in 3 variables with many zero values, we used the value for the transformation constant c which 4 resulted in the lowest value of x for presence (lowest original value > 0) of zmin = 0.200 after 5 transformation and ranging. The highest transformed presence value, zmax = 1.000 then 6 became 5 × zmin (the lowest presence value). 7 Transformed and ranged variables can be back-transformed to the original scale by 8 solving equations (4) or (5) inserted in (6) for x: 9 10 = * (max( )−min( ))+min(y)) − for = ln( + ) (7) 11 12 and 13 14 = ln ( * (max( )−min( ))+min( )) for = (8) 15 16 Relationships between original (untransformed) and transformed values for analytic 17 variables that are particularly relevant for assignment of OUs to segments along complex 18 landscape gradients, delineation of major types, etc., are given in Tables 6 and 7.  19  20  21  22  ORDINATION OF LANDSCAPE PROPERTY MATRICES FOR IDENTIFICATION OF  23  COMPLEX LANDSCAPE GRADIENTS  24  25  26 Ordination methods 27 28 Ordination methods were used to summarise the main gradients of coordinated variation in 29 landscape properties among the observation units, as characterised by the 85 analytic 30 variables. Ordination analyses were performed by a data-and result-driven procedure, starting 31 with the total data set (3966 OUs), and followed by analyses of subsets. The splitting into 32 subsets was based upon interpretation of the results of previous ordination analyses in terms 33 of tentative major-type groups, major types and groups of OUs of special interest. 34 Ordination methods sort observation units ( between the compared OUs (Swan 1970;Williamson 1978;De'ath 1999;Bouttier et al. 6 2003)]; monoMDS with maximum number of iterations (maxit) = 1000, convergence ratio for 7 stress reduction (smin) = 1•10 -7 and stopping criterion for lowest stress-value (sfgrmin) = 8 1•10 -7 . A minimum of 100 GNMDS solutions, all from random start configurations, were 9 obtained. The GNMDS solution with lowest stress was compared with the solution with 10 second-most lowest stress by a Procrustes test (Oksanen et al. 2016), by which the two 11 ordinations are compared after origo-translation, mirroring, and linear rescaling of axes. A 12 correlation coefficient (the Procrustres sum of squares) was used as a measure of the degree of 13 correspondence between the compared ordinations. A permutation test (protest) was used 14 to test the hypothesis: 'the degree of correspondence between the two ordinations is not better 15 than between two random point configurations'. The test was conducted by use of 999 16 permutations. A significant test (rejection of the hypothesis at the α = 0.001 level) was taken 17 as a necessary, but not sufficient, condition for ordination results to be potentially meaningful. 18 The graphic representation of the results of Procrustes analysis, by which the origo-translated, 19 mirror-imaged and linearly rescaled axes of one of the compared ordinations was plotted in 20 the first ordination (the plot(procrustes)command) was used to identify OUs with 21 unstable location (i.e. very different positions in the two ordinations). The fraction of OUs 22 with unstable location was used together with the procrustres sum of squares to judge the 23 stability and reliability of ordinations. 24 The solution with lowest stress was accepted as the best solution when, based on the 25 results of the Procrustes test and the number of OUs with unstable location, it was considered 26 as essentially equal to the ordination with the second lowest stress. In the opposite case, new 27 GNMDS ordinations were obtained, based upon new random initial configurations, until 28 either (1) two similar, lowest-stress solutions were found and the best was accepted, or (2) no 29 best solution was obtained from 1000 random initial configurations. 30 The best (lowest stress) solution was subjected to varimax rotation by PCA and 31 rescaled to half-change (H.C.) units using the postMDS function of vegan. Varimax rotation 32 translates the ordination space so that GNMDS axis 1 goes through the longitudinal axis of 33 the cloud of OUs in the ordination space (i.e. captures the maximum possible amount of 34 variation in the ordination space, expressed as sum of squared OU scores), axis 2 is 35 perpendicular to axis 1 and passes through the longest axis that is orthogonal to the first axis, 36 etc. Rescaling to H.C. units scales the axes in units that express landscape property 37 compositional turnover. One H.C. unit corresponds to an average difference in landscape 38 property composition of 0.5 units on the PD scale. 39 For each analysed data set or subset, GNMDS ordination was first performed with 40 number of dimensions k = 4 (i.e. to produce a four-dimensional GNMDS solution). 41 The best 4-dimensional solution was then compared to the DCA solution (DCA always finds 42 four axes) by calculating pair-wise Kendall's nonparametric correlation coefficients τ between 43 all pairs of axes. If all GNMDS axes had paired correlations with at least one DCA axis 44 greater than τ = 0.4, we considered the 4-dimensional GNMDS ordination to be confirmed by 45 the DCA ordination. In such cases, with 4 confirmed ordination axes, the 4-dimensional 46 GNMDS ordination was taken as basis for further interpretation. If the 4-dimensional 47 GNMDS ordination was not confirmed by DCA, the entire process was repeated for k = 3 48 and, unless the three axes of the 3-dimensional GNMDS ordination were confirmed by 3 or 4 49 DCA axes, the process was repeated once more and the 2-dimensional GNMDS ordination 50 compared to DCA. For all analysed data sets, both axes in the 2-dimensional GNMDS 1 ordination were confirmed by DCA .  2  In the few cases where the GNMDS ordination process had to be terminated without a  3  confirmed ordination result, the DCA ordination was used as the basis for further  4 interpretation with great caution. The purpose of parallel ordination is to avoid subjecting to 5 interpretation ordination results that have high probability of being burdened with 6 mathematical artefacts. Such artefacts are not likely to be present when two very different 7 methods, such as DCA and GNMDS, give closely similar results (Økland 1996, van Son & 8 Halvorsen 2014). The few cases without a confirmed GNMDS ordination result were always 9 obtained for large data sets (i.e. with many OUs). In these cases, all ordinations had about the 10 same stress value, ca. 0.27, and the Procrustes figure (Fig. 5) revealed that the ordination 11 results were unrelated. In these cases, the OUs were randomly distributed along the ordination 12 axes within a sphere-shaped point cloud (exemplified by Fig. 6). Such a result, which, as far 13 as we know, has not previously been reported in the literature, probably represents a case 14 where the relationships in the data matrix are so complex that the GNMDS iteration procedure 15 fails 'to break out of' the initial configuration. Interpretation of ordination results 21 22 Ordinations summarise correlations between the analytic variables and, at the same time, 23 reveal relationships between observation units (OUs). Because the variables in this case 24 express landscape properties and because identification of gradients in landscape properties is 25 the purpose of the analyses, ordination results are, strictly speaking, not in need of being 26 interpreted using extraneous data (i.e. data other than those used to obtain the ordination 27 results). In this respect, the ordination analysis of landscape element compositional data in 28 this study differs from ordinations of vegetation data which require interpretation by use of 29 external environmental information. 30 Nevertheless, all ordination results obtained in this study were subjected to a 31 formalised interpretation, the main purpose of which being to gain a best possible 32 understanding of the complex, co-ordinated patterns of variation in the landscape data sets as 33 a basis for identifying complex landscape gradients (CLGs is used throughout the text to mean |τ| > 0.3 (the alternative, |τ| < 0.3, is referred to as 'weakly 4 correlated'). 5 Several graphic tools were used in the interpretation. Vectors pointing in direction of 6 the largest increase in variable value were identified for key variables and selected analytic 7 variables using the envfit function in vegan. For factor variables (e.g. soil types), the envfit 8 function returns the centroid for the location of OUs belonging to each class. As a measure of 9 the strength of the relationship between each vector and OU locations in the ordination 10 diagram, the fraction of variation in the ordination diagram explained by the vector, r², was 11 calculated. Furthermore, ordination diagrams were provided with isolines for selected 12 variables, calculated by use of the ordisurf function in vegan. Symbols of different shapes 13 and colours for different landscape types and landscape gradient segments were used in the 14 ordination diagrams in order to illustrate the variation in landscape properties. 15 16 17 Identification of complex landscape gradients 18 19 Interpreted ordination results were used to identify candidate major landscape types as 20 maximally well-defined groups of OUs in ordination diagrams. In order to ensure that functional subcategories could be identified as CLG candidates 49 even if they were represented by a small number of KVs, a relaxed criterion (2) replaced 50 criterion (1) for subcategories represented by 3 or fewer variables. According to criterion (2), 1 a sufficient condition for being considered as a CLG candidate was that the variable had a 2 'noticeable' correlation (|τ| > 0.3) with the axis (or axes). For CLG candidates identified by 3 use of (2), variables were included if they satisfied criteria (i) and (ii), while (iii) was replaced 4 by (v), that the CLG candidate should include all variables with a noticeable correlation (|τ| > 5 0.3) with the axis (or axes) in question. 6 All CLG candidates identified by the criteria (1) or (2) were subjected to a judgement 7 of homogeneity to ensure that they consisted of variables that, in addition to representing one 8 and the same functional subcategory, made up a group of sufficiently strongly correlated 9 variables. If not, the CLG candidate was split into two or more CLG candidates, each 10 consisting of one or more variables. Splitting into two groups was enacted when the 11 characterising variables could be divided into subgroups in such a way that no pair of 12 variables, one variable from each subgroup, was pairwise strongly correlated (|τ| > 0.4) with 13 each other. The corollary of this splitting criterion is that variables that were quite strongly 14 correlated [noticably correlated in the case of criterion (2)] with at least one other variable 15 from the same subcategory constitute one CLG candidate. Furthermore, variables that were 16 not correlated with at least one other variable from the same subcategory made up a CLG 17 candidate on their own. 18 For each CLG candidate so identified, the analytic and key variables were sorted from 19 highest to lowest |τ| value with the ordination axis (or axes) in question. 20 21 22 Identification of parsimonious complex variable groups and calculation of explained 23 variation for each CLG candidate 24 25 The requirement for CLG candidates to be represented by groups of correlated variables (KV 26 groups) necessarily implied that the variation explained (in statistical sense) by the variables 27 in a group overlap extensively. In order to obtain a realistic estimate of the relative amount of 28 the variation in a data set that was explained (still in statistical sense) by a CLG candidate, 29 useful as a measure of the relative importance of the CLG candidate (as a source of variation 30 at the landscape type level), we obtained for each CLG candidate a parsimonious KV-group 31 (PKV group) consisting of the KV variables that provided a significant and independent 32 contribution to explaining variation in the analytic data set in question. The variable selection 33 process used to obtain the parsimonious KV-group is explained below. 34 The PKV group, and estimates for the variation it explained, were obtained using 35 constrained ordination by the Redundancy Analysis (RDA) method (Rao 1964, ter Braak 36 1985. RDA is the constrained variant of Principal Component Analysis (PCA), an ordination 37 method that finds axes which successfully maximises the explained sum of squares ('linear  38 variation') in a data set. In constrained ordination, the axes are constrained to express the 39 variation also explained by a set of constraining variables (Økland 1996). For a PKV group 40 with s variables, a maximum number of s constrained ordination axes can be extracted. These 41 are ordered by decreasing variation explained, as expressed by the eigenvalues of the axes. 42 The sum of the variation explained by each of the s constrained ordination axes was used as 43 an estimate for the total variation explained by a CLG candidate, represented by its PKV 44 group. 45 For each CLG candidate in each major type, the KV group was reduced to a PKV 46 group by forward selection of variables (e.g. Blanchet et al. 2008). Forward selection is an 47 iteration process that starts with calculating, for each of the m variables in the group, how 48 much of the variation in landscape-element composition the variable explains. The variation 49 explained (VE) by a variable was found by an RDA analysis with the variable in question as 50 the only constraining variable, as VEi = λ1/TI, where λ1 is the eigenvalue of the constrained 1 ordination axis and TI is total inertia, that is, the sum of the eigenvalues of all PCA axes that 2 can be extracted for the data set in question. TI is the total sum of squares for all analytic 3 variables [after centering (to mean = 0) and standardisation (to SD = 1)]. For PCA with 4 centred and standardised variables, TI = m, the number of analytic variables in the data set. 5 For simplicity, all variations explained were standardised to TI = 1. After the variation 6 explained by each KV had been found, the KV with the highest VE was incorporated in the 7 PKV group. In the second round of the iteration procedure, the analysis was repeated with the 8 remaining m -1 KVs as constraining variables and the KV selected in the first iteration cycle 9 as 'conditioning variable' in m -1 RDA analyses. In each of these RDA analyses, the method 10 first removed the variation explained by the chosen KV and then calculated how much of the 11 residual variation (the variation remaining after the variation explained by the selected KV is 12 accounted for) that was explained by each of the remaining variables. After all variables had 13 been subjected to RDA analyses, the variable which explained the largest proportion of the 14 residual variation was selected. In principle, the iteration process can continue until there are 15 no variables left or until no variation in the data set remains unexplained. In practice, 16 however, the process is stopped when none of the remaining variables provides a sufficiently 17 large independent contribution to explain the residual variation. To avoid inclusion of 18 variables that are highly correlated with variables already included in the PKV group, we 19 adopted the default practice in forward selection, to set a lower threshold value for the 20 (residual) explained variation that a variable must explain to be included. This limit can, in 21 principle, be determined by use of statistical tests. However, given the large data sets analysed 22 in this study (i.e. with many OUs), the null hypothesis that the explained variation is not 23 significantly higher than expected of a random variable, will be rejected with low p values in 24 almost all cases. In addition, testing is made difficult by the fact that we do not distinguish 25 between constraining variables and analytic variables (the latter is contained in the data set 26 subjected to RDA and therefore contributes 1/m of TI). We instead used as a lower limit ω0 = 27 g • λ0 where g is a constant and λ0 is the expected variation explained by a random variable, 28 expressed as a fraction of the residual variation. In round 1 of the iteration process, i.e. before 29 any variable has been included in the PKV group and the residual variation (before the round) 30 equals the total inertia, ω0 was set equal to: 31 32 ω 0 = 2 • 1 = 2 (9) 33 34 The residual variation after iteration cycle j, when j variables have been included in the PKV 35 group and the variation explained so far (the sum of the eigenvalues of the j constrained axes) 36 is λj, is 1 -λj. In cycle j + 1, the threshold value is given by the equation: 37 38 The parameter value 2 was selected after careful assessment of the results for the first 41 analysed data sets. For data sets with fewer than 200 OUs, with a more prominent stochastic 42 component of variation, the threshold value 2.5 was used. 43 The variable selection process was stopped when no more KVs met the threshold 44 criterion. The CLG candidate was represented by its PKV group (consisting of s variables) in 45 subsequent analyses. The variation explained by the CLG candidate was obtained as λs, the 46 sum of the eigenvalues of the s constrained ordination axes extracted with the PKV group as 47 conditioning variables. The first axis of the RDA ordination of the PKV group is the linear 48 combination of the s variables in the PKV group which explained the maximum amount of 1 variation in the data set. This axis is termed the CLG candidate's primary key variable. 2 3 4 Forward selection among CLG candidates for a set of orthogonal CLGs 5 6 An important lesson from the 'pilot Nordland project ' (Erikstad et al. 2015) was that the large 7 number of landscape gradients that were considered to be important for each major type, and 8 hence used to define minor types by the standard NiN procedure (see Halvorsen et al. 2020: 9 Appendix S2), resulted in an unmanageably high number of minor types, unsuited for 10 practical descriptive purposes. The practical solution chosen for the 'pilot Nordland project' 11 was to aggregate minor types subjectively into types at a level between minor and major 12 types. This procedure did, however, not comply with the standards of the EcoSyst framework 13 (of which NiN version 2 shall be an implementation; Halvorsen et al. 2020). Therefore, in 14 order to construct a system of minor types at the landscape level that may be theoretically 15 well founded and practically useful, the principle adopted at the ecosystem level in NiN 16 version 2 of major type-independent local complex-gradient concepts was not adopted at the 17 landscape level. Instead, the set of t accepted CLG candidates was reduced to a set of u 18 orthogonal (i.e. independent) CLG's. Forward selection (as described above) was used, but 19 with CLG candidates (each represented by its parsimonious KV group) as units in the 20 selection process instead of individual KVs. CLG selection was undertaken successively for 21 the three functional variable categories, first the basic geo-ecological variables (category G) 22 which describe the fundamental conditions for soil and vegetation; thereafter bio-ecological 23 variation (category U), which describe basic conditions for human exploitation of the 24 landscape; and, finally, land-use variables (category A). 25 In the first phase of this selection process, the CLG candidate in category G with the 26 largest amount of explained variation λs was chosen. Then, for each of the other CLG 27 candidates in the G category (represented by their PKV group), RDA was used to calculate 28 how much of the variation in the data set that was explained in addition to the variation 29 already explained by the first chosen CLG. After a careful assessment of preliminary results, 30 6% (0.06) variation (in the data set) explained was chosen as a lower threshold value for 31 inclusion of a CLG candidate in the group of orthogonal CLGs. For data sets with fewer than 32 200 OUs, with a more prominent stochastic component of variation, the value 0.08 was used. 33 When no more CLG candidates met the 0.06/0.08 criterion, the selection process was stopped. 34 Thereafter, the procedure was repeated for CLG candidates in the bio-ecological U category, 35 with all selected CLGs in the G category, represented by their PKVs, as conditioning 36 variables. The 0.06/0.08 criterion for inclusion of additional CLGs was applied. Finally, the 37 procedure was repeated for CLG candidates in the A category, with all selected CLGs in the 38 G and U categories, represented by their PKVs, as conditioning variables. 39 40 41 Orthogonal primary key variables for each CLG 42 43 The set of u CLGs found by forward selection of CLG candidates represents a parsimonious 44 set of orthogonal CLGs for the major type. Since each CLG is represented by a set of s = 1-4 45 variables, these orthogonal CLGs constitute a series of orthogonal subspaces of dimension s 46 within the data-dimensional space. CLGs were operationalised to one stepwise gradient each 47 by reduction of each of these subspaces to one dimension. In line with the procedure 48 described above, RDA was used to find the vector in data-dimensional space that captured the 49 maximum possible variation in the s-dimensional space spanned by the CLG. This vector, 50 referred to as the orthogonal key variable (ONV) for the CLG, was given by the first RDA 1 axis in the constrained ordination of the PKV group of the CLG, with all variables included in 2 all PKV-groups of all previously identified CLGs as conditioning variables. 3 The ONVs were extracted as actual RDA score vectors from the respective 4 constrained ordinations by the vegan function scores using the following parameter 5 options: display = "lc" [extracts scores that are linear combinations of the constraining 6 variable ('linear combinations = "lc scores"'), while the alternative 'weighted average scores 7 = "wa scores" extracts scores after one iteration cycle towards the first PCA axis for the data 8 set in question]; scaling = 1 (provides correlation biplot scaling of scores, which optimises 9 the relationship between the cosine of the angle between the analytic variable vectors in 10 ordination space and their Pearson's product-moment correlation coefficient). 11 After extraction, the ONVs were ranged onto a standard 0-1 scale by the procedure 12 used to transform analytic variables. The term ONV is used also for these ranged 'primary 13 ONVs'. 14 15 16 Estimation of landscape distance 17 18 The requirements for a standardised division of complex landscape gradients ( The parallel to the species as key characteristic at the ecosystem level is landscape 31 elements (defined as 'natural or human-induced objects or characteristics, including spatial 32 units assigned to types at an ecodiversity level lower than the landscape level, which can be 33 identified and observed on a spatial scale relevant for the landscape level of ecodiversity'). 34 The parallel to the species composition, which is used to quantify differences between 35 ecosystem-type candidates, is the landscape-element composition, defined as 'the landscape 36 elements occurring within a specific area, quantified by an appropriate performance measure '. 37 In this study, the landscape-element composition of each OU was represented by the 85 38 analytic variables, quantified by use of the transformed and ranged performance scale. 39 In EcoSyst, the unit of compositional turnover of the key characteristic at an 40 ecodiversity level along a complex variable in the key source of variation at this level, is 41 termed the 'ecodiversity distance unit' (EDU; Halvorsen et al. 2020 composition also at the landscape-type level. 7 The so-called 'generalisation challenge' in calculating differences in species 8 composition is also present when differences in landscape-element composition are 9 quantified. This challenge implies that the expected inequality in landscape-element 10 composition between OUs located at the same position along a CLG differs from zero 11 because of random variation ('noise') in landscape-element composition. Accordingly, 12 measurements of compositional differences between OUs cannot be added and subtracted 13 unless the 'inequality between repeats', or 'internal association', i.e. the compositional 14 difference between OUs placed at the same point along the underlying CLG (INA; Whittaker 15 1952, 1960, Økland 1990), can be quantified. At the ecosystem level, the generalisation 16 challenge is resolved by using generalised species-list data to calculate ecodiversity distance 17 units in ecosystems (EDU-E;Halvorsen et al. 2020). Such data contain 'smooth' ('noise-18 free') species abundance data for each candidate ecosystem type, derived from models of 19 species' abundance distributions along local complex environmental gradients. These models 20 are obtained by consensus among experts, following a standard procedure which makes use of 21 all available data as well as informal knowledge (see Halvorsen et al. 2016: chapter B2c). Our 22 knowledge about variation at the landscape-type level is, however, so fragmentary that the 23 method for using generalised data sets assembled by expert judgement was inapplicable. 24 Instead, we chose an alternative, five-step procedure for estimation of ecodiversity distance in 25 landscapes: 26 27 (1) Division of CLG's ONV into segments (ONV segments) into which the OUs are 28 filed.

