Monitoring fish spawning sites in freshwater ecosystems using low-cost UAV data: A case study of salmonids in lakes in Iceland

Low-cost unmanned aerial vehicles (UAV), widely known as drones, have become ubiquitous and improved considerably in their technical capabilities and data quality. This opens new opportunities for their utilisation in scientific research that can help to reduce equipment and data collection costs. Remote sensing methods in ecological fieldwork can be a suitable approach to complementing, augmenting or even replacing certain aspects of fieldwork. In this study we tested the suitability of UAV for the detection of salmonid spawning grounds in two lakes in Iceland, Thingvallavatn and Ellidavatn. Salmonids are very susceptible to environmental changes, especially during embryonic development when highly oxygenated water flow and low temperatures are required. Monitoring the changes of the redd density over time will help understand the population dynamics of salmonid species and create strategies for species conservation. As part of this pilot study, we conducted aerial surveys during the spawning seasons in both locations in 2018 recording standard photographs in the visible spectrum (red, green, blue) to fully cover the respective areas of interest. Different flight altitudes were recorded to test the effects of image resolution on the final analyses. The images were then processed by applying standard remote sensing analyses in the software ENVI. The maximum likelihood classification combined with post-classification improvement methods resulted in satisfactory accuracies that are valuable for further monitoring efforts. From these findings we established a workflow that allows the implementation of UAV in ecological fieldwork for regular long-term observations. We discuss our experiences with regards to their utility, their limitations and identify future directions of research for implementing the potential that low-cost approaches have in supporting ecological studies of freshwater ecosystems.


Introduction
Salmonids are cold temperate species that are expected to be negatively impacted by the 35 warming of aquatic habitats and the correlated oxygen depletion due to climate change 36 Highly oxygenated water flow and low temperatures are required for salmonids' successful 43 embryonic development, making them especially vulnerable during this time (Moyle, 2002). showing that our current climate is changing more rapidly than before (Ficke et al., 2007). 49 Due to their long generation time (Willson, 1997), the monitoring of salmonid populations 50 during their spawning and subsequent embryonic development of their offspring would 51 provide information on the effects of climate change much faster than the data that would be 52 otherwise gathered on later life stages. 53 Most species belonging to the Salmonidae family are characterised as gravel nests 54 spawners (Nika et al., 2011). Due to their appearance as irregular or regular shapes that 55 contrast with the undisturbed area surrounding them (Bjornn & Reiser, 1991), the spawning 56 redds (i.e. nests) are visible from the air. This makes monitoring of the spawning redds from 57 the shore and directly above the water possible. Although redd count data are an important 58 source of information for management purposes, including monitoring population size and 59 estimating carrying capacity of spawning habitats (Rieman & Mcintyre, 1996), redd estimates 60 are predominantly done by manual counting from the shoreline. This method is often 61 to the inherent high variability of the method and the inability to compare results, the 70 shortcomings of the manual redd count make monitoring using this method rather difficult 71 (Schuett-Hames et al., 1996). 72 The use of unmanned aerial vehicles (UAV), also known as drones, for counting salmonid 73 redds is being explored as an answer to these shortcomings (Groves et al., 2016). Groves  applied machine learning and object-based classification techniques to UAV-based imagery 85 to assess the use of both RGB and hyperspectral imagery. They found that both sources of 86 imagery could be used to identify redds, but with varying degrees of accuracy. Despite the 87 current development of techniques that use remote sensing to detect redds, the application 88 of such techniques is not yet an integral part of large-scale monitoring programmes. 89 The implementation of UAV in monitoring programmes requires the establishment of routines 90 and procedures that can easily be applied and adapted to specific project needs. The cost of 91 such efforts must also be taken into account in order to keep monitoring efforts manageable 92 and effective. This gives low-cost UAV an advantage, and provides further benefits due to 93 their quick and relatively easy data collection (Calvario et al., 2017). The goal of this study 94 was to develop an easy, accessible and low-cost method to map salmonid spawning redds. 95 We created a pipeline for detecting salmonid redds using a semi-automated approach by 96 applying a pixel-based classification technique to UAV derived imagery. In this study, we 97 present the pipeline developed to detect salmonid spawning grounds which was tested at 98 two different Arctic charr (Salvelinus alpinus) spawning grounds in Iceland. 99 represent contrasting environments (Table 1) The study was realised in four main stages which are described below: The initial data 146 acquisition at the first case study area was followed by the main data processing steps which 147 were then concluded with an accuracy assessment to evaluate the quality of the results 148 ( Figure 1). Once a successful workflow had been successfully established, the procedure 149 was transferred to our second case study area for validating the method. in Thingvallavatn. Data collection was undertaken using a DJI Mavic Pro drone (Table S1)  The software ENVI version 5.1 (Exelis Visual Information Solutions, Boulder, Colorado) was 184 used for processing and classification analysis. For data processing, only images with the 185 least sunlight reflection and wind ripples were selected. Eight supervised image classification accuracy when classifying spawning redds, it was decided that this approach would be taken 189 forward for further analyses in this study to improve the results specifically for the redds 190 feature class. The maximum likelihood classification method assigns individual pixels to the 191 class with the highest a posteriori probability under the assumption of a normal distribution 192 (Richards & Jia, 1999). When using this method in ENVI, no probability threshold was set to 193 classify all pixels in the image. 194 Three post-classification methods were applied in order to improve the accuracy by 195 correcting isolated or misclassified pixels. First, the classified output was filtered by removing 196 spurious pixels with a majority-minority analysis (Gurney & Townshend, 1983). This analysis 197 changes spurious or "false" pixels to the class value that the majority of the pixels in the 198 manually indicated kernel belong to. The following parameters were selected: majority for the 199 analysis method, a kernel size of 3, and a centre pixel weight of 1. In addition, isolated pixels 200 were corrected using the sieve classes method (Exelis Visual Information Solutions, 2009; 201 Su, 2016). This method looks at neighbouring pixels to determine if a pixel is grouped with 202 the same endmember classes surrounding the pixel. For this study pixel connectivity was set 203 to four and the minimum size to two. Lastly, the clump classes method was applied to clump 204 similarly classified areas together adjacent from each other (Su, 2016). Following a visual 205 examination of the initial classification results, the size parameter for this method was set to 206 three. 207 Overall accuracy of the applied supervised algorithm is expressed as the percentage of 208 correctly classified pixels of all endmember classes. Producer's accuracy (PA) is defined as 209 the probability that each endmember class is classified correctly. User's accuracy (UA) is 210 defined as the probability that the classification map represents the ground truth data. 211 Furthermore, the kappa coefficient was used to evaluate the classification accuracy and can 212 be interpreted as a value ranging from 0 to 1 that explains the difference between the 213 observed classification of the endmember classes and the reference data (Campbell, 2002). 214 215

