Wetland birds in the northern prairie pothole region may show sensitivity to agriculture

Wetland losses in the Northern Prairie Pothole Region (NPPR) are largely attributed to agriculture. Since land-use is known to influence bird habitat selection, bird community composition is likely sensitive to the extent of neighboring agricultural activity. We determined which local and landscape habitat variables are most predictive of wetland bird assemblage occurrence in southern Alberta. We:1) identified distinct bird assemblages with a cluster analysis, 2) identified which species were indicative of these assemblages using an indicator species analysis and 3) predicted which bird assemblage would occur in a wetland with a classification and regression tree. Avian assemblages were more loosely defined and had few indicator species. Importantly, assemblages were specific to the natural region in which the wetland occurred. Also, landscapes with higher agricultural activity generally supported waterfowl and shorebirds, likely because agricultural activities excluded wetland-dependent birds that nest in upland habitat. Though waterfowl and shorebirds show poor sensitivity to surrounding landscape composition, edge-nesting wetland avifauna may make good indicators of ecological integrity.


Study Design 95
We selected 72 natural wetlands than ranged in permanence class from temporarily-96 ponded to permanently-ponded (sensu Stewart and Kantrud 1971) and were evenly distributed 97 among six, randomly selected sub-watersheds (three in each Natural Region -Grasslands and 98 Parklands) of glaciolacustrine or glaciofluvial-derived surficial geology (Fig. 1). Our selected 99 wetlands reflected the frequency distribution of wetland sizes within each sub-watershed, based 100 on the Alberta Merged Wetland Inventory (Government of Alberta 2014a), and so were 101 generally small (mean size 0.66 ± SE 0.07 ha). To guard against spatial dependencies, sites were 102 spaced a minimum of 3.5 km apart. Independent of their hydroperiod, wetlands were selected to 103 span a gradient in the extent of agricultural activity in the surrounding landscape (i.e., the 104 percentage cropping, haying, and pastureland covers within 500 m buffers around each wetland's 105 perimeter). Land cover data were derived from the Agriculture and Agri-Food Canada Annual 106 Crop Inventory Data (AAFC 2015) and supplemented with information from the provincial 107 Grassland Vegetation Inventory (Government of Alberta 2014b). We used a 500 m buffer 108 because this scale has been reported as the most influential of bird community integrity in 109 permanently-ponded wetlands in Alberta's Parkland (Rooney et al. 2012 Consequently, we classified wetlands with less than 25% agricultural land cover as being in the 113 least disturbed condition (sensu Stoddard et al. 2006) and used these as a reference condition 114 against which medium (25-75% agricultural cover) and high disturbance (greater than 75% 115 agricultural cover) wetlands could be compared. 116

Bird Surveys 117
Bird surveys were conducted by pairs of observers in 2014 and 2015, following the 118 method described in Wilson and Bayley (2012). In brief, surveys comprised a 10-min visual 119 survey followed by an 8-minute acoustic fixed-radius point count survey, with a radius of 100 m. 120 Because most wetlands were less than 1 ha, a single 100-m radius point count covered the entire 121 wetland. Larger sites were surveyed from two-point count locations, providing they could be 122 positioned at a minimum of 200 m apart, in which case counts were summed to reflect the 123 wetland as the sample unit. 124 Surveys were conducted twice at each wetland during the breeding season (May 19 th -125 June 24 th ) to account for any temporal partitioning of breeding activity among species within the 126 general breeding season. Consequently, we summed counts across surveys. Generally, birds in 127 our study region sing and call between sunrise and 11:00 am (Farr et al. 2012). Thus, all surveys 128 were restricted to this time period. 129 All birds visually observed foraging or nesting or heard singing or calling at the site were 130 enumerated and identified to species. Bird identifications followed the American Ornithologists 131 Union Standard Information used to determine guild membership of bird species, such as feeding 132 traits, preferred habitat, and migration patterns, were retrieved from Birds of North America 133 Online (CLO 2014). We distinguish between this complete bird assemblage and the subset of 134 birds observed using the marsh that are classified as wetland-dependent species (Online 135 Resource 1); only these wetland-dependent species were included in our subsequent analyses. 136

