New Technology II

Contributed Paper
ROOM: HCCC, Room 26A
SESSION NUMBER: 62
 

8:10AM Understanding How Landscape Features Affect Gene Flow: Advances in Resistance Surface Optimization for Landscape Genetic Studies
William E. Peterman; Kristopher Winiarski; Kevin McGarigal
Movement and dispersal are essential to the long term persistence and viability of populations, but habitat loss and fragmentation threaten these processes and present challenges to the management of species on the landscape. Further, movement is often difficult or impossible to directly observe, necessitating indirect measures, such as genetics, to infer successful movement of individuals across the landscape. Landscape genetics has emerged as a field especially suited to questions related to spatial population genetic processes. From its inception, an allure of landscape genetics was the potential to use spatial genetic data to determine how landscape features affect gene flow. However, it has remained exceedingly challenging for researchers to objectively determine how landscape features affect movement and gene flow across landscapes. To address this issue, the R package called ResistanceGA (Resistance optimization using Genetic Algorithms) has been developed, which provides a framework for conducting unbiased analyses of landscape surfaces to determine their effect on gene flow. Through the use of genetic algorithms, ResistanceGA is able to optimize the resistance values of categorical (e.g., land cover, roads) and continuous (e.g., temperature, slope) resistance surfaces, as well as multiple surfaces simultaneously. Simulations revealed that accuracy increases with sample size, but accuracy decreases as variance in pairwise genetic data increases. ResistanceGA is capable of correctly identifying the data generating resistance surface even when highly correlated surfaces are present. Optimized landscape resistance surfaces can provide novel insights into how landscapes affect movement. Such information is invaluable to the spatial management of populations. Identifying how and which landscape features affect gene flow has been a formidable challenge facing researchers. ResistanceGA provides a framework for optimizing resistance surfaces, and is a viable solution to overcoming previously encountered challenges.
8:30AM See What You’ve Been Missing: an Assessment of Reconyx® Game Cameras
Rachael E. Urbanek; Holly J. Ferreira; Colleen Olfenbuttel; Casey Dukes
For camera trap studies to produce accurate data, cameras should reliably detect animals within their field of view. Reconyx® cameras are often touted as being the most effective for research, but we found that they have detectability issues. We reviewed 1,503,330 pictures across 3 North Carolina coastal game lands that were obtained during August-September 2016 and February-March 2017 using 36 Reconyx PC900 Hyperfire cameras. We pooled all independent timed and triggered events, regardless of species, and evaluated factors related to temperature, wind speed, moon phase, and which detection band(s) and zone(s) an animal was observed in that may have increased the likelihood of triggering any animal within each game land. We also analyzed the data by pooling across the sites for 10 species to determine species-specific reasons for not triggering the camera, regardless of site. Different variables affected whether the camera would trigger at each game land, but at all 3 sites the likelihood of triggering the camera increased with the more zones an animal was detected in (Likelihood ratio = 23.2-143 P <0.0001) and there were more triggers than expected if the animal was observed in both detections bands as opposed to none (χ2 = 26.0-94.5 P <0.0001). Triggering of the camera favored larger species (G8 = 24.5-123 P <0.0002). The trigger missed 14-16% of the independent events of large mammals and was overall ineffective at triggering for small mammals. Our results indicated 47-64% of independent events of mesocarnivores would have been missed if only the infrared-motion trigger was used. The lack of detection of these species is alarming because it would severely underestimate occupancy and abundance estimates of these populations. Regardless of brand, researchers should assess detectability of their target organisms when designing camera trap studies and use both infrared-motion trigger and time-delay options to improve detection rates.
8:50AM Wildlife Camera Image Identification Combining Deep Learning and Crowd-Sourced Identifications
Patrick Lorch; Matthew Sochor; Josiah Olson
Cleveland Metroparks has deployed a large network of wildlife cameras to monitor animal communities and activity throughout our 23,000 acre park system. This has generated over 6 million images. A two-pronged approach to identify animal species in the images involving Michigan State University undergraduate students and a crowd sourcing project on Zooniverse has failed to identify images faster than new ones come it. This led us to develop an artificial intelligence model involving deep learning neural networks called Transfer (https://github.com/matthew-sochor/transfer). The objective of this project was to identify the optimal way to combine a deep learning model and oversight by human volunteers. Initial training of the model on a test set of images identified by MSU undergraduates resulted in over 96% accuracy at identifying images with no animal and 95% accuracy at identifying animals to species. We then developed a more comprehensive training data set involving all cameras with identifications from both MSU and Zooniverse volunteers with over 200,000 images. After retraining the model, we will develop a way to use confidence information from the model to decide what images require further scrutiny by human volunteers. We expect confidence thresholds to depend on species and season, but may also depend on location and lighing conditions, among other factors. The results should inform all wildlife camera projects on how best to deal with identifying large volumes of images, as well as general ways to improve deep learning models.