29
(2) Calculation of landscape element profiles (LEPs) for each ONV segment. The 30 landscape element profiles are the parallels to the generalised species lists at the ecosystem 31 level. 32 (3) Calculation of proportional dissimilarity (PD) between the LEPs. 33 (4) Estimation of the 'inequality between repeats' (INA) from the PD values and 34 calculation of landscape distances (LD) between ONV segments. 35 (5) Estimation of the gradient length of the CLG, in EDU-L units. 36 37 (1) Segmentation of CLGs by their ONV. Ideally, we intend (i) to divide each ONV 38 into segments that are as small as possible, in order to be able to describe the variation in 39 landscape element composition along the ONV in the greatest possible detail; (ii) to make all 40 ONV segments equally wide (i.e. span similar landscape-element compositional turnover 41 along the CLG) and thus to make them ideal minor-type candidates; and (iii) that each ONV 42 segment is represented by as many OUs as possible to reduce random variation as much as 43 possible and, at the same time, increase the degree of generalisation. Conditions (i) and (iii)  44 are incompatible, calling for a best possible compromise. Furthermore, condition (ii) is 45 incompatible with both of (i) and (iii) because the distribution of OUs along most ONVs is 46 unimodal (more or less skewed). Accordingly, equal segment width (measured on the 0-1 47 scale of the ranged ONV) typically implies that segments near CLG ends contain very few 48 OUs. 49 Based on the above considerations, the following criteria were used to divide ONVs 1 into segments: (i) No segment should contain less than n = 25 OUs. (ii) The number of 2 intervals q = 8. (iii) Each segment should have the same width, measured in the units by 3 which the ONV is scaled. The criteria are arranged hierarchically, so that criterion (i) has 4 priority over criterion (ii) which, in turn, has priority over criterion (iii). Application of the 5 criteria can be exemplified by the affiliation of the 118 OUs in major-type candidate IF 6 (inland sediment plains) to segments. Regardless of which CLG is addressed, criterion (i) sets 7 an upper limit of four (q = 4) for the number of segments, criterion (i) overruling criterion (ii). 8 While criterium (iii) suggests that the four segments comprise OUs in the ranges 0-0.25, 9 0.25-0.5, 0.5-0.75 and 0.75-1 ONV units, criterion (iii) may be overruled by criterion (i), 10 implying that segments near ONV endpoints comprise the 25 OUs with the lowest, respective 11 highest, ONV scores. Thereafter, the remaining OUs are distributed on two segments of equal 12 width (criterion iii) if not again overruled by criterion (i) which demands a minimum of 25 13 OUs in each segment. 14 (2) Calculation of LEPs for each ONV segment. For each of the q ONV segments, the 15 mean value of each (zero-skewed transformed and ranged) analytic variable was calculated. 16 The vector of mean values for each analytic variable in the set of OUs that represent the ONV 17 segment, the 'centre of gravity-vector', was used as landscape element profile. 18 (3) Calculation of PD between the LEPs. The PD (proportional dissimilarity index) is 19 the one among the dissimilarity indices in common use, e.g. in vegetation ecology, which is 20 best suited for calculating ecodiversity distance in ecosystems (Økland 1986). This holds true 21 also for other situations in which the response variables (here: the analytic variables) are in 22 general linearly or unimodally related to the underlying complex gradients (here: CLGs). In 23 its simplest form, calculated for pairs of ONV intervals j and k, PD is given by the formula 24 25 26 where xij is the value of variable i in ONV segment j, xik is the value of variable i in ONV 30 segment k, and m is the total number of variables (i.e. the total number of variables with 31 variation in the major-type candidate in question). The PD index is bounded below by 0 and 32 above by 1; the value 0 indicates that the landscape-element composition of two ONV 33 segments is identical (i.e. that each variable has the same mean value in segments j and k), the 34 value 1 represents the situation (which in landscape contexts is completely unrealistic) that 35 one of the two compared LEPs has the value 0 while the other has a positive value for each 36 variable.

37
PD was calculated between all pairs of ONV segments j and k from 1 to q = 8, j ≠ k.

(4) Estimation of 'inequality between repeats' (INA) based on estimated PD values 39
and calculation of landscape distances (LD) between ONV segments. PD values between 40 completely generalised vectors (without random variation) of variables that vary linearly 41 along a gradient, are additive (Økland 1986). In mathematical terms, this means that PD 42 differences between three ONV segments, represented by centres of gravity A, B and C, 43 where A < B < C along the gradient, satisfy the equation If the LEP profiles satisfied demands for complete generalisation and additivity, PD 48 could be used directly as a measure of landscape distance between centres of gravity of 49 ONVs, after linear rescaling of PD values to EDU-L units (i.e. by multiplication with a 1 constant w). Inspection of landscape-element compositional data for all ONVs in all major-2 type candidates (see Results) did, however, show that PD(A,C) > PD(A,B) + PD(B,C) almost 3 throughout and, thus, that the requirement for (complete) generalisation is never fulfilled. 4 Moreover, the results show that the linearity requirement is fulfilled to varying degrees. 5 Extreme cases by which PD(A,C) < PD(A,B) + PD(B,C), that is, that the similarity in 6 landscape element composition between A and C is greater than between A and B or between 7 B and C, also exist. This calls for correction of raw PD values for INA, which was 8 accomplished in accordance with the following line of reasoning: 9 Let LD(A,C) denote the underlying, true landscape distance (after correction for INA) 10 between centres of gravity of segments A and C. It follows from the definition of landscape 11 distance EDU-L that 12 13 LD Equation (17)  gradient. Let us assume that the gradient has been divided into 8 segments with points of 6 gravity ABCDEFGH, each represented by an LEP. Equation (17)  For the outermost ranges, only one INA estimate is available, e.g. for INAAB given by 23 24 segments 3 and 4, 4 and 5, and 5 and 6. Method II clearly overestimated gradient length. 46 Inspection of the PD matrix (e.g. Fig. 7) reveals the reason for the failure of Method II: that 47 the landscape element composition of segment 1 is more similar to the composition of 48 segments 6-8 than to segment 5. Similarly, segment 8 is more similar to segments 3-5 than to 49 segment 6. The effect of nonlinear variation along the gradient is corrected for by Method III,1 which for this gradient yields gradient length estimates intermediate between those provided 2 by the two other methods. 3 (5) Estimation of total landscape gradient length. Total landscape-gradient length 4 (LGL), on the corrected PD inequality scale, was calculated by summing up LD values for all 5 intervals 12, 23, 34, 45, 56, 67 and 78 into which the ONV was divided. Additionally, based 6 on the same reasoning as for the ecosystem level (Halvorsen et al. 2016(Halvorsen et al. , 2020, we assume 7 that the LEPs represent the centre of gravity in each interval (centre of the interval), i.e. that 8 the underlying gradient extends beyond the centre of gravity of the intervals 1 and 8 by half 9 the LD to the centres of gravity of the neighbouring segments. Accordingly, the LGL was 10 estimated by the equation: 11 12 LGL = 0.5•LA(1,2) + LA(1,2) + LA(2,3) + .... + LA(6,7) + LA(7,8) + 0.5•LA(7,8) 13 (24) 14 15 LGL = 1.5•LA(1,2) + LA(2,3) + .... + LA(6,7) + 1.5•LA(7,8) Rescaling of ONVs and standardised division into segments 31 32 Each ONV, representing one CLG in one given major-type candidate, was rescaled by use of 33 the results of LEP dissimilarity analyses according to the following procedure: 34 (1) A new measurement scale from 0 to LGL was established for each ONV by first 35 placing the centre of gravity of ONV segment 1, TP1, at 0.5 • LD(1,2). Then the centre of 36 gravity TP2 of ONV segment 2 was placed at 0.5 • LD(1,2) + LD(1,2) = 1.5 • LD(1,2), TP3 at 37 1.5 • LD(1,2) + LD(2,3) etc. 38 (2) Borders between the intervals, GR12, GR23 etc., on the new measurement scale 39 were then estimated by 'smoothing' to further reduce the effect of stochasticity in landscape 40 element composition by the equations: GRx(x+1) = GR(x-1)x + 0.5•(LA(x-1,x) + LA(x,x+1)) (28) 45 46 (3) Positions of OEs along an ONV relative to this new scale were found by linear 47 interpolation within each segment: Let IGRxy denote positions of borders between ONV 48 segments x and y along the ONV as originally scaled from 0 to 1 and let y be the position of 49 one specific OE on this scale. Within each interval, a linear rescaling was then done by 50 converting the position t of an OU, given on the original measurement scale, into a new 1 position z on the PD measurement scale from 0 to LGL using the equation: The vector Z is denoted the rescaled ONV (rONV). 6 (4) For use in some analyses, the rONV vector was ranged to a scale from 0 to 1 to compare 7 relative placements along the rONVs. These ranged, rescaled ONVs are referred to as 8 rrONVs. 9 A standard segment along a CLG, represented by an rONV, is defined as an interval 10 along the CLG/ONV with a minimum extent of at least 1 EDU-L unit (= 0.08 PD units). This 11 definition accords with the general definition of standard segments in EcoSyst (Halvorsen et 12 al. 2020): 'one in a set of intervals into which a complex-gradient is divided, that is made up 13 by one, two or more elementary segments, each comprising at least 1 ecodiversity distance 14 unit (EDU) of variation in the key characteristic within the major type in question'. As with 15 the standardised division of LECs, the number of major-type adapted segments is set to 16 17 A fundamental principle in the 'pilot Nordland project', which is carried over to NiN version 48 2, is that the correlative analyses of variation in landscape-element composition in large data 49 sets is just a first (but crucial) step on the road towards a system of landscape types. After 50 thorough interpretation, the results are discussed and operationalised. The first step towards 1 operationalisation of the analytic results, was to 'translate' the rONV representation of CLGs 2 to key variables or other explicit criteria that could be used in practice for drawing borders 3 between standard segments. In some cases, these key variables or criteria can be used directly 4 to draw boundaries between spatial landscape units, in other cases, further iterations are 5 needed for 'operationalisation', e.g. by modelling or by defining operationalised key variables 6 (OpNV). In accordance with the stated aim, the present study includes a discussion of 7 potential OpNVs for each CLG. The use of these key variables to obtain a landscape-type 8 map for Norway is treated in a companion paper (Simensen et al., in prep.). 9 10 Ordination of Tot3966the total data set 8 9

