2023 AMSS Abstract Book

Bering Sea | Mammals

Improving accuracy of automated the Northern fur seal composition count on Tuleny Island, Russia, 2017-2022 Presenter: Ivan Usatov , Usatov.ivan.Alex@gmail.com, Kronotsky Biosphere Nature Reserve Anna Kirillova , canis7@yandex.ru, National Park the Commander Islands Egor Vasyukov , egor.vasyukov@list.ru, Kamchatka Branch of the Pacific Geographical Institute FEB RAS Alexey Altukhov , aaltukhov@gmail.com, North Pacific Wildlife Consulting Russel Andrews , russel.d.andrews@gmail.com, Marine Ecology and Telemetry Research Vladimir Burkanov , vladimir.burkanov@noaa.gov, Marine Mammal Laboratory, Alaska Fisheries Science Center Thomas Gelatt , tom.gelatt@noaa.gov, Marine Mammal Laboratory, Alaska Fisheries Science Center Over the last five years, we have been developing a computer vision algorithm to automatically count Northern fur seals (NFS) from aerial photos at Tuleny Island, Russia. Between 2017-2021 we conducted 201 aerial surveys using a DJI quadrocopter. To distinguish animals from the background we used a U-Net model and trained it using 23,638 hand- prepared tiles to make able to count all non-pup NFS with a median deviation of only 2.6% (min -5.0%; max=7.4%) from the count generated by observing the images. Age and sex composition counts were less accurate, and we discovered high false negative and false positive errors which cancel such that the total non-pup count was closer to the actual count than if the bias was one-tailed. We developed two VGG16 models to reduce the false identification rate and two other VGG16 models to better separate animals by age and sex. To test the new models, we used 29 new aerial surveys conducted in the summer of 2022 which weren’t used for model training. Incorporation of the new models reduced the error rate for non-pup counts to 1.2% (min -12.4%; max 7.9%). Female counts improved slightly from -3.3% (min=- 44,4%; max=108.3%) to -3,0% (min =-28.0%; max=3.9%). Median count for harem males changed from -1.1% (min=- 47.7%; max=65.8%) to -2.0% (min -27.5%; max=30.0%); for idling males from 20.3% (min=-12.5%; max=320.0%) to 27.8% (min=-22.2%; max=118.0%); and for juvenile males from -9.5% (min=-33.8%; max=118.1%) to -32.2% (min=-40.7%; max=37.9%). The incorporation of this new automated count algorithm improved the results by reducing the error rate for important age and sex groups historically used for NFS trend analysis, the harem males and females. However, the variation is still high and additional refinement is necessary.

Alaska Marine Science Symposium 2023 267

Made with FlippingBook flipbook maker