216
The method to classify spawning redds from UAV imagery was established using the well-217 studied Ólafsdráttur spawning grounds in Thingvallavatn where the contrast of the spawning 218 redds was high and reflectance similarities were low. Aerial images were taken at altitudes of 219 classes and the 100 m imagery due to reflectance and low accuracy results. The 50 m imagery was deemed suitable for the classification methods after a visual examination, as 224 these images provided the optimal contrast and lowest environmental reflectance in relation 225 to the spawning grounds as the main features of interest in this analysis. 226 227 Figure

258
An accuracy assessment was performed using the selected training data and test data to 259 estimate the quality of the classification results using the maximum likelihood classification 260 method after post-classification improvements were applied ( Table 2). The classification and 261 the ground truth data were combined in an error matrix to estimate the number of correctly 262 classified pixels. Most importantly, the producer's accuracy for classifying spawning redds 263 was 90.68% and 97.07% for the user's accuracy. While the spawning redds were mostly 264 correctly classified, 26 pixels were classified as underwater rocks, 8 pixels as shoreline, 5 265 pixels as deep water, and 1 pixel as surface rocks. 266

270
The above results are all reported after applying post-classification methods. The addition of 271 these classification improvements raised the overall accuracy of the maximum likelihood 272 classification from 91.27% to 93.40% (Table S2). to each other. In the UAV image this group can be recognised due to its light colouring. Class 285 density and location of the class. Furthermore, some pixels are located in a small shaded 289 part of the land area. Class 5 (ice; in blue) consists out of a high-density main group of pixels 290 located in the North-western edge of the image with an irregular shape and furthermore, few 291 pixels distributed not larger than a few pixels together. The UAV image shows that this class 292 is fully distributed on land. Lastly, class 6 (human structure; in white) is very heterogeneous 293 and does not cover a large area. The class has a main group of pixels in a square shape and 294 some more pixels spread out in small clusters consisting of only a few pixels. The UAV 295 image shows this class located on land recognised by a grey colour.