Local-level Habitat Characterization 137
We surveyed the vegetation at each wetland during peak aboveground biomass between 138 late July and August. First, we used a sub-meter accuracy GPS (Juno Trimble T41; SXBlue II 139 GPS/GNSS Receiver) to delineate the wetland boundary such that the perimeter of the wetland 140 lay where vegetation transitioned to <50% cover by wetland-obligate plant species. Next, we 141 sub-divided the wetland into zones based on vegetation form (woody vegetation, drawdown, 142 ground cover, narrow-leaved emergent, broad-leaved emergent, robust emergent, open-water 143 area) and the associated dominant or co-dominant macrophyte species. These vegetation zones 144 were delineated in the same manner as the wetland and their area calculated in the field to inform 145 quadrat-based sampling intensity. Each vegetation zone was then characterized by a minimum of 146 five 1 m 2 quadrats. If a zone was larger than 5000 m 2 , we added one quadrat per 1000 m 2 . 147 Finally, we estimated the mean percent cover among plots and then relativized our estimates to 148 100%, for a site-level estimate of vegetation cover. 149 In addition to vegetation surveys, we monitored abiotic variables known to influence 150 bird habitat selection. From May and September 2014, we measured water depth using staff 151 gauges, providing ponded water remained in the wetland. This was used to calculate the 152 wetland's maximum water, minimum water depth, and seasonal amplitude (maximum depth 153 minus minimum depth). 154

Statistical Analysis 155
Our analysis objectives were 1) to test whether the birds grouped into distinct 156 assemblages; and 2) to determine which local-and landscape-level variables were most 157 predictive of bird assemblage occurrence by developing a model to classify wetlands in terms of 158 their expected bird assemblage based on local-and landscape-variables, with particular emphasis 159 on the level of agricultural activity surrounding each wetland. 160 We used a square-root transformation and relativized our wetland-dependent bird count 161 data by the maximum value in each column to improve multivariate normality and reduce the 162 influence of numerically-dominant species. To reduce data sparsity, we removed rare avian 163 species (<2 occurrences out of 72 wetlands). Following the recommendations of McCune and 164 Grace (2002), we used a Bray-Curtis dissimilarity measure to characterize distances in species 165 space of community composition among our wetlands. 166

Cluster & Indicator Species Analysis 167
We used a cluster analysis to identify distinct wetland-dependent bird assemblages 168 among our sites. We used a hierarchical agglomerative polythetic process for the cluster analysis 169 After implementing each NMDS, we used vector overlays to visualize how species 193 counts (r 2 > 0.2 with at least one axis) and counts of species possessing various functional traits 194 aligned (r 2 > 0.1 with at least one axis) with major trends in bird community composition. We 195 symbolized sites by assemblages identified in the combined cluster analysis and ISA. 196

Classification and Regression Tree 197
Finally, we developed a classification and regression tree to predict which wetland-198 dependent bird assemblage would occur at a marsh, using a combination of local-and landscape-199 level data. In our case, the classification and regression tree partitions the wetlands based on 200 local-and landscape-level characteristics to create nodes of wetlands such that the deviance 201 between node membership and bird assemblage cluster is minimized. We used local-level (size, 202 percentage cover of woody, robust emergent and broad leave plants, maximum water depth) and 203 landscape-level variables (percentage cover of grassland, forest and shrubs, water and wetlands, 204 cropland and human-related land use within a 500 m radius) that would be critical in influencing 205 the functional traits of wetland-dependent birds present in a wetland, as predictors in the 206 classification tree. 207 We used the "tree" package (Ripley 2016) in R statistical software (R Core Team 2017), 208 to implement the classification and regression tree. The classification tree implements binary 209 recursive partitioning, using the deviance index described in Breiman et al. (1984) to estimate 210 impurity for splitting, and stops splitting when the terminal node passes a size threshold for the 211 number of wetlands included (Ripley 2016). Next, we used k-fold cross-validation to prune the 212 tree, where k = 10, which was based on cost-complexity as measured by deviance. We also used 213 the "tree" package to determine the number of misclassifications for the overall tree, as well as 214 the number of misclassifications at each node. Because our small sample size could contribute to 215 unstable k-fold cross-validation errors with increasing tree size, we repeated the test 100 times 216 and found the mean and standard error across iterations. 217 We used goodness of fit tests to measure if our classification and regression tree 218 predictions differed from the groups generated by the combined cluster analysis and ISA. Using 219 the DescTools package (Signorell 2017) in R, and a Williams correction for our small sample 220 size, we performed a G-Test. Next, we used the caret package (Kuhn 2017) in R to examine 221 whether there was strong agreement between our classification and regression tree predictions 222 and ISA assemblages, using kappa statistics. 223

