Arctic | Mammals
Automated classification of walrus trail camera imagery: Towards in-season haulout monitoring
Presenter: Jeff Tracey , jatracey@usgs.gov, U.S. Geological Survey Alaska Science Center Janice Gordon , janicegordon@usgs.gov, U.S. Geological Survey Alaska Science Center Lydia Kieine , lydia_kleine@fws.gov, U.S. Fish and Wildlife Service Jonathan Snyder , jonathan_snyder@fws.gov, U.S. Fish and Wildlife Service Anthony Fischbach , afischbach@usgs.gov, U.S. Geological Survey Alaska Science Center
The Pacific walrus occupies the continental shelf of the Pacific Arctic. Walruses must rest out of water between foraging bouts. They rest on sea ice when it is seasonally available and on land when it is not. When walruses gather in large numbers to rest on shore at locations that are termed “haulouts”, they are at risk of large mortality events resulting from natural and human disturbances that cause stampedes. They are also at increased risk from vessel spill events because walrus distributions are so densely concentrated at and around coastal haulouts. To reduce these risks, managers require a means to monitor large coastal haulouts. Managers have stationed trail cameras at large haulouts to collect images throughout the haulout season. However, reviewing these images has only been possible after the season and has required substantial staff time. We developed convolutional neural network (CNN) models to classify walrus occupancy in time-lapse trail camera imagery frames collected at several Alaskan haulouts with varying local terrain and walrus attendance. We trained models based on four different CNN architectures (InceptionResNet50V2, InceptionV3, ResNet50V2, and Xception). Each model consisted of a feature extraction module that passed its output into a classification module that predicted the probability of walrus in the image. The weights of the feature extraction module were pre-trained on the ImageNet dataset (using “transfer learning”), and the weights of the classification module were trained on our training and validation datasets. Finally, we evaluated the performance of each model on an independent test dataset. The Xception and InceptionResNet50V2 models yielded the best performance for classifying walrus presence or absence (95.7 and 95.5 percent accuracy, respectively). In the coming season we plan to develop a data processing pipeline to classify satellite-linked trail camera imagery to deliver in-season monitoring to walrus managers.
Alaska Marine Science Symposium 2023 293
Made with FlippingBook flipbook maker