A novel deep-learning analysis of multiple wildlife imagery datasets for cross-species generalization

Link To Share This Poster: https://cdmcd.co/Yk3dmp

Thomas Y. Chen, Academy for Mathematics, Science, and Engineering, NJ

Background/Question/Methods
As human activity across the planet increases, wild species are increasingly losing their natural habitats, causing population decreases, loss of genetic variation, and therefore their progression to becoming endangered and even extinct. 85% of imperiled species suffer from this habitat loss phenomenon, notwithstanding the pollution, climate change, invasive species, and excessive hunting and fishing that also contributes to declines in populations. The IUCN Red List, a gold standard for official species designation and biodiversity loss mitigation responses that ensue, lists 19,817 species as “threatened,” 3,947 as critically endangered, and 5,766 as endangered. In order to rehabilitate endangered and critically endangered species and to prevent threatened species from attaining these statuses in the first place, accurate and efficient mechanisms must be in place to identify individuals. Particularly, deep-learning neural networks have been used for automated identification and analysis, namely convolutional neural network (CNN) architectures such as AlexNet, VGGNet, and ResNet, pretrained on ImageNet. Classification efficacy and efficiency has been greatly improved by deep learning, as compared with conventional methods. Many datasets containing camera trap wildlife imagery have been published, each containing different combinations of species categories. The Labeled Information Library of Alexandria: Biology and Conservation (Lila BC) is an open data repository housing many such datasets, including “Great Zebra and Giraffe Count,” “Whale Shark,” “BIRDSAI,” “Amur Tiger Re-identification in the Wild,” etc.
Results/Conclusions
Dataset diversity is not only important for more powerful, accurate data analysis but is also associated with more ethical practices in data science. Therefore, in this work, we develop a convolutional neural network trained on 10 datasets within the Lila BC catalogue, consisting of camera trap imagery of animals from multiple continents. For our baseline model, we use the ResNet50 architecture, pretrained on ImageNet data and use the Adam optimizer with a learning rate of 0.01. We train for 100 epochs on Tesla K80 GPUs. Preliminarily, we achieve a 0.85 weighted F1-score on all species/taxa categories. While this is relatively high, future work includes investigating the individual F1-scores of each taxon category. This further study will reveal deficiencies in the machine learning model we developed and lead to improvements in model specifications or changes in architecture. Ablation studies will be conducted. This research allows for gaining insights into population trends and individual conditions, subsequently inspiring directed resource and personnel allocation for aid in conservation, which is a major necessity for successful endangered species recovery.