Cluster & Indicator Species Analysis 225
We differentiated five distinct wetland-dependent bird assemblages (dendrogram in Fig. 2; 226 indicator values listed in Table 1), using agglomerative hierarchical clustering and ISA. We 227 assigned each assemblage a name reflecting the life history traits of the birds that were the 228 strongest indicators of the assemblage (indicator values listed in Table 1). Only a few species 229 were considered significant indicators of the five wetland-dependent bird assemblages. Note that 230 all but one wetland-dependent bird assemblage had at least one significant indicator species that 231 was both faithful and relatively exclusive to that assemblage of birds. The exception is the 232 Hummock Nesters (Table 1) in the cluster and indicator species analyses is presented in Table 1. 237

Visualizing Community Composition 238
Based on an assessment of the marginal decline in stress with increasing dimensionality, 239 we concluded that a three-dimensional solution was optimal for both our NMDS ordinations. The 240 NMDS, final stress was 18.58, and the NMDS converged in fewer than 20 runs. 241 The abundance of wetland-dependent birds differed among assemblages, based on their 242 nesting or habitat preferences (Fig. 3) with the Wetland Edge Nester assemblage ( Fig. 3C; 3D). Wetlands classified as supporting the 246 Shorebird Assemblage contained abundant shorebird and ground nesting species (Fig. 3A; 3B). 247 The Hummock Nester-classified wetlands shared species with all assemblages except the shrub 248 associates ( Fig. 3A; 3B), and they were not strongly associated with any bird nesting or habitat 249 preferences ( Fig. 3C; 3D). Conversely, only marsh (e.g. Sora) or pond species (e.g. American 250 Coot) were associated with the Pond and Reed Nesters assemblage ( Fig. 3A; 3C). 251 The NMDS axes were indicative of various local-and landscape-scale wetland 252 characteristics. Axis one in the NMDS reflected a disturbance gradient (Fig. 3C; 3D), wetlands 253 differed in which Natural Region they were located along axis two (Fig. 3C) and axis three 254 reflected a hydroperiod gradient (Fig. 3D). 255