RESULTS AND INTERPRETATION
The 3966 OUs were distributed on landscape-type variables (major types in the 'pilot 10 Nordland 2076. Note that the IS (inland plains) major type was not recognised in Nordland and was 13 therefore not included among landscape-type variables. 14 The total data set was too large to be ordinated by GNMDS, and unconfirmed DCA 15 ordination results are therefore reported. The DCA ordination had two strong axes, as 16 indicated by a considerable reduction of eigenvalues and gradient lengths from DCA axis 2 to 17 DCA axis 3 ( Table 8). The ratio of gradient lengths for DCA axes 2 and 1 was 0.945. Further 18 description and interpretation was therefore based upon DCA axes 1 and 2. 19 The ordination diagram for DCA axes 1 and 2 ( Fig. 8) had no visible artefacts and 20 separated coast from inland and, accordingly, fjord (KF) from valley (ID); see Fig. 9. In order 21 to deduce precise criteria for the border between coast and inland (e.g. at the major-type group 22 level), OUs with special combinations of characteristics (in the borderland between coast and 23 inland) were studied in more detail (Table 9; Figs 10-17). With few exceptions, OUs without 24 coastline were clearly separated from OUs with coastline also in the sparse area in the 25 ordination diagram (Fig. 10). This also applied to OUs which, due to inaccuracy in the 26 delineation of OUs, possessed a small element of coastline (Figs 10,14); also these OUs 27 possessed 'coastal characteristics'. The only group in Table 9 that was divided into almost 28 equally large subgroups on either side of the sparse area, was group 2 of used to sort subset X162 OUs on inland and coastal groups, the lowest misclassification rate 6 (proportion of OUs that were misclassified to the inland and coastal groups, based upon 7 Oyst_i values), 14.8%, would be obtained for Oyst_i = 0.733, i.e. by including OUs of islands 8 < 10 km 2 without coastline in the 'coastal group' (Fig. 17). 9 Conclusion. The improvement of precision in assignment of island OUs without 10 coastline obtained by splitting these OUs onto 'inland' vs. 'coastal' groups in the DCA 11 ordination of the Tot3966 data set by transferring OUs on islands < 10 km 2 to coastal 12 landscapes, is marginal. Hence, we judge the disadvantages of a more complex definition of 13 these two major landscape-type groups larger than the advantages of a marginally reduced 14 misclassification rate. We therefore conclude that island OUs without coastline should be 15 affiliated with the inland group of OUs. 16 17 18 Examination of data subset X64 -OUs assigned to LT-type KS (coastal plains) located on 19 islands 20 21 Of the 64 OUs in data subset X64, 7 were located to the right of the dotted line through the 22 sparse region in the Tot3966 DCA ordination diagram (Fig. 18) and the remaining 57 OUs 23 were located to the left of that line. Neither island size nor primary key variables were 24 correlated with positions of these coastal-plain OUs in the ordination diagram ( Fig. 18; results 25 of correlation analyses not shown). 26 Conclusion. Island OUs assigned to LT-type coastal plains should be affiliated with 27 the inland group of OUs. 28 29 30 Conclusion on division into major-type groups 31 32 The analyses support a principal division of OUs into two major-type groups; inland 33 landscapes (I) and coastal landscapes (K). There are good reasons to keep strictly marine 34 landscapes (i.e., without coastline and terrestrial land) as a third major-type group (M). 35 The results of the analyses of Tot3966 and subsets X162 and X64 support separate 36 analyses of one coastal data set (K810) and one inland data set (I3150) as the next step in the 37 analysis. Six inland OUs with traces of coastline or incomplete records of some variables 38 were removed before further analysis. Ordination of K810the coastal landscapes data subset 46 47 All four axes of the 4-dimensional GNMDS ordination of the K810 data subset (Fig. 20) were 48 confirmed by the corresponding DCA axis ( Fig. 19; Table 10). The gradient lengths indicated 49 two strong axes (Table 10; axis 1 and 2, respectively). The GNMDS ordination had Procrustes 50 SS = 0.0092 and no unstable OUs (i.e. OUs that obtained widely different positions in the two 1 lowest-stress GNMDS solutions). The first two GNMDS ordination axes separated OUs 2 assigned to 'pilot Nordland' major types coastal plains (LT-variable KS) and fjord (KF) along 3 a line from the upper left to the lower right corners of Fig. 21. The black, continuous line in 4 Fig. 21 indicates the transition from preponderance of KS OUs (lower left) to preponderance 5 of KF OUs (upper right). Taking this line as the truth, misclassification rates were 47/350 = 6 13.4% for KF and 53/457 = 11.6% for KS OUs, respectively. 7 Isolines for Oyst_i in Fig. 21 established island size as an important variable for 8 explaining variation among KS OUs, while island size was less important in fjords. The three 9 OUs assigned to 'pilot Nordland' major type KA (coastal hill and mountain landscapes) were 10 placed inside the cluster of KF OUs. This suggests that there is no fundamental difference 11 between the side of a fjord and a mountain coast, most likely because, at the outset, there is no 12 logical reason why a steep fjord side and a steep coastline not situated in a fjord-like landform 13 should differ in other landscape features than, perhaps, terrain roughness. Furthermore, this 14 indicates that the criterion adopted in the 'pilot Nordland project', to divide fjord segments 15 into separate OUs for the two sides when they have different landscape element composition 16 and the width of the fjord exceeds 1 km, should be carried over to NiN version 2. 17 Conclusion. The analyses support a principal distinction between coastal plains and 18 fjords, based on the criteria that were used to distinguish two major types in the 'pilot 19 Nordland project' (Appendix S1). The sparse data material does not support discrimination of 20 coastal hills and mountains as a major type on its own. 21 22 23 Ordination of I3150the inland landscapes data subset 24 25 After removal of analytic variables that express coastal characteristics, 67 variables were 26 retained in the inland data subset I3150. This data subset was too large to be ordinated by 27 GNMDS, and unconfirmed DCA ordination results are therefore reported (Table 11). DCA 28 axis 1 had an eigenvalue more than twice as large as that of DCA axis 2, which in turn had an 29 eigenvalue 1.8× that of DCA axis 3. The gradient length of DCA axis 1 was ca. 1.5× that of 30 DCA axis 2. This indicated existence of one strong and one less strong, but potentially 31 interpretable, gradient in landscape-element composition in the I3150 data subset. The DCA 32 ordination diagram for axes 1 and 2 ( Fig. 22) did not clearly separate the three relevant major 33 types of the 'pilot Nordland project': inland hills and mountains (IA), (inland) valleys (ID) 34 and coastal plains (KS), i.e. landscapes close to the coast but without coastline. A slight 35 tendency for concentration of IA-affiliated OUs at low DCA axis 1 scores (to the left) and ID-36 and KS-affiliated OUs to the right in Fig. 22 could, however, be observed. Affinity to IA, ID 37 or KS did, however, only explain 3.7% of the variation in I3150 (ANOVA: F2,3147 = 60.35, r 2 38 = 0.0369, P < 0.0001). 39 Correlations with primary key and analytic variables established DCA axis 1 as an 40 infrastructure gradient, while DCA axes 2-4 had no obvious interpretation in terms of 41 explanatory variables (Table 12). Furthermore, the correlation analysis also suggested that 42 variation in the intensity of human exploitation was conditioned on underlying geo-ecological 43 variation. Geo-ecological variables correlated with DCA axis 1 (τ values in parentheses) were 44 (Table 14): areal coverage of rugged terrain (Guro_t_a; τ = -0.4391); relief (RR1; τ = -45 0.4169); areal coverage of flat terrain (Flat_a; τ = 0.4032); areal coverage of steep terrain 46 (Steep_a; τ = -0.3890); mean terrain ruggedness (Rug3_m; τ = -0.3554); and areal coverage 47 of terrain depressions (TPI1h_a; τ = -0.3151). This supported interpretation of DCA axis 1 48 also as a gradient from a rough and coarse-scaled, alpine terrain towards a more gentle, less 49 'rugged' terrain. Glaciers, as indicated by variables BrI and Glac_a, and several soil variables, 50 were also correlated with DCA 1 (τ> 0.30; see Table 12). While glaciers were concentrated to 1 low DCA axis 1 scores, the opposite was true for mires. 2 Conclusion. Besides identifying a clear main gradient in landscape-element 3 composition related to land use and observable infrastructure, the ordination results (DCA 4 axis 1) support existence a rough, geomorphologically based subdivision of inland landscapes 5 into valleys and 'non-valley landscapes'. However, the variation in landscape element 6 composition in the inland data set I3150, as revealed by DCA ordination of all 67 analytic 7 variables, appears so complex that important structural patterns may be hidden in this 8 complexity. Also, the fact that the ordination of the I3150 data subset indicated that the 9 dominant infrastructure gradient is conditioned on variation in geo-ecological 10 (geomorphological) properties, may indicate that variation in landscape-element composition 11 relative to basic geomorphological variation is concealed behind, or overshadowed by, land-12 use related variation. A new ordination of I3150, using the basic geo-ecological analytic 13 variables (category G) only, was therefore undertaken, in order to clarify the basis for 14 dividing inland landscape into major types. 15 16 17 Ordination of I3150Gthe inland landscapes data subset, using basic geo-ecological 18 analytic variables only 19 20 The reduction from 67 to 35 variables did not change the fact that the I3150 data subset was 21 too large to be ordinated by GNMDS, and results of an unconfirmed DCA ordination are 22 therefore reported (Table 13). DCA axis 1 had an eigenvalue almost twice as large as that of 23 DCA-axes 2-4, while the gradient lengths of the four DCA axes were in the same range 24 (Table 13). The OUs were evenly distributed in the space spanned by DCA ordination axes 1 25 and 2, without obvious artefacts (a very weak tongue effect can be traced on DCA axis 2). 26 Analysis of correlations between DCA axes and the primary key and analytic variables 27 (Table 14) revealed a very strong relationship between DCA axis 1 and terrain ruggedness 28 (relief-related variables) and a distinct relationship between DCA axis 2 and valley shape 29 (DNI variable, expressing the relationship between valley depth and valley width). The results 30 of this ordination were examined graphically in greater detail , with the aim of 31 finding a supported division into major types. 32 Fig. 23 shows a relatively sharp differentiation between valleys (green symbols) and 33 other landscapes (hill and mountain landscapes and plains) along DCA axis 2 (the shift from 34 preponderance of red or blue vs. green dots in Fig. 23 is indicated by the blue line). Despite 35 the relatively abrupt shift, none of the variables were more than noticeably correlated with 36 DCA axis 2 ( almost complete lack of 'inland coastal plain' OUs near the low-score end of DCA axis 1 47 (Fig. 23) suggests existence of inland landscapes with features more typical of plains, such as 48 lower relative relief and higher proportion of flat terrain, than the 'inland coastal plain'. This 49 result thus supports separation of inland plains as a major type on its own, parallel to the 1 separation of coastal plains (KS) from fjords (KF) in the coastal major-type group. 2 The degree of sediment sorting (landscape-factor variable SK1: category S; cf. Table  3 3) increased towards the left in the ordination diagram (low DCA axis 1 scores), i.e. towards 4 OUs with characteristics typical of plains ( Fig. 25) such as low relative relief and high 5 proportion of flat terrain (Figs 23-24). Fig. 25 shows that OUs with a high proportion of their 6 area covered by flat terrain also had relative relief << 100 m, suggesting that these two 7 characteristics, separately or in combination, can be used to separate plains from 8 hills/mountain landscapes. Detailed examination of selected OUs from different parts of 9 Norway (not shown) did not show clear patterns with respect to position in the ordination 10 diagram on one hand and sediment category SK1 or relative relief RR1 on the other. 11 Close examination of Fig. 26, designed to fine-tune criteria for separating plains from 12 hills and mountains on the basis of relative relief and proportion of flat terrain (note that the 13 values for Flat_a of 0.7 and 0.8 correspond to back-transformed area proportions of 100×100 14 m pixels with slope < 2° of approx. 20% and 35%, respectively; see the chapter 15 'Transformation of analytic variables') shows (i) that RR = 50 is too strict to be used as a 16 separating criterion; most likely because the RR1 index is vulnerable to edge effects; (ii) RR 17 = 100 is too liberal, affiliating with plains many OUs that do not comply with a common 18 understanding of a plain (e.g. at Romerike, SE Norway); and (iii) that Flat_a = 0.75 is likely 19 to be the one among the three potential criteria that provides the sharpest borderline (note that 20 all large, but not all small blue rings in Fig. 26  Ordination of IS520the inland plains data subset 30 31 Two confirmed ordination axes were found for the IS520 data subset ( between GNMDS axis 1 and land-use variables (infrastructure as well as agriculture, cf. Table  38 16), support the interpretation that the main gradient in landscape-element composition 39 primarily expresses variation in characteristics related to human influence, and to a gradient 40 from lowlands to mountain plains. Furthermore, the strong correlation between this axis and 41 the proportion of OU area covered by marine deposits (Kmar_a) may indicate an underlying 42 geo-ecological structure related to dominant soil type (Fig. 36, Table 16). The 520 OUs were 43 distributed onto soil classes (JK1) as follows; B (exposed bedrock): n = 60; E (glaciofluvial Fine-sediment plains formed a distinct cluster near the high-score end of GNMDS axis 1 (Fig.  50 40), also characterised by high values for the infrastructure (Iflu; Fig. 37) and agricultural 1 land-use indices (Fig. 38). This shows that agriculture and other types of human exploitation 2 are more or less confined to fine-sediment plains. The NiN principle that major (landscape) 3 types shall differ with respect to complex-variable groups suggests splitting of inland plains 4 into two major types. 5 Fig. 40 indicates that marine sediments have a particularly strong structuring effect on 6 landscape-element composition. This is substantiated by examination of OUs dominated by 7 bare rock (JK1 = B) and co-dominated by marine sediments (JK2 = M), which cluster with 8 JK1 = M OUs. 9 Conclusion. The results lend support to a division of inland plains into two major 10 types; (1) fine-sediment plains (dominated by marine and glaciofluvial deposits); and (2)  11 'other plains'. Based on ordination results for the IS530 data subset, inland plains were 12 assigned to fine-sediment plains if at least one of the following five criteria were satisfied: (1) 13 proportion of area with marine deposits > 50%; (2) proportion of area with glaciofluvial 14 sediments > 50%; (3) if no sediment type covers 50%, that the total proportion of area with 15 fine sediments (marine deposits and glaciofluvial sediments) > 50%; (4) proportion of area 16 with marine deposits > 20% and cover of exposed bedrock > 50%; (v) proportion of area with 17 marine deposits > 10%, and no dominant soil class (i.e. no single soil class covers more than 18 50% of OU area). 19 Three OUs with positions in the ordination diagram that did not match their sediment 20 composition were excluded from further analyses of the inland plains data subsets. A total of 457 OUs were affiliated with the coastal plain (KS) major-type candidate. All four 31 axes in the 4-dimensional GNMDS ordination of the KS457 data subset were confirmed by 32 the corresponding DCA axes (Table 17). The GNMDS ordination had Procrustes SS = 0.0363 33 and seven unstable OUs. The OUs were evenly spread out in the space spanned by DCA ( Fig.  34 41) and GNMDS ( Fig. 42) ordination axes 1 and 2, and no visible artefacts could be observed. 35 Gradient lengths of both GNMDS and DCA axes and eigenvalues of the latter showed three 36 relatively strong axes (i.e. relative length > 0.7× that of the longest axis; Table 17). 37 Correlations between key and analytic variables on one hand, and GNMDS axis 1 scores on 38 the other, supported by vector (Figs 42-43) and isoline diagrams (Figs 44-49) revealed that 39 the first axis was very strongly related to land-use intensity (infrastructure and agricultural 40 land use; see Figs 44 and 45, respectively), with strong relationships also to the geo-41 ecological gradient from inner to outer coast (Fig. 47, Table 18). 42 The OU (Id 2916) that occupied the extreme position along GNMDS axis 1 (lower 43 right corner, Fig. 44) was affiliated with OI class 6 ('big city'), actually comprising parts of 44 the centre of Oslo, the largest city in Norway. Most bio-ecological variables (functional 45 category U) were strongly correlated with GNMDS axis 1, increasing from 0 (or near 0) for 46 low GNMDS axis 1 scores towards higher GNMDS axis 1 scores (Fig. 42). An example of 47 the pattern of variation of a category-U variable is given by the proportion of area covered by 48 deciduous forest (Arlov_a; Fig. 46). The proportion of OU area covered by bare rock 49 decreased with increasing island size (Fig. 48, Table 18). A similar pattern along GNMDS 50 axis 1 was obtained for the LG-variable coastal/archipelago properties (SN; see Table 2). Fig.  1 47 shows a clear (inverse) island-size gradient from low to higher GNMDS axis 1 scores. The 2 segregation of OUs along GNMDS axis 1 according to island size was particularly clear-cut 3 for islands smaller than ca. 20 km 2 (Oyst_i > 0.7; SN ≥ 3). This opens for the possibility that 4 the complex gradient expressed on GNMDS axis 1 actually consists of more than one CLG. 5 Among the 69 OUs in the KS457 data subset without infrastructure (IfI = 0), 56 6 possessed a largest island that was smaller than 1 km 2 . Islands with extensive infrastructure 7 were generally much larger; an exception being Id 4038 which had Oyst_i = 0.8647 and IfI = 8 9.34. This OU comprises the SW part of Nøtterøy (Vestfold), which has been subjected to 9 extensive summer-cabin development. 10 Figs 44 and 45 show that agricultural land use is conditioned on a minimum of 11 infrastructure; JP ≥ 2 implies OI ≥ 3. A closer examination revealed a unimodal relationship 12 between JI and IfI ( Fig. 50), which implies that peak agricultural land-use intensity was 13 associated with a well-developed infrastructure, decreasing towards cities (Figs 44-45; Fig.  14 50). 15 The second GNMDS axis was strongly related to terrain shape; values of relative 16 relief, ruggedness and related variables increased towards high GNMDS axis 2 scores ( Fig.  17 43, 49). Furthermore, the proportions of flat vs steep 100×100 m pixels decreased and 18 increased, respectively, along this axis. The strong signal from topographic variables in 19 coastal plains (Fig. 49) at a first glance contrasts the general idea of a plain. 20 GNMDS axis 3 was correlated, but not strongly, with variables that expressed 21 presence and areal cover of mires, lakes and boreal heaths ( Fig. 51; Table 18). Mires and 22 lakes became more prominent with increasing island size and were more or less absent from 23 small islands (Figs 52-53). 24 GNMDS axis 4 was correlated with variables that express coastal complexity, e.g. the 25 number of marine islands (Table 18, Fig. 55). The noticeable correlation of GNMDS axis 4 26 with proportion of area covered by marine deposits illustrates that sediments are mainly 27 deposited on larger land units (Fig. 55). 28 Variation in dominant soil/sediment class within the KS457 data subset was explored 29 to assess if coastal plains, like inland plains, should be divided further. The 457 OUs were 30 distributed onto dominant (> 50% cover) soil classes (JK1) as follows; B (> 50% exposed 31 bedrock): n = 346; H (marine deposits): n = 63; 0 (unsorted sediments): n = 61. In addition, 83 32 OUs belonged to JK2-class (> 25% cover) H. The distribution of JK1 classes in the GNMDS 33 ordination diagram for axis 1 and 2 showed a weak relationship between JK-class and 34 GNMDS axis 1 (Fig. 56) and did not support further division of coastal plains. 35 Parallel RDA analyses with soil-class factor variables and relevant continuous analytic 36 variables as alternative ways of representing variation in dominant soil properties showed that 37 the analytic variables Kmar_a and Kbf_a explained more variation in the KS457 data subset 38 (0.0600 and 0.0597) than the factor variables JK1 and JK2 (0.0429 and 0.0344, respectively). 39 We therefore used the analytic variables to represent the CLG candidate soil/sediment class. 40 Conclusion. Based on ordination analyses and interpretation, we identified the 41 following 9 CLG candidates in the three functional variables categories G, U and A 42 ( Fjords and other coastal landscapes (KF) 25