9:10AM Hotspotter Is More Accurate Than Wild-Id in Identifying Individual Animals By Coat Patterns, Especially for High-Quality Camera-Trap Images
Robert B. Nipko; Brogan E. Holcombe; Marcella J. Kelly
Camera-trapping has become a widespread, non-invasive technique for monitoring wildlife populations. It is especially useful for species like jaguars (Panthera onca) and ocelots (Leopardus pardalis) with distinct spot patterns that can be used to identify individuals, allowing use of capture-mark-recapture techniques to estimate abundance. However, sorting photographs and identifying animals by eye is time-consuming and tedious. HotSpotter and Wild-ID are two freely downloadable software programs that match images of individual animals from trail camera data. With the proliferation of camera-trap studies, and the large quantity of data produced, there is a need for such programs that automate and accelerate identification. There has been little work comparing the effectiveness of such programs, thus our objective was to provide such an assessment. We built reference databases of jaguars and ocelots in HotSpotter and Wild-ID, then tested each program’s accuracy and efficiency in identifying individuals. We used 359 jaguar and 354 ocelot test images collected from camera stations in Belize from 2015-2016. Test images were ranked on a scale from 1 (low-quality images) to 3 (high-quality images). We found that HotSpotter was significantly more accurate than Wild-ID in identifying jaguar images of all qualities combined (77% accuracy, P = 0.0056), but this result was mostly driven by higher accuracy for high-quality images (96% accuracy, P = 0.0039). In comparison, Wild-ID was 68% accurate for all quality images combined and 85% accurate for high-quality images. HotSpotter was 81% and 51% accurate identifying medium- and low-quality images, respectively, compared to 74% and 41% for Wild-ID, but these differences were only marginally significant. Results suggest HotSpotter is more effective for identifying individual animals in remote camera data, but human discernment is still necessary for identifying lower-quality images. We discuss when pattern-recognition software provides advantages over manual identification and provide advice on building databases and potential pitfalls.
9:30AM A Geometric Estimator to Calculate Triangulations From Radio Telemetry Data
Landon R. Jones; Benjamin C. Colteaux; Paul L. Leberg; Scott M. Duke-Sylvester
Triangulating on transmitter signals is a common method to estimate animal locations. Lenth’s maximum likelihood estimators (LMLE), the most common method to calculate triangulations from radio telemetry data, perform well when telemetry operators surround the signal and angle bearings between triangulations are large. However, achieving these conditions in field studies can prove difficult, particularly for fast-moving animals or when terrain constrains operators to take bearings at points along one side of the animal. We created a simple geometric method and script in program R called TriAD to calculate triangulations based on two sources of error, 1) the distance from operators to transmitter signals, and 2) angle bearing range of the strongest signal for operators. Overlapping error polygons created by the variation in these two parameters provide signal location estimates that are resistant to erroneous bearings. To compare the accuracy of estimated locations from TriAD to LMLE, we used beacon tests and empirical data from two radio telemetry studies, one tracking two species of toucans (Ramphastos sulfuratus, Pteroglossus torquatus) and another tracking snapping turtles (Chelydra serpentina), where operators took bearings to transmitters while moving only short distances between mobile receiver points arranged along linear transects. Location estimates from beacon tests were 40 m closer to true locations for toucans and 65 m closer for turtles with TriAD compared to LMLE. Over 99% of 513 toucan triangulations and 100% of 54 turtle triangulations from beacon tests converged for TriAD compared to 93.2% and 33.3%, respectively, for LMLE. For empirical data, 98.0% of 4445 toucan triangulations and 93.3% of 15 from turtle triangulations converged for TriAD compared to 88.6% and 73.3%, respectively, for LMLE. Our results indicate that TriAD provides an alternative to calculating triangulations using LMLE that may improve location estimates when tracking fast-moving species or any species in difficult terrain.

 

Contributed Paper
Location: Huntington Convention Center of Cleveland Date: October 10, 2018 Time: 8:10 am - 9:50 am