All Birds 257
Using a combination of local and landscape-level variables (comprehensive list of 258 variables in Online Resource 2), we predicted which of the bird assemblages would occur at a 259 given wetland. The classification tree had ten terminal nodes (Fig. 4), with low total residual 260 deviance (12.48) and a misclassification error rate of 27%. Based on 10-fold cross validation 261 error, we trimmed the tree from ten (cross-validation error = 60%) to eight (cross-validation error 262 = 59%) terminal nodes (Table in Online Resource 3). The pruned tree had a marginally higher 263 total residual deviance (12.94), but the same misclassification error rate (27%). 264 The pruned model predicted all five assemblages. The model predictions had strong 265 agreement with the assemblages from the ISA (kappa = 66%). More, any differences between the 266 classification tree predictions and the observed assemblages were not statistically significant (G-267 Test: G = 10.19, df = 63, p-value =1.00). 268 The Wetland Edge Nesters and Shorebirds assemblages were the most distinct (in local 269 and landscape characteristics), occurring in a single terminal node (Fig. 4). The other 270 assemblages each occurred in two terminal nodes but differed in the distances between nodes 271 (Fig. 4). The Shorebirds assemblage was the third-most distinct assemblage, though predicted at 272 different tree heights. The Pond and Reed Nesters had the furthest vertical distance between the 273 nodes. However, the Hummock Nester-classified wetlands were predicted in both regions, 274 suggesting they were the least distinct assemblage. 275 Misclassification error rates were moderate (0 -46%) ( Table 2). Error rates were highest 276 for adjacent assemblages (e.g. Wetland Edge Nesters vs. Hummock Nesters), supporting birds 277 with similar foraging and nesting preferences. Conversely, error rates were low for assemblages 278 that were restricted to a region (e.g. Shorebirds vs Pond and Reed Nesters) ( Fig. 4; Table 2). 279 Local wetland characteristics were most predictive of assemblages (Fig. 4). Similar to our 280 analysis on all birds using a wetland, Natural Region was the strongest predictor of assemblages. 281 Apart from Hummock Nesters and Wetland Edge Nesters, the assemblages were predicted by 282 proxies for wetland hydroperiod (e.g., size, depth) and vegetation characteristics (e.g., robust 283 emergent vegetation cover, broadleaf vegetation cover). sensitive to the proportion of agricultural land cover in the surrounding landscape. We found 295 some support for this prediction -one of our five wetland-dependent bird assemblages were 296 absent from wetlands in agriculturally-dominated landscapes. We attribute this to differences in 297 the nesting behaviors of this assemblage -they nested in upland habitat and consequently 298 selected for landscapes with more natural land cover, mainly grassland. Conversely, dabblers, 299 divers, and waders were indicative of assemblages in landscapes with more agricultural activity 300 and longer hydroperiods. Consequently, waterfowl and shorebirds are less sensitive to this land 301 conversion, and so come to dominate wetlands situated in agricultural landscapes. However, the 302 importance of surrounding land cover in predicting which wetland-dependent assemblage would 303 occur at a wetland was less evident than we anticipated. 304 The most significant predictor of bird assemblage occurrence was the Natural Region that 305 the wetland fell in -Parkland vs. Grassland. The Grassland and Parkland natural regions differ in 306 both their landscape-and local-level characteristics. At the landscape-level, the Parkland 307 supports copses of aspen forest and more shrubland than the Grassland (Downing and Pettapiece 308 2006). Further, while there is more cropland in the Parkland, pastureland and haying are more 309 common in the Grassland (Downing and Pettapiece 2006). Also, because of differences in 310 climate, we observe a higher abundance of wetlands with longer hydroperiods in the Parkland 311 (Government of Alberta 2014a). In the more arid Grassland, the magnitude of difference 312 between potential evapotranspiration and precipitation is larger, resulting in the dominance of 313 shorter-hydroperiod wetlands (e.g., temporary and seasonal). Thus, we likely find Natural 314 Region to be a strong predictor of avian assemblage occurrence because of the preference of 315 some bird species for shorter-hydroperiod wetlands in mixed-grass prairie typical of the 316 Grassland (e.g., Shorebirds) versus the preference of other bird species for longer-hydroperiod 317 wetlands in landscapes with more shrubland and forest typical of the Parkland (e.g., Pond/Reed 318 Associates). 319 The composition of the landscape surrounding a wetland is not strongly predictive of 320 which wetland-dependent assemblage we find occupying a wetland. Local-level factors such as 321 hydroperiod (wetland permanence class, depth and size) and vegetation characteristics 322 These assemblages were characterized by birds that dive, dabble, and wade to feed. More, these 331 assemblages were predicted to occur in wetlands that were deeper, larger, and had longer 332 hydroperiods and nearly all their indicator species were ground, pond or reed nesters that nest in 333 the wetland proper. Thus, we conclude that these assemblages were most sensitive to in situ 334 factors about the wetland, rather than the character of its surrounding landscape. 335 Waterfowl and shorebirds may dominate wetlands in landscapes heavily influenced by 336 agriculture not necessarily because they profit from cropping and grazing activities, but because 337 species reliant on upland habitat for nesting are excluded. Both the Wetland Edge Nesters and 338 Hummock Nesters assemblages were predicted to occur in deeper (>0.53 m), larger (>10745 m 2 ) 339 wetlands in the Grassland, but it was the Hummock Nester assemblage that occurred in wetlands 340 with higher cropland activity in the surrounding landscape (>42 %). Consequently, species 341 belonging to the Hummock Nester assemblage come to dominate these wetlands because their 342 nesting habitat is still available when upland habitat is lost to agriculture, while wetland birds 343 that typically nest in upland habitat are now unable to do so (e.g., species the Wetland Edge 344 Nester assemblage). Similarly, Anderson and Rooney (2019) reported that significant differences 345 in bird community composition between natural and restored wetlands in the Parkland region of 346 Alberta were only evident when all birds were considered. They also reported that any difference 347 in the composition of wetland-dependent birds were negligible because restored wetlands were 348 similar to natural wetlands in their size, hydroperiod, and vegetation zonation, but differed 349 significantly in terms of landscape context. Therefore, by using a more comprehensive bird 350 survey data, we can develop bird-based wetland monitoring and assessment tools that reflect the 351 community-wide impacts of land cover change on bird assemblage occurrence. 352 Our CART and NMDS can be useful tools in designing wetlands for wetland-dependent 353 birds. Though the species pool did differ between Natural Regions, landscape composition can 354 be important when designing wetlands for birds. For example, if a practitioner aimed to provide 355 habitat for a Shorebird assemblage in the Grassland Natural Region, the wetland should be deep 356 (> 0.53 m) or large (>10745 m 2 ) (i.e., CART results) and have lower human activity (i.e., NMDS 357 results). However, if the said practitioner was targeting Wetland Edge Nesters, the wetland can 358 be smaller (<10745 m 2 ) but should have moderate to low cropping activity in the landscape 359 (<42.9 %). 360

