Insights and approaches using deep learning to classify wildlife

Miao and colleagues trained deep convolutional neural networks to classify species in camera-trap images from an African dataset and, importantly, used interpretability techniques to probe what features the models rely on for their decisions. They achieved high classification accuracy and showed that the networks often attend to ecologically sensible body parts, lending credibility to automated identification. By opening the black box, the study addressed a common concern that deep-learning predictions are unaccountable, and it offered guidance on building trust in automated systems. The work contributed both performance benchmarks and methods for model interpretation in ecological image analysis. It is a cited reference in the literature on applying and understanding deep learning for camera-trap species classification.