The use of remote cameras to study animal populations remains active in ecology. In the Applied Wildlife Ecology (AWE) lab, we deploy cameras across Michigan to understand the distribution and activity patterns of mammalian carnivores including species of conservation concern (gray e.g., wolves). However, after our laborious field efforts associated with the camera survey then emerges the new challenge of processing millions of the resultant images. We have to manually sort images that have no animals, non-target animal species, or target species, quantity individuals, and describe behavior. To assist, we created Michigan ZoomIN on the Zooniverse platform, a virtual citizen science project to crowd-source identification of images. Each image is classified by different users before “retirement” and then we employed various consensus algorithms to obtain a final classification with systematic confirmation from research team. Here, we obtained images from surveys conducted October 2015 – December 2016 and compared four consensus methods obtained from user-classified data through Michigan ZoomIN. Methods included: 1) full agreement with 100% of users choosing the same classification, 2) Cohen Kappa above 0.60 indicating inter-user reliability, 3) Cohen Kappa with Pielou’s evenness coefficient above 0.50 indicating high accuracy, and 4) majority classification if an expect identification was included.
Results/Conclusions
Thanks to participation from >3,000 volunteers, we obtained >655,305 classifications from season #1. We concluded consensus for 40310 images with ~15% false positives (i.e., empty image). The contribution across methods was: consensus (80%), Kappa and evenness (15%), Kappa only (4%), and majority with expert (1%). Despite high user-agreement, consensus methods were species-specific with carnivore consensus relying more on Kappa and evenness statistics. The remaining 3,377 images (8%) require further review. For these images, both Pielou’s evenness index (0.80 ± 0.01) and Kappa coefficient (0.39 ± 0.02) were higher than thresholds set for agreement. Moreover, no consensus from user-classified data were obtained for beaver (Castor canadensis), cougar (Puma concolor), gray fox (Urocyon cinereoargenteus), mink (Neovison vison), marten (Martes americana), and Canada lynx (Lynx canadensis). We will revise our educational material on Michigan ZoomIN to improve future classifications of these “difficult species”. Ultimately, Michigan ZoomIN proved an effective tool for engaging citizen scientists and reduced our workload for image processing (though time for development was substantial). Despite less reliable identifications for our target species, filtering out false-triggers (5,931 images) and deer (22,124 images) represented a significant contribution from our volunteers to investigating carnivore community throughout the state of Michigan.