Conclusion 361
We show that, generally, wetland-dependent assemblages show poor sensitivity to 362 agricultural activity. While waterfowl and shorebirds were sensitive to in situ properties of the 363 wetland, such as water depth and wetland size or vegetation zonation patterns, edge-nesting birds 364 were excluded from wetlands with higher cropping activity. Waterfowl and shorebirds seem to 365 dominate wetlands in landscapes with more agricultural activity because other avian species are 366 excluded, despite being at greater risk of predation in these landscapes (Emery et al. 2005). 367 SWhen designing wetlands for use by these wetland avifauna, our concurrent analyses using a 368 CART and NMDS are useful tools in determining the landscape context and wetland 369 characteristics suitable for assemblages that may be the target in restoration policy.  Table 1 This table provides indicator values for species belonging to each of the five bird  assemblages identified via cluster analysis of the dataset including only birds categorized as  wetland-dependent species. Each species is grouped under the assemblage for which it had the  highest indicator value, and the table includes all 38 species, regardless of whether it was a significant indicator of an assemblage. However, only 13 species were significant indicators (p<0.05), indicated by "*". The associated p-value indicates the probability that an indicator value that large could be obtained from the data by chance alone. Note that the Hummock Nester assemblage was the first assemblage to merge (into the Pond and Reed Associates assemblage) during agglomerative clustering analysis ( Figure 2) and had no significant indicators.   Fig. 1 A map of study our region in the northern prairie pothole region. Our 72 wetland sites occupied both the Grasslands and Parklands region, belonging to temporary (n=11), seasonal (n=18), semi-permanent (n=10), and permanence (n=9) permanence classes. Birds categorized as terrestrial species were excluded from this analysis (see Online Resource 1).
Symbology of the sites at the tips of the dendrogram reflects the optimal dendrogram pruning level, determined using indicator species analysis. Group names were based on the life history traits of the dominant indicator taxa for each group.  Online Resource 3 Cross-validation error for the classification tree for birds categorized as wetland-dependent species, based on the number of terminal nodes. We found the mean and standard error for cross-validation across 100 iterations.