26
A total of 353 OUs were affiliated with the major-type candidate fjords and other coastal 27 landscapes (KS). The three axes in the 3-dimensional GNMDS ordination of the KF353 data 28 subset were confirmed by DCA axes, with three strong axes in both ordinations (i.e. relative 29 length > 0.792× that of the longest axis; Table 19). Initially, however, the 4-dimensional 30 GNMDS was used as a basis for further analyses and graphical interpretation. Because the 31 first three axes in 4-dimensional GNMDS are almost identical to the corresponding axes in 3-32 dimensional GNMDS, the figures and interpretation were not re-made for three dimensions. 33 The GNMDS ordination had Procrustes SS = 0.0003 and no unstable OUs. Visible artefacts 34 could neither be observed in the spaces spanned by DCA ( Fig. 57) nor by GNMDS ( Fig. 58) 35 ordination axes 1 and 2, but the OUs were more spread out towards the periphery in both. 36 Correlations between key and analytic variables on one hand (Table 20), and GNMDS 37 axis 1 scores on the other, supported by vector diagrams (Fig. 58) and isoline diagrams (Figs 38 59-60), revealed that the first axis was strongly related to land-use intensity (infrastructure 39 and agricultural land use; Figs 59 and 60, respectively). Few cities ( Fig. 59; OI = 5 & 6) were 40 located in coastal fjords and there was hardly any difference in location along axis 1 between 41 OI = 4, OI = 5 and OI = 6. OUs with moderate land-use intensity (OI = 3; Ifl between 6 and 42 12) were spread along the entire GNMDS1 axis. Agricultural land-use intensity was clearly 43 related to infrastructure; the isolines for these variables followed the same pattern (Figs 59-44 60). 45 The second GNMDS axis was strongly related to proportion of terrestrial area 46 (Land_a), presence of river networks (R_net_a) and partially also to terrain shape (Table 20;  47 Figs 58 and 63). The relatively short vector arrows for terrestrial area (Land_a) and river 48 networks (R_net_a), running parallel to GNMDS axis 2, indicated relatively weak co-1 variation with other analysis variables. 2 GNMDS axis 3 was clearly related to relief (Table 20; Fig. 64). Figs 58 and 62 show 3 that amount of infrastructure (IfIu) and agricultural land-use intensity (JI) on one hand, and 4 the relief-related variables (DNI, RR1 and TPI1) on the other hand, were independent CLGs; 5 the vector arrows were more or less perpendicular to each other. Fig. 65 show that narrow and 6 steep-sided fjords (DN = 3-4) differed from other OUs. The relatively low number of OUs (n 7 = 21) with high abundance of mire (MP = 2; Fig. 66) indicate that amount and cover of mires 8 are less important properties in fjords. Number of lakes (Inns_s) and mires (Mire_a) were 9 correlated with both of GNMDS axes 2 and 3, and strongly correlated with each other (τ = 10 0.4390). The weak signals from the variable 'island size' neither support a separation of the 11 mainland from islands in fjords nor suggests that gradients from outer to inner coast are 12 important in the fjord major type. 13 Conclusion. Based on ordination analyses and interpretation, we identified the 14 following 9 CLG candidates in the three functional variables categories G, U and A 15 ( Arover_a ( A total of 118 OUs were affiliated with the inland fine-sediment plains (IF) major-type 1 candidate. The two axes in the 2-dimensional GNMDS ordination of the IF118 data subset 2 were confirmed by the corresponding DCA axes (Table 21). A possible but weak tongue-3 effect towards the right could be observed in the DCA ordination diagram for axes 1 and 2 4 (Fig. 68). 5 The gradient length of GNMDS axis 2 was 0.741× the length of the longest axis; Table  6 21). The GNMDS ordination had Procrustes SS = 0.0002 and no unstable OUs. The OUs 7 were relatively evenly spread out in the space spanned by GNMDS ordination axes 1 and 2, 8 with OUs more sparsely spread out towards the periphery (Fig. 69). 9 The GNMDS-ordination (Fig. 69) show a clear division between with glaciofluvial 10 (lower left) and marine (upper right) deposits. Correlations between key and analytic variables 11 on one hand, and GNMDS axis 1 and 2 on the other hand, were in general weaker than for the 12 other major-type candidates, but Table 22 show that the first axis was strongly related to 13 infrastructure (Table 22; Fig. 70). Variation in infrastructure was more noticeable within OUs 14 on marine sediments. The isoline diagram and distribution of OUs in Fig. 71 show that high 15 agricultural land-use intensity (JI = 2) was a common property of all OUs affiliated with this 16 major-type candidate. Table 22 and Figs 69 and 72 show that mire abundance was negatively 17 related to the amount of infrastructure. 18 The continuous variables exposed bedrock (Kbf_a) and glaciofluvial deposits (Kelv_a) 19 were more strongly related to the GNMDS axes than the factor variables for soil types (JK1 20 and KJ2). Thus, we conclude that these mutually correlated variables (Kbf and Kelv; τ = -21 0.4082) are better suited for subsequent analyses with the aim of identifying CLG candidates. 22 However, this does not prevent use of dominance of marine vs glacial sediments in the 23 subsequent operationalisation process. 24 Conclusion. Based on ordination analyses and interpretation, we identified the 25 following 9 CLG candidates in the three functional variables categories G, U and A 26 ( Comment: Bplu_a and Brich_a are treated as separate CLG candidates because the 4 variables are not correlated with each other (τ = 0.0730). 5 6 7 Inland plains without dominance of fine sediments (IX) 8 9 A total of 399 OUs were affiliated with the 'other inland plains' (IX) major-type candidate. 10 The two axes in the two-dimensional GNMDS ordination of the IX399 data subset were 11 confirmed by the corresponding DCA axes ( the in the ordination space spanned by GNMDS ordination axes 1 and 2 (Fig. 74). A tendency 14 for reduced OU density towards higher scores along both axes was observed. Relative 15 gradient lengths of GNMDS and DCA axes, respectively, were in the same range (0.79-0.85; 16 cf. Table 23). 17 Correlations between key and analytic variables on one hand, and GNMDS axis 1 18 scores on the other (Table 24), supported by vector diagrams (Fig. 74) and isoline diagrams 19 (Figs 75-78) revealed a first axis that was very strongly related to a gradient in vegetation 20 cover from boreal to alpine OUs ( Fig. 77), with strong relationships also to increasing land-21 use intensity (infrastructure and agricultural land use; Figs 75 and 76, respectively). However, 22 the variation in the amount of infrastructure was generally low within this major-type 23 candidate; only 6 OUs had above intermediate infrastructure ( Fig. 75; OI > 4). The same 24 pattern applied to agricultural land-use intensity; only 39 OUs had high agricultural land-use 25 intensity. These OUs made up a cluster at the right-hand (i.e. high-score) end of GNMDS 1 26 ( Fig. 76; JI = 3). Very few OUs belonging this major-type candidate had large proportional 27 cover of marine deposits (Fig. 78), while OUs with glaciofluvial deposits (Fig. 79) were 28 evenly distributed throughout the ordination space. The few OUs with more than 10% marine 29 deposits were located in the periphery of the ordination space (high GNMDS 1 scores; Fig.  30 78), indicating a transition towards inland fine-sediment plains. 31 We interpret GNMDS axis 2 as related to variation from coastal to inland geo-32 ecological properties, as expressed by the variables distance to coast, inverse island size, 33 bedrock and number of lakes (Table 24). This gradient was also related to the gradient from 34 boreal to alpine landscapes revealed by GNMDS axis 1. 35 Conclusion. Results of ordination and correlation analyses support the separation of 36 fine-sediment plains (IF) from other inland plains (IX). Based on ordination analyses and 37 interpretation, we identified the following 6 CLG candidates in the three functional variables 38 categories G, U and A (note that most CLG candidates have a weak definition basis) for the 39 IX major-type candidate: which are in turn strongly correlated (|τ| > 0.4874) with each other. Coniferous forest and 8 boreal heaths (Arbar_a and Bohei_a), which are weakly negatively correlated with each other, 9 make up the extremes of this group. This group is therefore regarded as one CLG candidate. 10 Boreal heath (Bohei_a), which has the strongest correlation with GNMDS axis 2 among 11 potential 'members' in a group for open areas, |τ| = 0.3610, was left out because more than 8 12 variables were at least quite strongly correlated (|τ| > 0.4) with this axis. 13 14 15 Inland valleys (ID) 16 17 A total of 1012 OUs were affiliated with the inland valleys (ID) major-type candidate. The 18 three axes of the 3-dimensional GNMDS ordination of the ID1012 data subset were 19 confirmed by the corresponding DCA axes (Table 25). Gradient lengths of the second and 20 third GNMDS axes were 0.760× and 0.664× the length of GNMDS axis 1, respectively. The 21 GNMDS ordination had Procrustes SS = 0.1796 and one unstable OU. No visible artefacts 22 could be seen in the ordination diagrams (Figs 80 and 81), but a tendency for reduced density 23 towards the periphery of the ordination space was observed. Gradient lengths of the first axes 24 of both ordinations indicate existence of a strong major gradient in the ID1012 data subset 25 (Table 25). Correlations between key and analytic variables on one hand, and GNMDS axis 1 26 scores on the other (Table 26) GNMDS axis 2 was clearly related to terrain form; three key variables and seven 36 analytical variables that expressed terrain variables had correlation coefficients |τ| > 0.4 with 37 this axis (Table 26; Fig. 81). The vector diagram (Fig. 81) shows that the material can be 38 arranged in two very distinct variable groups; (i) the gradient from above to below the treeline 39 (i.e. from mountains to mainly forested lowlands), along which the amount of infrastructure 40 and agricultural land-use intensity increased; and (ii) relative relief and other terrain-form 41 variables. OUs with presence of glacier ( Fig. 86; BA = 2) occurred in the lower left corner in 42 the space spanned by GNMDS ordination axes 1 and 2, associated with high relief. OUs with 43 glacier did, however, not form a separate group but occurred intermixed with high-relief OUs 44 without glaciers. Thus, glacier presence apparently varies within the 'high-relief end' of the 45 relief gradient. 46 GNMDS axis 3 was strongly correlated with variables that expressed presence and 47 areal cover of freshwater lakes (Table 26; Figs 87-88) although OUs characterised by lakes 48 did not segregate clearly from other OUs along GNMDS axis 3 (Fig. 88). 49 Since the proportions of areas of both glacial deposits and bare rock were correlated 1 with GNMDS axis 1, we plotted the distribution of the 1018 OUs on primary soil classes JK1. 2 A clear, but not strong, relationship between soil category (JK1) and positions in the 3 ordination diagram can be seen in Fig. 89. The distribution of soil types was related to relief; 4 areas with thick layers of till are associated with low relative relief, while bare rock and 5 landslide soils are correlated with high relative relief. The only soil type that was clearly 6 associated with the CLG from mountain to lowland was, not unexpectedly, the abundance of 7 marine sediments (11 OUs in the lower right corner of the ordination diagram). 8 Based on ordination analysis and interpretation, we identified the following 8 CLG 9 candidates in the three functional variable categories G, U and A (note that most CLG 10 candidates have a weak definition basis): Comments: The hypsographic index (Er_m) included in CLG-candidate G2 was 36 correlated with GNMDS axis 3 as well as with lake variables and was assigned to this CLG 37 candidate despite the pairwise correlations were weak [τ (Er_m, Lake_a) = -0.5129]. The 38 amount of freshwater lakes and abundance of mires (G2 and G4, respectively) make up two 39 separate CLG candidates as evident from the lack of a relationship: τ (Dismire, Lake_a) = 40 0.0678. Soil, expressed as a factor-type variable (G3), was included among hCLG candidates 41 with classes for dominance of bare rock (B) and of glacial sediments (E). These two dominant 42 soil types were well represented in the data material and the centroids of both were placed 43 clearly off the origin in the ordination diagram (Fig. 89 Inland hills and mountains (IA) 49 A total of 1618 OUs were affiliated with the inland valleys (IA) major-type candidate. The 1 three axes in the 3-dimensional GNMDS ordination of the IA1618 data subset were confirmed 2 by corresponding DCA axes (Table 27). The GNMDS ordination had Procrustes SS = 0.3208 3 and no unstable OU. No visible artefacts could be seen in the ordination diagrams (Figs 90  4 and 91), but a tendency for reduced density towards the periphery of the ordination space was 5 observed. The cloud of points in the DCA ordination diagram (axes 1 and 2) had a distinctive 6 triangular structure. Gradient lengths of the first axes of GNMDS and DCA ordinations 7 indicated existence of a strong first axis (Table 27), while the relative length of both of 8 GNMDS axes 2 and 3 were > 0.8× the length of GNMDS axis 1. 9 Correlations between key and analytic variables on one hand, and GNMDS axis 1 10 scores on the other (Table 28), supported by vector diagrams (Fig. 91) and isoline diagrams 11 , revealed a first axis that was very strongly related to a bio-and geo-ecological 12 gradient in landscape element composition from mountains to lowlands, that also included 13 variation in human land-use intensity. Along this composite gradient, the relative relief of the 14 terrain decreased, the fractional area covered by forest and mire increased and the amount of 15 infrastructure (including agricultural land-use intensity) increased sharply (Fig. 91). OUs with 16 medium and high agricultural land-use intensity form a clearly differentiated group in the 17 lower right corner of the ordination diagram ( Fig. 93; JP > 1). The amount of infrastructure 18 followed the same pattern but in a more gradual manner, increasing from the upper left 19 (mountains) to the lower right (lowlands) in the diagram of GNMDS-axes 1 and 2 ( Fig. 92). 20 Vegetation cover followed the same pattern: OUs from alpine areas (BA = 3, Alp_a > 0.75), 21 with high values for exposed bedrock (Kbf_a) and open treeless areas (Sn_flekk) prevailed at 22 low GNMDS axis 1 values while coniferous (Arbar_a) and mixed (Arbla_a) boreal forests 23 (BA = 1) prevailed at the high-score end of this GNMDS axis (Table 28; Figs 91 and 94). 24 Isolines for the proportion of lime-rich bedrock (Brich_a; Fig. 98) show a similar, but weaker 25 pattern. OUs with glacier ( Fig. 95; BP = 2) were restricted to the low-score end of GNMDS 26 axis 1, where they occurred intermixed with OUs with glacier. Unlike in inland valleys, OUs 27 with glacier occurred in inland hills and mountains without any relationship to terrain 28 variation (i.e. along GNMDS axis 2). 29 GNMDS axis 2 was clearly related to (residual variation in) relief. Variation in the 30 abundance of mires followed the variation in relief (Fig. 96). OUs with distinct peaks (Fig 97;  31 TP = 2) to some extent followed the same pattern as glaciers and occurred intermixed with 32 other OUs with high relief at low GNMDS axis 2 (and GNMDS axis 1) scores. 33 GNMDS axis 3 was correlated with few analysis variables, and only weakly so. No 34 clear, easily interpretable pattern was observed along this axis, which identified six OUs with 35 outlier positions in the upper left corner of the ordination diagram (low scores for GNMDS 36 axis 1 and high scores for GNMDS axis 3). 37 Based on ordination analyses and interpretation, we identified the following 8 CLG Arover_a ( The analyses for identification of independent significant analytic variables for each CLG 31 candidate by forward selection from the n = 84 variables using RDA show that all nine CLG 32 candidates satisfied the requirement for explaining at least 6% of the variation ( The three variables included in the ONV that represents CLG 1, inner-outer coast (IYK), 41 were: (1) number of rivers (R_net_a); (2) areal coverage of exposed coast (Ekspve_a); and (3) 42 inverse island size (Oyst_i). This geo-ecological CLG had an estimated gradient length of 43 0.4026/0.08 = 5.033 EDU-L (ecodiversity distance units in landscapes) and was, accordingly, 44 divided into 5 standard segments (Table 30). The distribution of OUs along this ONV (in its 45 original scaling) was left-skewed ( Fig. 101). 46 This CLG expresses variation in coastal landscapes from areas with 'inland properties' 47 on the inner side of larger islands, hardly exposed to the harsh conditions of the open sea 48 (protected inner coast), to the outer coast, directly exposed to the actions of wind, waves and 49 ocean currents, i.e. strongly wave-exposed outer coast (Figs 102-104). The intermediate 1 segments are termed: moderately protected coast, moderately wave-exposed coast; and wave-2 exposed outer coast. 3 Differences in landscape-element composition among OUs from the outer coast 4 (relatively small islands, SN-classes 3-5; cf. Fig. 102, Table 30), which was represented in the 5 data subset by relatively few OUs, contributed strongly to gradient length. The gradient length 6 estimate is therefore uncertain. 7 A closer examination of variation along the rescaled ONV for this CLG (Fig. 102) 8 motivated for using the following logarithmic sequence of (largest) island sizes as criteria for 9 separating standard segments 4 and 5, 3 and 4, 2 and 3, respectively: 1 km²; 10 km 2 ; 100 km². 10 The border between standard segments 1 and 2 was less distinct in terms of island vs. 11 mainland properties. 12 Infrastructure and agricultural land use followed CLG 1 IYK closely ( The two variables included in the ONV that represents CLG 2, relief (RE), were: (1) mean 22 terrain ruggedness (Rug3_m) and (2) areal coverage of convex terrain (Tpi1h_a). This geo-23 ecological CLG had an estimated gradient length of 0.3003/0.08 = 3.754 EDU-L units and 24 was, accordingly, divided into 3 standard segments (Table 31). The distribution of OUs along 25 this ONV (in its original scaling) was slightly left-skewed (Fig. 105). 26 This CLG expresses variation in terrain-form within coastal plains from (relatively) 27 flat coastal plains (via moderately rugged coastal plains) to rugged coastal plains, often with 28 cliffs or remnant peaks. The CLG had good linearity, with an even distribution of OUs along 29 the CLG (Figs 106-107). A slight increase in PD between adjacent initial CLG segments in 30 both directions from the middle of the candidate CLG indicates decreasing degree of 31 generalisation towards gradient endpoints ( The three variables included in the ONV that represents CLG 3, abundance of wetlands (VP), 42 were: (1) mean distance to mire (Dismire); (2) number of lakes (Inns_s); and (3) mean 43 distance to lake (Dislake). This geo-ecological CLG had an estimated gradient length of 44 0.1794/0.08 = 2.243 EDU-L units and was, accordingly, divided into 2 standard segments 45 (Table 32). The distribution of OUs along this ONV (in its original scaling) was right-skewed 46 (Fig. 108). 47 This CLG expresses variation in the areal cover of wetlands (including mires) and the 48 abundance of small lakes and tarns, which are often associated with wetlands. The two 49 segments are characterised by low to medium abundance (VP•1) and high abundance (VP•2), 1 respectively, of wetlands and associated ecosystems (Figs 109-110). 2 We identified some nonlinearity in initial segments 6-7 (cf. Table 32). The increase in 3 PD between adjacent initial CLG segments from mid-gradient towards the gradient's upper 4 end (i.e. with increasing abundance of lakes and wetlands) indicates decreasing 5 generalisation. This suggests that there is some variation in lake occurrence also on small 6 islands ( Fig. 108), while mires are more or less completely confined to larger islands and the 7 mainland (Fig. 109). The variation along this CLG is most clearly expressed by the number of 8 lakes, as indicated by isolines for the Inns_s variable in Fig. 109. 9 10 CLG 4 -(U1) Forest cover: SkP (Table 33, Figs 111-112) 11 12 The two variables included in the ONV that represents CLG 4, forest cover (SkP), were: (1)  13 areal coverage of deciduous forest (Arlov) and (2) areal coverage of mixed forest (Arbla_a). 14 This bio-ecological CLG had an estimated gradient length of 0.1726/0.08 = 2.158 EDU-L 15 units and was, accordingly, divided into 2 standard segments (Table 33). The distribution of 16 OUs along this ONV (in its original scaling) was unimodal and symmetric ( Fig. 111). 17 This CLG expresses variation in forest cover from bare rock without or with sparse 18 forest cover, e.g. or heaths (SkP•1) to forested areas (SkP•2) ( Fig. 112). 19 The increase in PD between adjacent initial CLG segments from mid-gradient towards 20 the lower end of the gradient (open landscapes) indicates decreasing degree of generalisation 21 towards this gradient endpoint ( The two variables included in the ONV that represents CLG 5, amount of infrastructure (OI), 28 were: (1) the infrastructure index (IfI) and (2) areal coverage of built-up areas (Gab_a). This 29 land-use related CLG had an estimated gradient length of 0.1344/0.08 = 1.68 EDU-L units 30 and was, accordingly, not divided into standard segments (Table 34). The distribution of OUs 31 along this ONV (in its original scaling) was unimodal and symmetric. 32 This CLG expresses variation in human impact as reflected in man-made 33 infrastructure, expressed by the abundances of buildings, roads and other visible traces of 34 human activity. Low PD values show that there is little residual variation attributable to 35 infrastructure left after extraction of CLGs 1-4. The weak foundation for OI as an 36 independent CLG in coastal plains motivated for not dividing this CLG into standard 37 segments. 38 39 Tentative division of coastal plains into minor types 40 41 Correlations coefficients between ONVs and ordination axes, primary key variables and 42 analysis variables were calculated as a basis for operationalisation of the CLGs (Table 35). 43 The theoretical number of gradient combinations (tentative minor types) is 5 × 3 × 2 × 2 (× 1) 44 = 60. Of these, 51 were represented in the total data set ( Coastal fjords (KF) 2 3 The analyses for identification of independent significant analytic variables for each CLG 4 candidate by forward selection from the n = 84 variables using RDA show that seven of the 5 eight CLG candidates satisfied the requirement for explaining at least 6% of the variation 6 ( Table 37). The parsimonious set of CLGs, found by RDA with forward selection among the 7 CLG candidates separately for each functional variable category, contained 4 CLGs (Table  8 37). The properties of the four CLGs and their associated (ranked primary) ONVs are 9 described below on the basis of Tables 37-43 and Figs 114-122.  10  11 CLG 1 -(G1) Relief in fjords: RE ( The four variables included in the ONV that represents CLG 1, relief (RE), were: (1)  The two variables included in the ONV that represents CLG 2, abundance of wetlands (VP) 30 were: (1) areal coverage of mire (Mire_a) and (2) number of lakes (Inns_s). This geo-31 ecological CLG had an estimated gradient length of 0.1860/0.08 = 2.325 EDU-L units and 32 was, accordingly, divided into 2 standard segments (Table 39). The distribution of OUs along 33 this ONV (in its original scaling) was right-skewed ( Fig. 116). 34 This CLG expresses variation in the fraction of an OU's area that was covered by 35 wetland systems (predominantly mires) and in the abundance of small lakes and tarns which 36 are often associated with wetlands: from low to medium abundance to high abundance of 37 mires and tarns . Variation along this CLG was found for OUs with low 38 relative relief (Figs 117-118). Isolines for number of lakes (Inns_s) ran almost exactly 39 parallel with the CLG (Fig. 118). The CLG VP in fjords also captured variation in vegetation 40 cover; this CLG was more strongly correlated with the fractional cover of boreal heaths 41 ( The four variables included in the ONV that represents CLG 3, open area cover (AP), were 48 the fractions of OU area covered by: (1) boreal heath (Bohei_a); (2) exposed bedrock 49 (Kbf_a); (3) open areas (Araaf_a); and (4) 'impediment' (i.e. areas not suitable for agriculture 50 or forestry; Sn_imp), respectively. This bio-ecological CLG had an estimated gradient length 1 of 0.2036/0.08 = 2.545 EDU-L units and was, accordingly, divided into 2 standard segments 2 (Table 40). The distribution of OUs along this ONV (in its original scaling) was slightly left-3 skewed (Fig. 119). 4 This CLG (Fig. 120)  sparse forest cover. The linearity of this CLG was generally low (Table 40). CLG AP 7 separated a relatively small group of low-score OUs from the majority of OUs which were 8 placed in segment 2. Only three variables had correlations coefficients |τ| > 0.3 with CLG AP 9 (Table 42): areal coverage of impediment (Sn_imp; τ = 0.5276), areal coverage of boreal 10 heaths (Bohei_a; τ = 0.3190) and relative abundance of boreal vs. alpine landscapes (AlpA; τ 11 = 0.3728). The CLG AP explained variation in vegetation cover not accounted for by CLG 1, 12 RE. Based on the weak correlation between primary key variable AlpA and this CLG, we 13 regard the foundation for AP as an independent CLG in fjords as relatively weak. 14 15 CLG 4 -(A1) Amount of infrastructure: OI ( The two variables included in the ONV that represents CLG 4, amount of infrastructure (OI), 18 were: (1) the amount of infrastructure (IfIu) and (2) areal coverage of built-up areas (Build_a 19 and Gab_a). This land-use related CLG had an estimated gradient length of 0.2054/0.08 = 20 2.567 EDU-L units and was, accordingly, divided into 2 standard segments (Table 40). The 21 distribution of OUs along this ONV (in its original scaling) was right skewed (Fig. 121). 22 This CLG expresses variation in the abundance of buildings and other infrastructure 23 from low and intermediate to settlement (village, small town or city; Fig. 122). The CLG has 24 good linearity with decreasing degree of generalisation towards both gradient ends ( Correlations between ONVs and ordination axes, primary key variables, and analysis 38 variables were calculated as a basis for operationalisation of the CLGs (Table 42). The 39 theoretical number of gradient combinations (tentative minor types) is 4 × 2 × 2 × 2 = 32. Of 40 these, 31 were represented in the total data set ( Inland fine-sediment plains (IS) 47 48 The analyses for identification of independent significant analytic variables for each CLG 49 candidate by forward selection from the n = 84 variables using RDA (Table 44) show that five 50 out of nine CLG candidates satisfied the relaxed criterion for accepting CLG candidates set to 1 8% of the variation (as opposed to 6% for the other major types). Use of a relaxed criterion 2 was motivated by the small data set (118 OUs). For the RDA analysis with forward selection 3 of variables within each functional variable category, undertaken to find a parsimonious set of 4 CLGs, the CLG candidate soil type (JA) was subjectively chosen as CLG 1. The reason for 5 this was (1) the sharp separation of the two dominant soil type classes in the ordination 6 diagram, and (2) that the variation explained by relief (RE) was inflated by the large number 7 of intercorrelated topographic variables in the data set. The parsimonious set of CLGs 8 contained 4 CLGs, while abundance of mire (MP) was very close to fulfil the requirement. 9 The properties of the four CLGs and their associated (ranked primary) ONVs are described 10 below on the basis of Tables  The two variables included in the ONV that represents CLG 1, soil type (JA), were the areal 15 coverage of: (1) exposed bedrock (Kbf_a) and (2)  The two variables included in CLG 2, relief (RE), were: (1) relief (RR1) and (2) areal  29 coverage of convex terrain (Tpi1h_a). This geo-ecological CLG had an estimated gradient 30 length of 0.1543/0.08 = 1.929 EDU-L units and was, accordingly, not divided into standard 31 segments (Table 38). The distribution of OUs along this ONV (in its original scaling) was 32 right-skewed (Fig. 124). 33 While this ONV expresses variation associated with relief within inland fine-sediment 34 plains, the total variation in landscape characteristics was insufficient (as estimated by 35 Method III) to motivate for a division of this CLG into two segments. 36 37 CLG 3 -(U1) Open area cover: AP ( The single variable included in the ONV that represents CLG 4, amount of infrastructure (OI), 8 was amount of infrastructure (IfIu). This land-use related CLG had an estimated gradient 9 length of 0.1176/0.08 = 1.47 EDU-L units and was, accordingly, not divided into standard 10 segments (48). The distribution of OUs along this ONV (in its original scaling) was slightly 11 right-skewed (Fig. 128). 12 This CLG expresses variation in amount of buildings and other infrastructure in inland 13 fine-sediment plains, but the total variation in landscape characteristics was insufficient (as 14 estimated by Method III) to motivate for a division of this CLG into segments. 15 16 Tentative division of inland fine-sediment plains into minor types 17 18 Correlations between ONVs and ordination axes, primary key variables, and analysis 19 variables were calculated as a basis for operationalisation of the CLGs (Table 49). The 20 theoretical number of combinations of standard segments is basically only 3 because only one 21 CLG reached the threshold for segmentation of 2 EDU-L units. However, the clear separation 22 of the two primary soil categories along CLG 1 supported recognition of 6 gradient 23 combinations based upon division of each AP segment into two (sub)segments by dominant 24 soil type (JA). All 6 combinations of AP interval and soil type were represented in the IS118 25 data subset (Table 50). 26 27 28 Inland plains without dominance of fine sediments ('other inland plains'; IX) 29 30 The analyses for identification of independent significant analytic variables for each CLG 31 candidate by forward selection from the n = 65 variables using RDA show that five of the six 32 CLG candidates satisfied the requirement for explaining at least 6% of the variation (Table  33 51). The parsimonious set of CLGs, found by RDA with forward selection among the CLG 34 candidates separately for each functional variable category, contained 4 CLGs ( The two variables included in the ONV that represents CLG 1, distance to coast (KA), were: 41 (1) inverse island size (Oyst_i); and (2) distance to coast (Discoast). This geo-ecological CLG 42 had an estimated gradient length of 0.2959/0.08 = 3.699 EDU-L units and was, accordingly, 43 divided into 3 standard segments (Table 52). The distribution of OUs along this ONV (in its 44 original scaling) was strongly left-skewed (Fig. 129). 45 This CLG separated a 'coastal group' with 23 OUs (segment 1; values before rescaling 46 < 0.52) from an inland group (the rest). The variation along KA was strongly non-linear, 47 indicating presence of little systematic variation. Accordingly, we interpret this CLG as an 48 indication that inland plains close to the coast with presence of marine deposits, should be 49 kept separate from other inland plains (Fig. 130). Several of the OUs from inland plains close 50 to the coast were located on islands, but lacking coastline. Except for the correlation with 1 marine deposits (Table 56; Kmar_a; τ = -0.3076), this CLG had no strong correlations with 2 other analysis variables than those already included in the CLG. While Table 52 shows the 3 tentative division of the CLG into three segments suggested by the analyses, the interpretation 4 motivates for a division into two segments only. A critical assessment of the basis for 5 segmentation of this CLG is required prior to practical implementation in a landscape-type 6 system. 7 8 CLG 2 -(G2) Freshwater lake properties: IP (Table 53, Figs 131-132) 9 10 The single variable that represented CLG 2, freshwater lake properties (IP), was number of 11 lakes (Inns_s). CLG 2 had an estimated gradient length of 0.3218/0.08 = 4.022 EDU-L units 12 and was, accordingly, divided into 4 standard segments (Table 53). The distribution of OUs 13 along this ONV (in its original scaling) was slightly left-skewed, with few OUs representing 14 the high-score end of the rONV (Fig. 131). 15 This CLG expresses variation within inland plains in the abundance of small lakes and 16 tarns (which are often associated with wetland systems); from low or medium abundance to 17 high abundance. This CLG was negatively correlated with mean distance to lake (Dislake; τ = 18 -0.5349) and mixed boreal forest (Arbla_a; τ = -0.3062), and positively correlated with areal 19 coverage of freshwater lakes (Lake_a; τ = 0.4261), number of freshwater lake islands 20 (Innoy_s; τ = 0.3329) and areal cover of dry heath/open areas (Sn_torr; τ = 0.3541). All of 21 these variables represent landscape properties typical of mountain plateaus (Norwegian: 22 'fjellvidde') with a high abundance of small lakes. 23 While the gradient-length estimate (Table 53) suggests a tentative division of the CLG 24 into four segments, we question the basis for such a fine division. Most likely, some variation 25 in infrastructure, perhaps also other characteristics, which are correlated with lake abundance, 26 contribute to the large estimated gradient length. It is first and foremost the presence of many 27 small lakes that define the 'end' of this CLG. 28 29 CLG 3 -(U1) Forest cover: SkP (Table 54, Figs 133-134)  30  31 The three variables included in the ONV that represents CLG 3, forest cover (SkP), were the 32 areal coverage of: (1) coniferous forest (Arbar_a); (2) open areas (Araaf_a); and (3) open  33 areas with dry heath (Sn_torr). CLG 3, SkP, had an estimated gradient length of 0.4619/0.08 = 34 5.774 EDU-L units and was, accordingly, divided into 5 standard segments (Table 54). The 35 distribution of OUs along this ONV (in its original scaling) was slightly right-skewed ( Fig.  36 133). 37 This CLG was strongly correlated with the key variable amount of alpine areas 38 (Alp_a; τ = 0.5538) and six analytical variables related to vegetation cover (Table 56). 39 Accordingly, this CLG expresses variation in vegetation cover within inland plains, from 40 barren mountain plains without or with sparse vegetation cover to forested or potentially 41 forested plains below the climatic forest line (Fig. 134) four land-use and building abundance variables (Table 56). 47 Although the gradient-length estimate (Table 54) suggests a tentative division of the 48 CLG into five segments, the basis for such a fine division is questionable. 49 50 CLG 4 -(A1) Amount of infrastructure: OI (Table 55, Figs 135-137)  1  2 The two variables included in the ONV that represents CLG 4, amount of infrastructure (OI), 3 were: (1) the amount of infrastructure (IfIu) and (2) areal coverage of built-up areas (Gab_a). 4 This land-use related CLG had an estimated gradient length of 0.3475/0.08 = 4.344 EDU-L 5 units and was, accordingly, divided into 4 standard segments (Table 55). The distribution of 6 OUs along this ONV (in its original scaling) was right-skewed and leptokurtic (Fig. 135). 7 This CLG separates the few OUs in other inland plains with a high amount of 8 buildings and other infrastructure from other OUs. Most of the variation related to density of 9 infrastructure occurred in the lowlands (Fig. 136). Variables that express variation in 10 agricultural land-use intensity were relatively weakly correlated with this CLG (Table 56; τ < 11 0.3), but seemed to follow the variation in the amount of infrastructure (Fig. 137). 12 Although the gradient-length estimate ( Correlations between ONVs and ordination axes, primary key variables, and analysis 21 variables were calculated as a basis for operationalisation of the CLGs (Table 56). The 22 theoretical number of combinations of intervals along gradients is 3 × 4 × 5 × 4 = 240. Of 23 these, 75 were represented in the total data set ( Inland valleys (ID) 32 33 The analyses for identification of independent significant analytic variables for each CLG 34 candidate by forward selection from the n = 67 variables using RDA show that six of the eight 35 CLG candidates satisfied the requirement for explaining at least 6% of the variation. The 36 parsimonious set of CLGs, found by RDA with forward selection among the CLG candidates 37 separately for each functional variable category, contained 4 CLGs ( The three variables included in the ONV that represents CLG 1, relief (RE), were: (1) relief 44 (Rr1_m); (2) mean terrain ruggedness (Rug3_m); and (3) areal coverage of moderate slope 45 (Meant_a). This geo-ecological CLG 1, relief (RE), had an estimated gradient length of 46 0.5313/0.08 = 6.641 EDU-L (ecodiversity distance in landscapes) units and was, accordingly, 47 divided into 6 standard segments (Table 59). The distribution of OUs along this ONV (in its 48 original scaling) was almost symmetric (Fig. 138). 49 This CLG expresses variation in valley shape, as expressed by the depth/width ratio of 1 the valley relative to its surroundings; from wide valleys via open valleys to narrow and steep-2 sided valleys. The CLG has good linearity, but the degree of generalisation is reduced towards 3 the low-score end of the gradient, i.e. segments 1 and 2 (of 6), corresponding to relative relief 4 RR < 100 m. 5 Segments 1 and 2 contained relatively few OUs compared to the other segments ( Fig.  6 139), suggesting that these OUs are atypical for the major type. This opens for the possibility 7 that these OUs are transitional to the major types inland plains and inland mountains. They 8 may, however, also result from inconsistencies or uncertainties in the OU delineation process. 9 Segments 5 and 6 are characterised by the same threshold values as landscape-gradient 10 segments DN2/DN3 and DN3/DN4, respectively, in the 'pilot Nordland project' (see 11 Appendix S1). 12 This CLG apparently captures all variation related to valley form and relief in the 13 major type. Three key variables and 6 analysis variables had correlations with |τ| > 0.6 with 14 the sONV. The strong signal from RE was also correlated with variation in areal coverage of: 15 moderate slope (from flat to steep terrain; τ = -0.5420), landslide soil (τ = 0.5077), exposed 16 bedrock (Kbf_a; τ = 0.3452) and mires (Mire_a; τ = -0. 3164). The two variables included in the ONV that represents CLG 2, freshwater lake properties (IP), 23 were: (1) mean distance to lake (Dislake) and (2) abundance of freshwater lakes (II). This 24 CLG had an estimated gradient length of 0.3784/0.08 = 4.73 EDU-L units and was, 25 accordingly, divided into 4 standard segments (Table 60). The distribution of OUs along this 26 ONV (in its original scaling) was left-skewed (Fig. 140). The CLG has a good linearity and 27 reduced degree of generalization towards gradient endpoints. 28 This CLG expresses variation within valleys from landscapes without lakes or with 29 small lakes to valley landscapes with large lakes; the typical 'inland fjords'. Four analytical 30 variables related to hydrology were strongly (|τ| > 0.3) correlated with this CLG ( The two variables included in the ONV that represents CLG 3, forest cover (SkP), were: (1)  40 areal coverage of impediment (i.e. areas not suitable for agriculture or forestry; Sn_imp) and 41 (2) areal coverage of coniferous forest (Arbar_a). This bio-ecological CLG had an estimated 42 gradient length of 0.3285/0.08 = 4.106 EDU-L units and was, accordingly, divided into 4 43 standard segments (Table 61). The distribution of OUs along this ONV (in its original scaling) 44 was bimodal and left-skewed (Fig. 142). The linearity was low. The noteworthy 'gap' in 45 landscape-property composition in the middle segments, giving rise to the bimodal frequency 46 distribution, was probably due to a clear-cut differentiation between mountains and forested 47 lowlands. 48 Four vegetation variables that express vegetation cover were correlated with this CLG 49 (Table 63) boreal/alpine landscapes (AlpA; τ = -0.5309). In addition, this CLG captured most of the 2 variation in infrastructure and agricultural land-use intensity ( Fig. 144-145), with nine land 3 use variables correlated at |τ| > 0.3 with the CLG (Table 63). Fig. 144 indicates that 4 agricultural land-use intensity was non-linearly related to this CLG. 5 6 CLG 4 -(A3) Amount of infrastructure: OI (Table 62, Figs 146-148) 7 8 The two variables included in the ONV that represents CLG 4, amount of infrastructure, were: 9 (1) the infrastructure index (IfI) and (2) areal coverage of built-up areas (Gab_a). This land-10 use related CLG had an estimated gradient length of 0.2626/0.08 = 3.282 EDU-L units and 11 was, accordingly, divided into 3 standard segments (Table 62). The distribution of OUs along 12 this ONV (in its original scaling) was right-skewed (Fig. 146). Non-linearity occurred towards 13 the high-score end of the gradient; the landscape element composition of segment 8 was more 14 similar to that of segments 6 than with that of segment 7. 