From pixels to conservation: Deep learning for automated monitoring of rare and endangered wildlife

Abstract

Summary Conservation of rare and endangered species requires scalable monitoring solutions to address accelerating biodiversity loss. Deep learning has transformed wildlife monitoring, enabling automated species identification, individual tracking, demographic inference, and behavioral analysis from camera traps, drones, and video surveillance. This review systematically analyzes 159 wildlife monitoring studies (2014–2026) across four technical domains: species recognition, individual re-identification, soft biometrics (age/sex estimation), and behavior/pose analysis. We examine architectural evolution across three architectural paradigms—CNN-based models, Attention & Sequence Architectures, and Foundation Models- documenting performance advances and ecological applications. Key challenges—environmental interference, data scarcity, domain shift, computational constraints—and optimization strategies, including data augmentation, transfer learning, and model compression, are systematically addressed. Case studies demonstrate real-world impact across diverse taxa: giant pandas, Amur tigers, great apes, and marine mammals. We identify critical research gaps and emerging opportunities in foundation model adaptation, multimodal learning, spatiotemporal modeling, and edge deployment. This review bridges computer vision research and conservation practice, providing actionable insights for developing robust, deployable monitoring systems that translate algorithmic advances into conservation impact.