Streamlining ringed seal (Pusa hispida) abundance estimation through Computer Vision models

The Arctic is warming faster and experiencing a greater degree of physical transformation than other parts of the planet. Monitoring the status of wildlife populations under these conditions is essential to track the impacts of these changes on biodiversity conservation, environmental management and sustainable use of natural resources. Assessments of Arctic endemic marine mammals are particularly important because their sea ice habitats are declining precipitously and they are key resources for human populations in the Arctic. However, the remote nature of their habitats makes monitoring efforts logistically complex and financially costly. Herein, we present the development, testing and validation of a deep learning model applied to aerial imagery of ringed seals ( Pusa hispida ) in coastal sea ice environments. We conducted uncrewed aerial systems (UASs) surveys in several fjords in western Spitsbergen, Svalbard (Norway) and generated an annotated high-resolution dataset to train an object detection model - based on the YOLOv8 architecture. The results demonstrated the effectiveness of this model, which achieved an F1-score of 0.89 and a [email protected] of 0.90 on its own test set. When evaluated on a larger, external test set, the model was able to detect 96% of the ringed seals present when exposed to new images. Analysis of the efficacy of the model suggests that the human workload could be reduced by up to 99.7% from manual post processing analyses of images. This study demonstrates that our model is an effective tool for rapid, accurate and low-cost automated monitoring of ringed seals in aerial images from UAS surveys that has the potential to be useful for other sea-ice associated species after specific target-species training.