15 This CLG expresses residual variation in human land-use not captured by the other 16 gradients and not related to the gradient from lowlands to mountains (CLG 3, SkP). This 17 pattern clearly emerges from the six analytical human land-use related variables with |τ| > 0.3 18 with this CLG (Table 63), which include buildings, built-up areas and other technical 19 infrastructure (roads, power lines), and by the weak correlation of the CLG with the boreal-20 alpine variable (AlpA; τ = -0.0531). 21 22 Tentative division of inland valleys into minor types 23 24 Correlations between ONVs and ordination axes, primary key variables, and analysis 25 variables were calculated as a basis for operationalisation of the CLGs (Table 63). The 26 theoretical number of combinations of intervals along gradients is 6 × 4 × 4 × 3 = 288. Of 27 these, 98 were represented in the total data set (Table 64), 24 of which with > 11 OUs (> 1% 28 of the ID1012 data subset). 29 30 31 Inland hills and mountains (IA) 32 33 The analyses for identification of independent significant analytic variables for each CLG 34 candidate by forward selection from the n = 66 variables using RDA show that five of the 35 eight CLG candidates satisfied the requirement for explaining at least 6% of the variation 36 (Table 65). The parsimonious set of CLGs, found by RDA with forward selection among the 37 CLG candidates separately for each functional variable category, contained 3 CLGs (Table  38 65 The three variables included in the ONV that represents CLG 1, relief (RE), were: (1) relief 44 (RR1); (2) mean terrain ruggedness (Rug3_m); and (3) areal coverage of convex terrain 45 (Tpi1h_a). This geo-ecological CLG had an estimated gradient length of 0.5550/0.08 = 6.938 46 EDU-L units and was, accordingly, divided into 6 standard segments (Table 66). The 47 distribution of OUs along this ONV (in its original scaling) was right-skewed (Fig. 149). 48 This CLG expresses variation in terrain shape within inland hill and mountain 49 landscapes; from depressions in hills and mountains, via undulating inland hills and 50 mountains with low relief to steep and rugged hills and mountains (typical 'alpine 1 landscapes'). The CLG has good linearity, but weaker degree of generalisation at the lower 2 end of the gradient, i.e. segment 1, which is characterised by relative relief RR < 100 m. The 3 relatively few OUs in segment 1 (Fig. 150) may suggest that this segment represents a 4 transition towards the major type inland plains. 5 A total of 11 variables had |τ| > 0.6 with sONV_RE; the strongest correlation was 6 obtained for relief (Rr1_m; τ = 0.8606; Table 69). This was the strongest correlation observed 7 for a combination of a CLG and a variable in our material, showing that relief controls most 8 of the variation in landscape properties in inland hills and mountains. The fundamental role of 9 relief in this major type is further substantiated by the large number of strongly correlated 10 variables, the high total amount of variation captured by this CLG and its large estimated 11 gradient length. Mire abundance (MI; τ = -0.3778) was negatively correlated with this CLG, 12 following the same pattern as relief. Neither the amount of infrastructure nor the gradient 13 from lowland to mountain follow the relief gradient ( The two variables included in the ONV that represents CLG 2, forest cover (SkP), were: (1) 29 areal coverage of impediment (i.e. areas not suitable for agriculture or forestry; Sn_imp) and 30 (2) areal coverage of coniferous forest (Arbar_a). This bio-ecological CLG had an estimated 31 gradient length of 0.2648/0.08 = 3.31 EDU-L units and was, accordingly, divided into 3 32 standard segments (Table 67). The distribution of OUs along this ONV (in its original scaling) 33 was bimodal and right-skewed (Fig. 151 grazing). Fig. 152 indicates that the bimodal distribution of OUs along this CLG is due to 39 differences in landscape-element composition between mountains and other areas. 40 Correlated bio-ecological variables were ( The two variables included in the ONV that represents CLG were: (1) the infrastructure index 1 (IfIu) and (2) areal coverage of built-up areas (Gab_a). This land-use related CLG had an 2 estimated gradient length of 0.3052/0.08 = 3.815 EDU-L units and was, accordingly, divided 3 into 3 standard segments (Table 68). The distribution of OUs along this ONV (in its original 4 scaling) was strongly right-skewed (Fig. 154). 5 This CLG expresses residual variation in the importance of human infrastructure and 6 to some extent also agricultural land-use from low via intermediate to high. The OUs were 7 clustered into three groups, roughly corresponding to the standard segments, on the basis of 8 landscape property composition. The majority of OUs had low amount of infrastructure (Figs  9 155-156). Except for high scores along this CLG for OUs with large values for the 10 infrastructure index (Fig. 156: red symbols for high density of infrastructure, i.e. cities and 11 towns), this CLG does not clearly distinguish between OUs on the basis of explicit analysis 12 variables, or variable combinations. Tentative division of inland hills and mountains into minor types 18 Correlations between ONVs and ordination axes, primary key variables, and analysis 19 variables were calculated as a basis for operationalisation of the CLGs (Table 69). The 20 theoretical number of combinations of intervals along gradients is 6 × 3 × 3 = 54. Of these, 41 21 were represented in the total data set ( The division into coastal and inland landscapes 8 9 The analyses reveal very clear patterns of co-ordinated variation in landscape properties and 10 landscape-element composition in Norway, providing both general knowledge about the 11 distribution of landscape elements throughout the country and an empirical basis for the first 12 version of a landscape-type system. We find that the single most important property 13 explaining differences in the composition of landscape elements and features in Norway is the 14 location of the landscape in relation to the coastline. This finding supports construction of a 15 landscape-type hierarchy according to EcoSyst principles (Halvorsen et al. 2020), which is 16 performed by a top-down process, by separating coastal and inland landscapes at the topmost, 17 level of major-type groups. 18 19 20 Further division by geomorphological criteria 21 22 The statistical analyses of landscape element compositional data provide strong support for a 23 division of the study area into six 'major landscape types' by geomorphological criteria: 24 inland hills and mountains; inland valleys; inland fine-sediment plains; other inland plains; 25 coastal plains; and coastal fjords. These major types can be identified either by surface 26 geometry alone or by surface geometry in combination with soil types (inland plains). Back-27 transformed analysis variable values allow for fine-tuning of criteria for separation between 28 major types based on identified important terrain parameters such as relative relief and 29 proportion of flat terrain. The results of our analyses may thus provide an empirical basis for 30 elaboration of a high-resolution (100×100 m) map of meso-scaled landforms based on 31 consistent criteria. 32 In a Nordic context, the major types identified by us are recognisable in descriptions 33 (and some coarse-scaled maps) of major geomorphological features that date back to the 19th 34 century (see, e.g. Reusch 1894, Reusch 1905, Rudberg 1960, Gjessing 1978, Klemsdal 1982, 35 Erikstad et al. 2009). Nevertheless, the major types identified in our study differ in several 36 ways from earlier Norwegian geomorphological typologies. Thus Rudberg (1960)  types of inland plains, and six types of hills and mountains based on relief and location along 46 an elevation gradient. Most, but not all of the types described by Rudberg (1960) and 47 Mattsson et al. (1984), are identified as either major or minor types in our type-system. While 48 our analyses point to geomorphological features as more fundamental for landscape 49 compositional variation than the altitudinal gradient, variation such as between plains along a 50 gradient from the lowland to the mountains in the above-mentioned systems (Norwegian: 1 'lavlandssletter' and 'fjellvidder', respectively) are captured by the bio-ecological gradients 2 'forest cover', and hence included in the hierarchy of minor types. Our analyses do not 3 provide support for including 'fissure valley landscapes' or 'coastal archipelago ' (c.f. 4 Rudberg 1960;Mattson et al. 1984) as major types in a consistent system of major landscape 5 types. The fine-grained terrain variation characterising such landscapes will not be captured 6 by morphometric variables derived from a relatively coarse-grained DEM (100×100 m grid 7 cells, derived by interpolation of height data from 20m contour lines). We assume that both 8 fissure valley landscapes and coastal archipelago (i.e. partly submerged fissure valley 9 landscapes), will be recognised by the use of a high-resolution DEM. 10 Our analyses support the identification of 'inland valleys' as a major landscape type. 11 Inland valleys are widely acknowledged as an important geomorphological feature in 12 geomorphological literature (Gjessing 1978, Sulebak 2007, Erikstad 2009), but not included 13 as a separate type in earlier versions of geomorphological maps covering Norway. 14 Klemsdal (1982Klemsdal ( , 2010 identify six primary coastal types in Norway based on 15 morphogenetic criteria (i.e. landform-creating processes): (1) 'strandflat'; (2) 'fjord'; (3)  16 'fjärd,' (a Swedish term that refers to submerged valleys that do not have a glacial over-17 deepened character, often found in Swedish and Southeastern Norwegian coastal landscapes 18 with fissure valleys and 'skjaergård'); (4) 'cliff abrasion coast'; (5) 'flat abrasion coast'; and 19 (6) 'moraine topography coast'. The major type 'coastal plains' identified in our study 20 encompasses the strandflat, the fjärd, the flat abrasion coast and the moraine topography coast 21 in Klemsdal's typology. Klemsdal's 'cliff abrasion coast' is not identified as a major 22 landscape type on its own in our study. Neither do our analyses reveal distinct patterns of 23 landscape variation that motivate for distinguishing coastal hills and mountains as a separate 24 major type. Contrary to the analyses for inland plains, the ordinations neither reveal patterns 25 of variation along a gradient from coast to inland nor discriminate a coast-near group of 26 landscapes within the major type inland hills and mountains. This result may, however, partly 27 result from the rules for delineation of OUs: 'coastal hills and mountains' was represented in 28 the data material with only 3 OUs. In the geomorphological literature, cliff abrasion coast is 29 consistently described as a distinct coastal landform at the scale of our study (Norwegian: 30 'klippekyst'; 'naeringer'; see e.g. Gjessing 1977, Klemsdal 1982, Trømborg 2006, Sulebak 31 2007. We suggest that the distribution of this landform and its co-occurrence with other 32 landforms and landscape elements should be explored further by inclusion of morphometric 33 variables derived from a fine-grained DEM in the analyses. If new analyses still do not 34 provide support for identification of cliff abrasion coast as a separate landscape type, 35 occurrence of this landform can be recognised as a category in the attribute system (see 36 discussion later). 37 According to EcoSyst principles, a demand on major types is that they differ with 38 respect to the set of complex gradients needed to describe important within-type variation 39 (Halvorsen et al. 2020). This rule motivates for a division of inland plains into two major 40 types: inland fine-sediment plains and other inland plains, based on a clear separation in the 41 ordination diagrams between these two groups with respect to dominant soil type. 42 Nevertheless, since the relatively small number of OUs (118) from inland fine-sediment 43 plains (118 OUs) form a weak basis for identification of major CLGs, we recommend that 44 patterns of landscape-element composition in inland plains are explored further. 45 The discrepancies between the major types identified by our analyses and earlier 46 Norwegian meso-scale landform typologies may have (at least) three explanations: (1) the 47 major types in our study are identified by similarities in surface geometry (morphometry) 48 without explicitly accounting for the morophogenetic processes responsible for these patterns; 49 (2) finer-scale variation in surface geometry (e.g. gullies, fissures, cliffs, etc.) are not well 50 captured by the relatively coarse-scaled DEM used as the basis for our study; and (3)  1 properties of the data material, such as low representation of OUs from inland plains and 2 coastal hills and mountains or lack of high-quality data for important characteristics such as 3 soil type. On the other hand, the earlier typologies, which express subjective expert opinions, 4 do not provide definitive answers to typification questions. Anyhow, we recommend that the 5 issues raised above are addressed by morphometric studies based on a more detailed DEM, 6 with soil-type data of more consistent quality, and a sufficiently large set of representative 7 observation units. 8 In a global context, the major landscape types identified in our study largely resemble 9 commonly applied classifications of macro-and meso-scale landforms (e.g. plains, hills and 10 mountains) based on surface geometry, in the tradition of Hammond (1954), Dikau (1989), 11 Brabyn (2005) to 10 4 m and size of the delineated areas from 10 4 to 10 8 m 2 . 'Mesorelief' is the level between 15 'macrorelief' (e.g. larger mountain areas such as the Alps or the Rhine Graben) and 16 microrelief (typically identifying landforms such as gullies, dolines, sand dunes and terraces). 17 Sulebak (2007), on the other hand, refer to landforms comparable to the major types in our 18 study as 'macro-relief', exemplified by the strandflat, valleys, fjords and large river plains 19 with width of the units from 10 to 10 3 m and size of the delineated areas from 10 2 to 10 6 m², 20 resulting from processes operating over 10 3 -10 6 years. Complex gradients of landscape element composition 28 29 The analyses show that each major type suggested by the analyses possess a unique set of 2-5 30 important complex landscape gradients. Furthermore, all major types satisfy the criteria by 31 Halvorsen et al. (2020; Appendix S2.3) that EcoSyst major types have to possess at least one 32 CLG not shared by another major type, and that each major type shall comprise more than one 33 ecodiversity distance unit in landscapes (EDU-L) of variation in landscape-element 34 composition. Although CLGs were identified for each major type independently, many CLGs 35 expressed similar properties across major types. Accordingly, the 24unique CLG variables 36 used to model the proxies representing each major-type specific CLG can be aggregated into 37 7 'universal CLG candidates' ( Main geo-ecological landscape gradients in Norway 2 3 The geo-ecological CLGs identified in our study express variation in broad structural patterns 4 of abiotic conditions. The analyses reveal that relief (RE) is an important geo-ecological CLG 5 in all of the recognised major types except inland plains (which, by definition, has low 6 relative relief). The relief-related CLGs express finer-grained terrain properties than the 7 coarse-grained variation that forms the basis for the division into major landscape types. 8 Different combinations of variables are used to model the relief gradient in each of the major 9 types. Relief in inland hills and mountains expresses a gradient in steepness and ruggedness 10 from gently undulating to steep and rugged hills and mountains. Relief in valleys and fjords, 11 on the other hand, expresses the depth and width of the valley (or the fjord) relative to its 12 surroundings; from wide, open valleys and fjords to narrow, deep and steep-sided valleys and 13 fjords (see Sulebak 2007). Relief in coastal plains expresses a third, related but slightly 14 different, gradient in topographyruggedness on a finer scale, i.e. a gradient from flat to 15 rugged coastal plains, the latter often with remnant peaks (in Norwegian: 'restfjell'). 16 All of the relief gradients identified in this study are previously recognised and 17 described in the literature (Sulebak 2007). However, to our knowledge they have never been 18 related to other geo-and bio-ecological landscape properties by quantitative analysis at the 19 level and scale of our study. The extensive gradient lengths, the high number of strongly 20 correlated variables and the large fraction of the total variation in landscape properties 21 explained by the relief-related CLGs show that relief controls most of the landscape variation 22 in valleys, fjords and hills and mountains. Our division of the relief gradient in plains, hills 23 and mountains also accords with landform classifications applied at regional, continental or 24 global scale described by, e.g. Hammond (1954) types based on stepwise variation in relief; e.g. plains (from flat to irregular) and hills and 27 mountains (from moderate to very high relief). 28 Two geo-ecological variables identified in our study express a gradient from coast to 29 inland. The CLG inner-outer coast (IYK) is identified as the most important CLG in coastal 30 plains, while the CLG 'distance to coast' (KA) is identified as the most important gradient in 31 other inland plains. The first CLG explains variation in landscape element composition from 32 the outer coast, directly exposed to the actions of wind, waves and ocean currents towards 33 coastal plains with more typical inland properties. The latter CLG, distance to coast, may be 34 interpreted as a continuation of the former, separating inland plains < 5 km from the coastline 35 from plains further inland. Notably, the analytical variable mean distance to coast (Discoast) 36 was also strongly correlated (τ = -0.45) with the first GNMDS ordination axis in inland fine-37 sediment plains, although not identified as a separate CLG in this major type. Regional 38 environmental variation in climatic conditions and species distributions from coast to the 39 interior of a landmass is well documented in biogeography worldwide (Lomolino et al. 2017) 40 as well as in Norway (Gjaerevoll 1973, Moen 1999, Bakkestuen et al. 2008). Our study shows 41 that this gradient is important for landscape-level variation as well, as expressed by variation 42 in the composition of landscape elements. 43 All hydrological variables and properties are, as expected, strongly related to landform 44 variation, but the relative importance of various hydrological properties vary among major 45 landscape types. The abundance of wetlands (including small lakes and tarns; VP) is 46 identified as an important CLG within landforms with low relief (except inland fine-sediment 47 plains), and within fjords with low relative relief. This CLG identifies areas with high 48 abundance of mires, small lakes and tarns as different from areas with low abundance of these 49 elements. Variation in these landscape elements is a well-known property of boreal and 50 subarctic bioclimatic zones (Keith et al. 2020) and is well documented throughout Norway 1 (Moen 1999, Bryn et al. 2018. The CLG freshwater lake properties (IP) in valleys identifies 2 'inland fjords ' (Erikstad et al. 2009) and other hydrological features typically occurring in 3 inland valleys, while the CLG freshwater lake properties (IP) identified in other inland plains 4 separates areas with many small lakes from areas with few lakes, thus expressing a different 5 aspect of hydrographic variation. 6 7 8 Main bio-ecological landscape gradients in Norway 9 10 Our study identifies two bio-ecological CLGs: 'open area cover' (AP; identified as important 11 in fjords and inland fine sediment plains) and 'forest cover' (SkP; identified as important in 12 other inland plains, inland valleys and inland hills and mountains). Forest cover is also 13 identified as a CLG-candidate in coastal plains in which, however, the additional variation 14 explained by forest cover after accounting for variation along the gradient 'inner-outer coast is 15 too small to meet the criterion for inclusion as an important CLG. Interestingly, the bio-ecological CLG forest cover (SkP) in other inland plains was 23 characterised by an extensive gradient length and, accordingly, as many as five segments. 24 This is an indication that relief controls and constrains a large amount of the bio-ecological 25 variation in the major types where relief is important, while otherwise this variation has to be 26 attributed directly to a bio-ecological variable which then will comprise more segments. 27 Landforms such as mountain plateaus and lowland plains are clearly separated by this 28 CLG (e.g. Rudberg 1960, Winsnes et al. 1983; see discussion in the section about major 29 types). We interpret the bio-ecological CLG as a response to abiotic processes that scale up to 30 structural patterns of 'expressed ecological landscape elements' at landscape-relevant spatial 31 scales. Examples of such processes are the geomorphological processes responsible for 32 landforms (Sulebak 2007 The main land-use related landscape gradients in Norway 42 43 Variation in the density of buildings, infrastructure and built-up landcover types, here referred 44 to as 'amount of infrastructure', is identified as an important landscape gradient in all major 45 types. This category comprises gradients in the aggregated outcomes of past and present 46 exploitation of natural resources; typically expressing variation from natural via rural to urban 47 landscapes or from landscapes with extensive or little human land use to landscapes shaped 48 by intensive farming or other harvesting of natural resources. In the initial analyses for 49 identification of CLG candidates, the CLG 'amount of infrastructure' is strongly correlated (τ 50 > 0.55) with the first GNMDS axis in all major types. Nevertheless, the analyses for deriving 1 a parsimonious set of CLGs within each major type show that the additional amount of 2 variation explained by amount of infrastructure after the geo-ecological and, eventually, bio-3 ecological, variation is accounted for, is generally low. These results show that patterns of 4 human land use are strongly determined and controlled by geo-ecological variation. 5 Furthermore, our analyses reveal the broad-scale geo-ecological conditions at which the 6 intensity of human land use is at its highest. Xu et al. (2020) point out that humans, just like 7 other species, have an environmental niche, and demonstrate that humans throughout history 8 have concentrated in a surprisingly narrow subset of Earth's available climates. Our study 9 indicates potentially interesting relationships between geomorphology and land-use patterns 10 that should be explored further in a paleoecological context. 11 Agricultural land-use intensity is not identified as an independent CLG in our study, 12 although the major type 'inland fine-sediment plains', defined by soil type, to a large extent is 13 characterised by high agricultural land-use intensity. Interestingly, the variation in agricultural 14 land-use intensity consistently follows that of infrastructure. The CLGs identified in this study express variation in the distribution and abundance of 36 observable landscape elements and properties, and do not explicitly address the processes 37 behind these patterns. In this sense, they are parallel to the 'coenoclines' of Whittaker (1960), 38 the species compositional gradients, which do not explicitly that the conditioning 39 environmental complex-into account and therefore do not allow direct mechanistic modelling 40 of the processes which give rise to the observed patterns. Nevertheless, the order of historical 41 events and processes that have brought about landscape variation is well established by 42 geological, paleoecological and historical evidence (Delcourt et al. 1982, Birks 1993. The 43 identification of CLGs opens for further studies of the relative magnitude of the drivers 44 behind landscape variation, and the response to these drivers as expressed in landscape 45 element composition. Such knowledge may have potential importance also for future 46 predictions of landscape change. 47 The total set of CLGs identified as an important component of a parsimonious 48 characterisation of landscape variation, bears resemblance to 'factors', 'components', and 49 categories commonly applied in biophysical landscape characterisation and mapping. In a 50 review of 54 methods, Simensen et al. (2018) find that the factors most often used for 1 landscape characterisation and mapping are landform (included in 95% of the analysis 2 methods), general land cover (83%), vegetation cover (81%), geology (67%), soil (63%), 3 hydrography (57%), agriculture (54%) and buildings and infrastructure (52%), all of which 4 are included in the set of CLGs identified in our study. Since our study explicitly addresses 5 the physical outcome of the processes that constitute landscape variation (i.e. observable 6 landscape elements), drivers such as climate (included in 44% of the studies in the review) are 7 not among the analysis variables included in our study. 8 The 'key factors' applied by Puschmann (2005) in a structured qualitative description 9 of 45 coarse-scale 'landscape regions' and 444 'sub-regions' in Norway are: (1) major 10 landform; (2) small-scale landforms; (3) water and watercourses; (4) vegetation; (5)  11 agriculture; (6) buildings and infrastructure. Together, Puschmann (2005) use these 12 components to qualitatively derive and describe the 'landscape character' of each region, sub-13 region and/or 'landscape area', subjectively assigning to each component in an area or a 14 region a value from one to three to indicate the component's relative importance for the total 15 landscape character. Although capturing much of the same variation as our study does, the 16 scale and subjective procedure makes it difficult to develop this approach further in a more 17 quantitative setting. Our quantitative approach successfully identifies the properties that are 18 commonly regarded important in landscape characterisation and mapping, and, more 19 importantly, provide strong empirical evidence for the relationship between these variables, 20 i.e. the relative importance of each CLG. Thus, the outcome of our study lends strong support 21 to Strand (2011), who argue that quantification of landscapes by use of variables and proxies 22 for subsequent application in GIS-based mapping, opens for consistent landscape-type 23 mapping with a level of detail and at spatial extents unachievable by field-based landscape 24 type mapping. 25 Gradient analysis at the landscape level is, to a large extent, still unexplored [but see 26 Luck & Wu (2002)  space. However, many of the landscape gradients recognised in our study are probably 31 relevant over vast areas and likely to persist for long times, even in a changing climate (e.g. 32 terrain gradients, coast-inland gradients, gradients in human land-use intensity, gradual 33 changes in vegetation cover, the abundance of lakes, etc.). We suggest that the distribution of 34 ecosystems along complex-gradients in the landscape should be explored across larger areas, 35 as such studies may yield new insights of both theoretical and practical importance. 36 37 38 Explained and residual landscape variationthe need for a descriptive attribute system 39 40 Like ecodiversity models in general, the landscape space model partitions variation into 41 'explained' variation and residual, 'unexplained' or 'apparently random' (Økland 1990), 42 variation. A landscape-type system built on CLGs will disregard all variation not explained by 43 the important CLGs. Unless appropriately handled, this will have the unfortunate consequence 44 that the 'uniqueness' or idiosyncracies, of landscapes (Phillips 2007) which is expressed in 45 unexplained variation, will be lost. This motivates for a standardised attribute system by 46 which the residual variation can be described in a standardised way (see Halvorsen et al. 47 2020). No such system has, however, yet been developed for description of landscapes in 48 Norway. We recommend development of an attribute system according to EcoSyst principles,49 structurally similar to that for ecosystems (see Halvorsen et al. 2020). Based on our study, 50 good candidate variables for inclusion in such an attribute system, are CLG-candidates that 1 'did not make it' to the final set of important CLGs within each major type. An example of 2 such a gradient is 'agricultural land use-intensity', which failed to explain enough variation to 3 be identified as a separate, important, CLG, but which is obviously useful for a more fine-4 grained and precise description of the various aspects of human legacies in the landscape. 5 Furthermore, CLGs identified as important in other major types, can be used for a more 6 detailed assessment, such as vegetation cover in coastal plains. This CLG was strongly 7 correlated with geo-ecological variables, not explaining enough addtional variation to be 8 included as a CLG for coastal landscapes. Moreover, all the 85 variables used for the analyses 9 can in principle be useful as variables in an attribute system, for two at least reasons: (1) 10 because they are specifically selected to describe aspects of the landscape that are considered 11 relevant at the scale and level of our study; and (2) because they result from a pruning process 12 by which strongly correlated variables were omitted. All the 85 analysis variables therefore 13 account for some of the variation in landscape properties not also accounted for by other 14 variables. 15 Data about the presence, absence and relative abundance of major ecosystem types 16 (e.g. exposed ridge, open fen, mire and swamp forest) and microrelief landforms (e.g. gullies, 17 dolines, sand dunes and terraces) will have to be important components of an attribute system 18 for the landscape level. Although field-survey based mapping of such landscape elements is 19 time-consuming and expensive, alternative, efficient pathways to acquiring information about 20 their spatial distribution may be used for this purpose, such as remote sensing (Pereira et al. The tentative landscape-type system built on the results of our study highlights 24 landscape-element composition (the relative performance of different observable objects 25 within each spatial unit), with less emphasis of structure (the patterns of distribution of 26 observable objects within these spatial units) or function (the processes that give rise to the 27 variation). We therefore recommend that landscape structure and function are incorporated as 28 central components in the attribute system. This can be accomplished in many different ways 29 due to the rapid proliferation of available 'landscape metrics' for quantification of the spatial 30 configuration of categorical landscape patterns over recent decades ( determines the level of detail captured at the lowermost hierarchical level in the landscape-6 type hierarchy. When compositional variation along most CLGs is continuous, no a priori 7 reason exist for selecting a specific EDU-L value to define the edge of an ideal basic-type 8 hypercube in landscape space with important CLGs as axes. Standardised stepwise division of 9 landscape gradients can then, as for LCEs on the ecosystem level, be done by defining one 10 EDU-L unit as a specific threshold value on the PD dissimilarity scale. No other guidelines 11 for how much variation one EDU-L should account for, exist other than practical 12 considerations; the choice of the threshold value for landscape distance will determine the 13 level of detail of the type-system. A high threshold value will result in a coarse-grained type-14 system with few categories while a low threshold value will provide a fine-grained type-15 system with many categories. The choice between the two approaches is largely a matter of 16 preference; i.e. whether you are a 'lumper' or a 'splitter', in the words of George G. Simpson 17 (1945). 18 With the ambition to balance different considerations, we defined one ecodiversity 19 distance unit in landscapes (EDU-L), i.e. one landscape distance unit, to 0.08 PD units. The 20 number of major type-adjusted segments by definition equals the number of EDU-L between 21 CLG endpoints, rounded down to the nearest whole number (Halvorsen et al. 2020). This 22 means that the gradient length must be at least 2 EDU-L units (> 0.16 PD units) in order for a 23 CLG to be used to divide a major type into basic types, at least 0.24 PD units to provide three 24 segments and a series of three basic types, etc. Experience from the analysis of the inland data 25 sets (especially the inland vallet data set ID1012), which resulted in gradient lengths up to 26 0.5313 for CLG 1 (Relief), corresponding to > 6 EDU-L, may indicate that the definition '1 27 EDU-L = 0.08 PD units' results in a type-system that is too fine-grained, i.e. with too many 28 types at the lowest level. An alternative view, supported by the fact that the swarm of OUs 29 becomes 'thinner' (shows reduced point density) from mid-points towards endpoints along 30 almost all ordination axes, is that the landscape gradient length estimates are 31 disproportionately heavily influenced by outlier OU. Outliers are OUs that occupy isolated 32 positions in an otrdination diagram (Økland 1990). Outlying OUs may represent: (1) 33 individual OUs with atypical properties for the major type in question; (2) OUs in the 34 transition zone between two adjacent major types; or (3)  Any statistical analysis is a product of the available data and the analytic methods used. This 16 study is no exception. Despite strong efforts to make the approach to landscape analysis 17 presented here observer-independent, the results of the analyses depend on the selection of 18 data sources and other subjective choices made during the analytic procedure (Yang 2020). 19 The 85 variables used in the multivariate analyses are derived from different sources that are 20 mapped, collected or modelled at different time-points and by use of different methods, 21 inevitably introducing variation in the quality of the input data. Even though we used 22 variables derived from standardised national mapping programs (e.g. AR 5 land-cover maps, 23 N50 topographical land-cover data, etc.), also data from these basic data sources are of 24 varying quality. This is exemplified by Bryn et al. (2018), who show that the cover of 25 wetlands in the basic Norwegian land-cover map N50 (data also used in our study) is strongly 26 underestimated while the cover of, e.g. forest, has been correct. Furthermore, the lack of area-27 covering and/or reliable data for potentially important landscape elements and properties is an 28 important source of uncertainty in the analysis. 29 Although we analysed a broad set of 85  for focal calculations). We assume that analyses based on a high-resolution grid (i.e. pixels < 49 100×100 m) will improve the overall sensitivity of the terrain analyses. 50 A potential source of uncertainty related to our statistical analyses is the number of 1 correlated variables representing different, but thematically and/or causally related properties. 2 Although strongly correlated variables (τ > 0.7) were removed prior to the analyses, there is 3 still a risk that related landscape properties represented by many correlated variables (e.g. 18 4 analytical variables related to buildings and infrastructure and 15 variables related to 5 topography) are disproportionately strongly emphasised in the analyses. 6 Another possible source of errors and uncertainty in our results is the sensitivity of the 7 analytical results to the definition and spatial delineation of data collection units. The 'zoning 8 problem' is an aspect of a well-known problem in the statistical analysis of spatial data known 9 as the 'modifiable areal unit problem' (MAUP; Gehlke & Biehl 1934, Openshaw 1984. 10 Jelinski & Wu (1996) demonstrate that, when a given set of areal units is recombined into 11 zones that are of the same size but located differently, variation in data values and, 12 consequently, different conclusions may result. According to Fotheringham & Wong (1991), 13 the effects of MAUP is a bigger problem in multivariate statistical analysis than in univariate 14 or bivariate analysis. We addressed the MAUP problem according to the recommendations of 15 Jelinski & Wu (1996) by applying an 'optimal' zoning system, i.e., a system that maximises 16 interzonal variation and minimises intrazonal variation. This was done by delineating 17 observation units by a set of rule-based criteria derived from the 'pilot Nordland study'. This 18 approach may reduce variation in results of an analysis caused by the MAUP, but has the 19 disadvantage that 'optimality' is a subjective criterion, both with respect to definition and to 20 operation. Furthermore, delineation errors may be introduced during the process by which 21 'optimal spatial units' are constructed. This is exemplified by inconsistencies or uncertainties 22 in the OU delineation process, which is mentioned as a possible cause of the 'thinning' of 23 OUs towards CLGs ends observed in this study, and by OUs transitional between inland 24 plains and inland mountains. The criterion-based delineation of OUs may also introduce an 25 element of circularity in the analysis. By defining OUs by spatial delineation based on a set of 26 pre-defined criteria, there is a risk that the analyses will identify the same OUs and confirm 27 the criteria used to delineate them. Use of standard sample units of a constant size (e.g. 1×1 28 km cells) removes the subjective judgement involved in the delineation of irregular sample 29 units, but does not solve the zoning problem (i.e. the infinite number of possible locations of 30 these zones). Furthermore, standardised sample units have the drawback that fundamentally 31 different landforms and/or ecosystems such as mountains and ocean systems will inevitably 32 be included in the same unit (Bunce et al. 1996). A lesson learnt from the pilot study in 33 Nordland ( show that predominantly abiotic processes, to a large extent, control and/or constrain biotic 48 processes and variation in human land use. Predictable patterns of co-occurrence exist 49 between landscape elements and other properties of landscapes, and most landscape elements 1 tend to have distinct performance (probability of presence and/or abundance) optima along 2 broad-scale compositional landscape gradients. Based on the results presented here, we 3 suggest that the gradient perspective, so often applied in studies at the ecosystem level, is a 4 conceptual framework appropriate for studies of landscape-level variation. Such studies may 5 contribute to a more coherent understanding of the relationship between connections between 6 non-living and living nature. 7 Although classification must be considered a tool, not a goal in itself (Økland &  8 Bendiksen 1985), any type-system developed with the aim of representing real systems or 9 processes should be based upon a consistent theoretical framework and the best available 10 empirical evidence. Our study provides, to our knowledge, the most comprehensive 11 quantitative assessment that simultaneously addresses biotic and abiotic variation at the 12 landscape level of ecological diversity in Norway. Although data for some of the major types 13 are sparse and should be explored further ( Pereira Simensen, T., Horvath, P., Vollering, J., Erikstad, L., Halvorsen, R., Bryn, A. 2020. 8 Composite landscape predictors improve distribution models of ecosystem types. 9 Divers Distrib. 00: 1-16. Prior to the study, a pilot study in Nordland county were conducted. Basic methodological 7 characteristics of the pilot study in Nordland landscape analysis was: 1) a rule-based division 8 of the landscape into 258 discrete spatial units; 2) recording of a broadest possible selection of 9 physical landscape attributes within each spatial unit (i.e. 173 landscape variables); and 3) 10 multivariate statistical analyses of the data set in order to identify 'landscape gradients', i.e., 11 'gradual variation in the presence and/or abundance of landscape elements'. The resulting 12 typology was obtained by dividing the identified, major complex landscape gradients that 13 explained most variation in the composition of landscape elements into segments, and then 14 defining landscape types by combining intervals along several gradients. 15 An overview of the method is presented in presented in Erikstad et al. (2015)