300
An accuracy assessment was performed to estimate the quality of the classification results 301 using the maximum likelihood classification method (Table 3). The classification and the 302 ground truth data were combined in an error matrix to estimate the amount of correctly 303 classified pixels. The producer's accuracy for classifying spawning redds reported 83.41% 304 and 58.83% for the user's accuracy. Misclassified pixels were classified as either human 305 structures, underwater rocks, or aquatic vegetation. 306

310
To obtain the best results using the maximum likelihood classification method the post-311 classification methods, majority-minority analysis, sieve classes method and clump classes, 312 were applied after classifying the pixels. The addition of these classification improvements 313 raised the overall accuracy of the maximum likelihood classification from 82.48% to 86.40% 314 (Table S2). 315 316

Discussion 317
This study aimed at developing an easy, accessible and low-cost method for mapping 318 salmonid spawning grounds in shallow freshwater areas. The analysis of UAV-derived 319 imagery in the contrasting case study areas situated in the subarctic region showed a 320 successful application of a pixel-based classification method that was able to identify 321 spawning redds from RGB imagery with high accuracy. Imagery taken from a height of 50 m 322 was deemed suitable for classifying spawning redds after which fifteen training areas were 323 selected for each spectral class with an average of 120 pixels each. The maximum likelihood classification method was run with resulting high accuracies for producer's and user's accuracy. By correcting isolated or misclassified pixels three post-classification improvement 326 methods were applied to improve accuracy: majority-minority analysis, sieve classes method 327 and clump classes. 328 The use of remote sensing for mapping and monitoring spawning grounds provides 329 improvement over traditional means of monitoring due to the method being safer (Groves et  dredging lake bottoms or exposing them to intolerable fluctuations of the water level 367 (Maitland, 1995), and the construction of dams that make spawning grounds inaccessible for 368 anadromous salmonids (Quiñones et al., 2015). In the case of our study system 369 Thingvallavatn, a small dam was built in 1959 on the southern outlet of the lake 370 (Sturlaugsson & Malmquist, 2011). The outflow of the lake had to be directed through a short 371 tunnel leaving the former outlet mostly dry. The installation of the dam and the tunnel had a 372 significant impact on the brown trout population of the lake. It is believed that a major 373 spawning site at the outlet was spoiled and the drying-up of the former outlet ruined an 374 extremely productive community of blackfly larvae (Simulium vittatum) that formed optimal 375 feeding grounds for juvenile trout (Malmquist, 2011). These however are only speculations 376 since the massive decline of the brown trout population came to light more than 20 years 377 later when in the 1980s during a lake-wide survey only a handful of trout specimen were 378 caught (Sandlund et al., 1987). These are just a few examples on how human impact has 379 negatively influenced spawning habitats and while in some places action has been 380 undertaken to, for example, remove dams to mitigate these effects, the removal is unlikely to 381 restore these ecosystems the way they were previously since the negative effects of the dam Our approach to detect spawning grounds has the potential for future use in complementing 387 fieldwork that aims at monitoring changes in salmonid spawning grounds in different 388 environmental conditions. Scaling this method to other freshwater ecosystems would allow to 389 assess the full potential of this method in detecting spawning redds from a range of different 390 species. We show the successful application of this method in two very contrasting lake 391 environments which indicates the promising potential of this low-cost method. 392 Remote sensing is a helpful tool for monitoring from small to large spatial and temporal 393 scales. Only relying on scientists and concerned institutions to gather data will be the limiting 394 factor when a lot of data is needed (Hunt et al., 2017). help gather information for analysis when the project is well conceptualised. Our method to 407 detect spawning grounds gives the opportunity to make use of crowdsourcing to gather 408 images of spawning grounds. The images for this study were taken using a commercially 409 available drone at a height of 50 m during specific weather conditions to minimise reflectance 410 that could interfere with classification. In principle an amateur drone pilot should be able to 411 accomplish this without complicated technical guidance. The ubiquitous existence of UAVs, 412 not least in a popular tourist destination like Iceland, provides a solid basis for crowdsourcing 413 more data to further assess the applicability of our method and to extent the observations 414 from our case study areas to other freshwater systems. 415 For further monitoring purposes, a framework building on the capabilities of Geographic 416 Information Systems (GIS) will need to be developed to create a spatial database from the 417 remote sensing analyses. This provides the basis for establishing change detection 418 procedures that help to assess the spatial changes in spawning grounds on a temporal 419 scale. Additional geographical observations from the shorelines along the lake are also 420 derived from the remote sensing analysis as basic indicators of the different endmember 421 classes (e.g., aquatic vegetation, vegetation, shoreline). This information provides additional 422 useful information for a GIS database which will be valuable for geomorphological mapping 423 of the study sites to further understand the ecosystem and see how its geography is 424 changing beyond the nature of the nesting sites. change and other stressors on the habitat of salmonid species and the freshwater ecosystem 432 as a whole. This study shows the suitability of using low-cost UAV and pixel based 433 classification approaches in remote sensing to detect salmonid redds in contrasting 434 environments. The developed method has potential to be scaled to other freshwater 435 ecosystems and salmonid species to assess the impact of anthropogenic stressors on a 436 wider scale and to develop monitoring procedures that enable the long term observation of 437 the changes that occur in these highly fragile environments. 438 the potential for spectrally based remote sensing of salmon spawning locations. River