Minimum mapping units
The spatial units (polygons) representing major landscape types have to satisfy the following main criteria: 1. Area > 4 km² 2. Minimum width > 2.5 km (geodesic distance between the outer boundaries of a spatial unit) Delineation 1. Spatial units representing a major-type candidate that satisfy the geomorphological definition of a major landscape type but do not meet the minimum size criteria shall, unless otherwise indicated, be assigned to adjacent spatial units at the best discretion. 2. Exceptions are described in the procedure in Table S1.
Delineation order: steps A1 -A7 below (applied strictly, no exceptions) Table S1. Delineation rules, major landscape types. Abbreviations: RR1 = relative relief 1 within a 1 km neighbourhood; TPI = terrain position index; TPI(mt) = terrain position index 2 based on marine and terrestrial areas (the complete DEM); TPI(t) = TPI based on terrestrial 3 areas only (marine areas set to meters above sea level = 0). The number after 'TPI' refer to the 4 size of the neighbourhood circle for TPI-calculations. 5 6 Step Task

Comments:
1. Upper border of spatial units for fjords shall be corrected by TPI(t) = 0, i.e., TPI in a calculation where all marine areas are set to 0 meters above sea level, instead of TPI(mt) = 0, when the line for TPI(mt) = 0 a. crosses the coastline b. is located closer to land than 250 m over a distance of 1 km, and/or c. is located in a continuous spatial unit with relative relief (RR1)>200 m (steep terrain) that crosses the coastline in a length of at least 1 km 2. For valleys TPI6(mt) = TPI6(t)

A2
Delineating spatial units for coastal plains Definition: 1. Coastal plains contain the 'Strand flat' along the coast in the classical (geomorphological meaning of the term, from Rogaland and northward, and 2. Other coastal landscapes with similar relief located other places along the coast

Comments:
1. Upper border towards hills and mountains above the coastal plains are set to TPI(mt) = 0, except cases where the line for TPI(mt) = 0 a. crosses the coastline b. is located closer to land than 250 m over a distance of 1 km, and/or c. is located in a continuous spatial unit with relative relief (RR1)>200 m (steep terrain) that crosses the coastline in a length of at least 1 km Step Task Main criterion Additional criteria, special cases, comments When the criterion in A, B or C is met, the line shall be corrected by use of TPI(t) = 0. 2. If further corrections are needed (without easily identified inflection points in slopes) the criteria A and B in the procedure for delineation fjords are applied.

A3 Delineating spatial units for residual areas with coastline
Definition: 'Residual areas with coastline' contain areas with coastline, not assigned to spatial units of the major landscape types fjords/valleys or coastal plains in step A1 and A2.
Delineation of spatial units for candidate major types include manual assessment of the minimum mapping unit criterion (2), interpreted as generalised coastline length > 2.5 km.
1. If the criterion for the minimum mapping unit is met, the spatial unit is assigned to the major type 'coastal hills and mountains'. 2. If the criterion for the minimum mapping unit is not met, and the affiliation to major type is equal on both sides of the spatial unit, the adjacent units are merged together. 3. If the criterion for the minimum mapping unit is not met, and the affiliation to major type is different on both sides of the spatial unit, the coast-line segment shall be split and merged by the two adjacent spatial units. The borderline between them shall follow a ridge or another natural terrain divide (i.e. ridge, depression, etc.). Note that the criterion 1 for size of minimum mapping unit is not applied in this step.

Main criterion:
The borderline shall follow a ridge or another natural geomorphological structure in the terrain across (perpendicular to) the

Exceptions:
The main criterion is overruled by continuous spatial units with amount of infrastructure (OI ≥ segment 5, i.e. village or city).
Step Task Main criterion Additional criteria, special cases, comments valley, maximum 2 km from the end of the fjord ('fjordbotn'). Indicative tool: map of maximum TPI1.
These spatial units shall not be split between fjords and valleys but shall be assigned as fjords.

Comment:
1. Some places, short valley segments will be generated between the fjord and the surrounding hills and mountains. These segments shall be delineated as valleys when the segment is longer than 2 km. When this criterion is not met, the valley segment shall be joined with the fjord unit.

Separating coastal hills and mountains from inland hills and mountains and inland plains
Main criterion: TPI6(mt) = 0 (similar to the criterion for separation between the major types 'fjords' and 'coastal plains'), secondarily TPI(t) = 0.
Manual delineation is applied in gentle or rugged terrain where TPI = 0 is not applicable. The borderline between coastal hills and mountains and inland hills and mountains shall follow the first natural geomorphological feature (normally a ridge), maximum 2 km from the coast-line. Indicative tool: map of maximum TPI1.

Manual correction of lower borderline (to marine landscapes)
is applied with similar criteria.

Exceptions:
1. Continuous coastal spatial units with amount of infrastructure (OI ≥ 5, i.e. village or city). These spatial units shall not be split, but be assigned as one spatial unit with affiliation to coastal plains. 2. If it is not possible to delineate inner spatial units on islands in line with the general size criteria for hills, mountains and inland, the island shall not be split between inland and coastal landscapes 3. If it is not possible to delineate spatial units in line with the general size criteria for hills, mountains and inland plains between a spatial unit for coastal hills and mountains and a spatial unit for inland valleys (i.e., a ridge/hilltop between the valley [perpendicular to the valley direction] and the coastal areas outside), the area in between shall be split between the two spatial units.

A6 Delineating spatial units for steep and rugged hills and mountains within coastal plains
Spatial units for hills, mountains and plains shall be delineated within a spatial units in coastal plains, regardless of the size criterion for minimum mapping units when the criterion for 'steep and rugged hills and mountains' (TP = 2, based on TPI1) is met.

Comment:
1. Segment A7 represent an exception from the minimum size criterion.

Delineating spatial units for inland plains Definition
Inland plains are defined by a combination of properties:

Comment 1. Borderline towards higher hills and mountains is drawn with use
Step     JK1 Primary soil/ sediment class B -exposed bedrock (> 50% of area covered by exposed bedrock) E -glacifluvial deposits H -marine deposits M -thick layer of till 0 -no soil/sediment class (> 50% not assigned to specific sediment/soil type) JK2 Secondary soil/ sediment class B -exposed bedrock (> 25% of area covered by exposed bedrock) E -glaciofluvial deposits H -marine deposits M -thick layer of till 0 -no soil/sediment class (> 25% not assigned to specific sediment/soil type)  Fig. 4. The principle of key variable calculation as a focal statistic (frequency) within 81 standardised grid cells of 100×100 m. The 81 cells (marked with pink shadow) are wholly or partly contained within a circle (marked with grey shadow) with a radius of 500 m around one focus point for which the key variable is calculated (marked with a black dot located in the middle of a cell). In the example, the property the key variable is based on (e.g., presence of buildings) indicated by red dots. In this example, the key variable has the value 10, or alternatively 10/81 = 0.123 if given as frequency.       Table 1).  Table 9. Characterisation of observation units in the transition zone between inland and coastal landscapes in the DCA ordination diagram of the Tot3966 data set (Fig. 8)   OUs (with coastline; KL = 2). Affiliation to LT-variables (LT variables, see Table 1): IA = inland hills and mountains; ID = Inland valleys; KA = coastal hills and mountains. SN = 1: mainland OU; SN = 2: observation unit on large island (> 20 km 2 ); SN = 3: observation unit on small or medium-sized island (1.5-20 km 2 ).
Figs 16-17. (16) Close-up showing the distribution of the 162 OUs in the X162 data subset (large, red dots) in the transition zone between inland (without coastline, left; red dots) and coastal landscapes (with coastline; blue dots) in the DCA ordination diagram, axes 1 and 2, for the Tot3966 data set (scaled in S.D. units). Isolines for medium relative relief (analytic variable RR1_m; black lines) and inverse island size (Oyst_i; green lines) are shown. OUs affiliated to LT-type IA, situated on islands, are indicated by large red dots. (17) Proportion of OUs misclassified (i.e., OUs without coastal line that are erroneously classified as coast), as function of inverse island size (Oyst_i index), provided that inverse island size is used to differentiate between coastal and inland major-type groups. The minimum value of 14.8% was obtained for Oyst_i = 0.733, which corresponds to an island size of 10 km 2 .    Table 1); isolines for island size (variable Oyst_i); and vectors pointing in the direction of maximum increase of Oyst_i and selected primary key variables (see Table 4 for explanation). The black, continuous '4.5' line marks equal probability (0.5) for an OU to be affiliated with KS and KF.  Table 1).       Table 16. Kendall's rank correlation coefficients (τ) between GNMDS ordination axes 1-2 for the inland plains data subset (IS520) and the 9 primary key variables (rows 1-10) and the 65 analytic variables used to characterise the 520 observation units.       Table 5. (44) Isolines (red) for the infrastructure index (Ifl). OI = stepwise variation in amount of infrastructure from low (OI 1-2) via medium (OI 3-4) to high amount of infrastructure, including towns and/or cities (OI 5-6). (45) Isolines (red) for agricultural land-use intensity (JI). JP = stepwise variation agricultural land-use intensity from low (JP 1) via medium (JP 2) to high (JP 3).  (46) proportion of area with deciduous forest (Arlov); (47) inverse island size (Oyst_i); (48) proportion of area with bare rock (Kbf_a); and (49) mean values for relative relief (RR1). SN 1 = terrestrial coast-line landscape; SN 2 = large islands (> 20 km 2 ); SN 3 = medium sized islands (1.5-20 km 2 ); SN 4 = small islands (0.1-1.5 km 2 ); SN 5 = skerries/islets (< 0.1 km 2 ). direction of maximum increase of the variable (relative lengths of arrows are proportional to the correlation between the recorded variable and the ordination axis). Variable names are abbreviated according to Table 4 and 5. (52-54) Variation in coastal/archipelago properties (SN) and isolines for (52) the number of lakes (Inns_s); (53) the mire index (MI); and (54) the proportion of area covered by boreal heaths (Bohei). SN 1 = terrestrial coast-line landscape; SN 2 = large islands (> 20 km 2 ); SN 3 = medium sized islands (1.5-20 km 2 ); SN 4 = small islands (0.1-1.5 km 2 ); SN 5 = skerries/islets (< 0.1 km 2 ).    Table 20. Kendall's rank correlation coefficients (τ) between GNMDS ordination axes 1-3 for the coastal fjords data subset (KF353) and the 9 primary key variables (rows 1-9) and the 84 analytical variables used to characterise the 353 observation units.  abbreviated according to Table 4 and 5. (59) Isolines (red) for the infrastructure index (Ifl). OI = stepwise variation in amount of infrastructure from low (OI 1-2) via medium (OI 3-4) to high amount of infrastructure, including towns and/or cities (OI 5-6). (60) Isolines (red) for agricultural land-use intensity (JI). JP = stepwise variation in agricultural land-use intensity from low (JP 1) via medium (JP 2) to high (JP 3). (61-63) Valley shape (DN) and isolines (red) for (61) Table 22. Kendall's rank correlation coefficients (τ) between GNMDS ordination axes 1-2 for the inland fine-sediment plains data subset (IF118) and the 10 primary key variables (rows 1-10) and the 66 variables used to characterise the 118 observation units.   Table 24. Kendall's rank correlation coefficients (τ) between GNMDS ordination axes 1-2 for the other inland plains data subset (IX399) and the 9 primary key variables (rows 1-9) and the 65 variables used to characterise the 399 observation units.  the variable (relative lengths of arrows are proportional to the correlation between the recorded variable and the ordination axis). Binary variables are represented by the name, abbreviated according to Tables 4 and 5, and placed at the centroid for the presence class (purple text). (75) Isolines (red) for the infrastructure index (Iflu). OI = stepwise variation in amount of infrastructure from low (OI 1-2) via medium  to high amount of infrastructure, including towns and/or cities (OI 5-6). (76) Isolines (red) for agricultural landuse intensity (JI). JP = stepwise variation in agricultural land-use intensity from low (JP 1) via medium (JP2) to high (JP 3). (77) Isolines (red) for the proportion of OU area situated above the forest line (AlpA). BA = stepwise variation along the gradient from boreal to alpine landscapes from areas dominated by forest (BA 1) via areas located at the forest line (BA 2) to mountainous areas above the forest line (BA 3). (78) Proportion of area covered by with marine deposits (Kmar_a), shown by coloured symbols and isolines (red). (79) Proportion of area with glaciofluvial deposits (Kelv_a), shown by coloured symbols.       Table 28. Kendall's rank correlation coefficients (τ) between GNMDS ordination axes 1-3 for the data subset inland hills and mountains (IA 1618) and the 10 primary key variables (rows 1-10) and the 66 analytic variables used to characterise the 1618 observation units.            Table 3. Isolines are given for relevant primary key variables (Table 4) and analytic variables (Table 5).     4  12  1322  3  13  2111  13  14  2112  5  15  2121  10  16  2122  6  17  2211  20  18  2212  28  19  2221  11  20  2222  20  21  2311  7  22  2312  6  23  2321  3  24  2322  3  25  3111  11  26  3112  1  27  3121  5  28  3211  24  29  3212  21  30  3221  8  31  3222  10  32  3311  11  33  3312  1  34  3321  3  35  3322  6  36  4111  2  37  4112  2  38  4121  1  39  4122  5  40  4211  9  41  4212  8  42  4221  10  43  4311  5  44  4321  1  45  4322  1  46  5111  1  47  5112  8  48  5122  1  49  5211  16  50  5212  8  51 5311 4   Table 3. Isolines are given for relevant primary key variables (Table 4) and analytic variables (Table 5).     Table 3. Isolines are given for relevant primary key variables (Table 4) and analytic variables (Table 5).     8  14  2212  1  15  2221  25  16  2222  17  17  3111  7  18  3112  2  19  3121  24  20  3122  24  21  3211  1  22  3212  2  23  3221  15  24  3222  4  25  4111  4  26  4112  3  27  4121  13  28  4122  5  29  4212  1  30  4221  6  31 4222 2     Table 46. Major type inland fine-sediment plains (IF): key properties * and segmentation ** of candidate complex-landscape gradient G2, relief (RE), as represented by its orthogonal key variable (ONV). OU = observation unit; Threshold = upper limit of class (RSC = rescaling). PD unit = Proportional dissimilarity unit. Segment threshold shows position of standard segment borders relative to initial segments.  Table 3. Isolines are given for relevant primary key variables (Table 4) and analytic variables (Table 5).     (IF): key properties * and segmentation ** of candidate complex-landscape gradient A1, amount of infrastructure (OI), as represented by its orthogonal key variable (ONV). OU = observation unit; Threshold = upper limit of class (RSC = rescaling). PD unit = Proportional dissimilarity unit. Segment threshold shows position of standard segment borders relative to initial segments.         Table 3. Isolines are given for relevant primary key variables (Table 4) and analytic variables (Table 5).  : key properties * and segmentation ** of candidate complex-landscape gradient U1 forest cover (SkP), as represented by its orthogonal key variable (ONV). OU = observation unit; Threshold = upper limit of class (RSC = rescaling). PD unit = Proportional dissimilarity unit. Segment threshold shows position of standard segment borders relative to initial segments.  Fig. 134. Major type inland other inland plains (IX): distribution of 399 observation units (IX399 data subset) along rescaled orthogonal key variables (rONVs) for complex landscape gradients CLG 3 (forest cover; SkP) and CLG 4 (land use intensity; OI). Symbols show affiliation of OUs to levels of relevant landscape-gradient (LG) variables, operationalised as ordered factor variables in accordance with Table 3. Isolines are given for relevant primary key variables (Table 4) and analytic variables (Table 5).   (IX): key properties * and segmentation ** of candidate complex-landscape gradient A1, amount of infrastructure (OI), as represented by its orthogonal key variable (ONV). OU = observation unit; Threshold = upper limit of class (RSC = rescaling). PD unit = Proportional dissimilarity unit. Segment threshold shows position of standard segment borders relative to initial segments.  Table 3. Isolines are given for relevant primary key variables (Table 4) and analytic variables (Table 5).  Table 56. Correlations between the four ONVs and ordination axes (rows 1-2), primary key variables (rows 3-11), and analytic variables in major landscape type other inland plains (IX).  Table 58. Identification of CLG candidates in inland valleys by forward selection of variables using RDA. Results of tests of single CLG candidates are shown in bold-face types. The best model at each step in the selection process is highlighted using light grey shading while the best model after testing all CLG candidates in a functional variable category is highlighted using darker grey shading. Abbreviations and explanations: Var = variable; Cst Var = constraining variable; Cond Var = conditioning variable(s); df = degrees of freedom; VE = variation explained (expressed as fraction of the total variation); ΔVE = additional variation explained; pseudo-F and P = F-statistic used in the test of the null hypothesis that adding the variable(s) do not contribute more to explaining variation in landscape-element composition than a random variable, with the associated P value . Variable names are abbreviated according to Tables 4 and 5. For further explanation, se Material and methods chapter.    (ID): key properties * and segmentation ** of candidate complex-landscape gradient G1, relief (RE), as represented by its orthogonal key variable (ONV). OU = observation unit; Threshold = upper limit of class (RSC = rescaling). PD unit = Proportional dissimilarity unit. Segment threshold shows position of standard segment borders relative to initial segments.  Fig. 139. Major type inland valleys (ID): distribution of 1012 observation units (ID1012 data subset) along rescaled orthogonal key variables (rONVs) for complex landscape gradients CLG 1 (relief; RE) and CLG 2 (freshwater lake properties; IP). Symbols show affiliation of OUs to levels of relevant landscape-gradient (LG) variables, operationalised as ordered factor variables in accordance with Table 3. Isolines are given for relevant primary key variables (Table 4) and analytic variables (Table 5).  Table 60. Major type inland valleys (ID): key properties * and segmentation ** of candidate complex-landscape gradient G2, freshwater lake properties (IP), as represented by its orthogonal key variable (ONV). OU = observation unit; Threshold = upper limit of class (RSC = rescaling). PD unit = Proportional dissimilarity unit. Segment threshold shows position of standard segment borders relative to initial segments.  Fig. 141. Major type inland valleys (ID): distribution of 1012 observation units (ID1012 data subset) along rescaled orthogonal key variables (rONVs) for complex landscape gradients CLG 1 (relief; RE) and CLG 2 (freshwater lake properties; IP). Symbols show affiliation of OUs to levels of relevant landscape-gradient (LG) variables, operationalised as ordered factor variables in accordance with Table 3. Isolines are given for relevant primary key variables (Table 4) and analytic variables (Table 5).   (ID): key properties * and segmentation ** of candidate complex-landscape gradient U1, forest cover (SkP), as represented by its orthogonal key variable (ONV). OU = observation unit; Threshold = upper limit of class (RSC = rescaling). PD unit = Proportional dissimilarity unit. Segment threshold shows position of standard segment borders relative to initial segments.  Table 3. Isolines are given for relevant primary key variables (Table 4) and analytic variables (Table 5).  Table 62. Major type inland valleys (ID): key properties * and segmentation ** of candidate complex-landscape gradient A1, land use intensity (OI), as represented by its orthogonal key variable (ONV). OU = observation unit; Threshold = upper limit of class (RSC = rescaling). PD unit = Proportional dissimilarity unit. Segment threshold shows position of standard segment borders relative to initial segments.  Fig. 147-148. Major type inland valleys (ID): distribution of 1012 observation units (ID1012 data subset) along rescaled orthogonal key variables (rONVs) for complex landscape gradients CLG 1 (relief; RE) and CLG 4 (amount of infrastructure; OI). Symbols show affiliation of OUs to levels of relevant landscape-gradient (LG) variables, operationalised as ordered factor variables in accordance with Table 3. Isolines are given for relevant primary key variables (Table 4) and analytic variables (   Table 65. Identification of CLG candidates in inland hills and mountains by forward selection of variables using RDA. Results of tests of single CLG candidates are shown in bold-face types. The best model at each step in the selection process is highlighted using light grey shading while the best model after testing all CLG candidates in a functional variable category is highlighted using darker grey shading. Abbreviations and explanations: Var = variable; Cst Var = constraining variable; Cond Var = conditioning variable(s); df = degrees of freedom; VE = variation explained (expressed as fraction of the total variation); ΔVE = additional variation explained; pseudo-F and P = F-statistic used in the test of the null hypothesis that adding the variable(s) do not contribute more to explaining variation in landscape-element composition than a random variable, with associated P value. Variable names are abbreviated according to Tables 4 and 5. For further explanation, se Material and methods chapter.   Fig. 150. Major type inland hills and mountains (IA): distribution of 1618 observation units (IA1618 data subset) along rescaled orthogonal key variables (rONVs) for complex landscape gradients CLG 1 (relief; RE) and CLG 2 (forest cover; SkP). Symbols show affiliation of OUs to levels of relevant landscape-gradient (LG) variables, operationalised as ordered factor variables in accordance with Table 3. Isolines are given for relevant primary key variables (Table 4) and analytic variables ( Table 5).   (IA): key properties * and segmentation ** of candidate complex-landscape gradient U1, forest cover (SkP), as represented by its orthogonal key variable (ONV). OU = observation unit; Threshold = upper limit of class (RSC = rescaling). PD unit = Proportional dissimilarity unit. Segment threshold shows position of standard segment borders relative to initial segments.  Table 3. Isolines are given for relevant primary key variables (Table 4) and analytic variables (Table 5).   (IA): key properties * and segmentation ** of candidate complex-landscape gradient A1, land use intensity (OI), as represented by its orthogonal key variable (ONV). OU = observation unit; Threshold = upper limit of class (RSC = rescaling). PD unit = Proportional dissimilarity unit. Segment threshold shows position of standard segment borders relative to initial segments.