AI in the sky - Advancing wildlife survey methods in Africa with deep learning and aerial imagery
Zeyu Xu is a PhD student in the Department of Natural Resources. (Co)Promotors are prof.dr. A.K. Skidmore and dr. T. Wang from the Faculty ITC.
Monitoring wildlife populations is critical for tracking biodiversity trends and informing conservation efforts, especially in ecologically rich but logistically challenging regions such as sub-Saharan Africa. Aerial imagery offers a scalable means of collecting wildlife data over large areas, yet translating these data into reliable ecological indicators—such as species abundance and distribution—remains complex. This thesis explores how deep learning can enhance aerial wildlife surveys by automating species detection and improving abundance estimation.
The research begins with a review of current deep learning approaches for animal detection in remote sensing imagery, highlighting key gaps in benchmark datasets, annotation practices, and ecological interpretability. It then investigates how annotation strategies and image-level factors influence model performance. Bounding box annotations improve accuracy but are labor-intensive, while point annotations offer a more efficient alternative under certain conditions. To address environmental complexity, the thesis develops a quantitative framework incorporating two critical image-level factors—spatial heterogeneity and target-background contrast—that significantly affect detection accuracy.
Building on these insights, a unified deep learning pipeline is proposed to support end-to-end aerial wildlife surveys. The system integrates three key components: (1) U-Net–based delineation of valid observation strips; (2) duplicate removal using geometric analysis and feature matching; and (3) contrast-sensitive detection of cryptic animals using a customized YOLO-based model. This framework improves detection accuracy, eliminates overcounting, and operates without reliance on GPS metadata, enabling use with historical and low-resource datasets.
The findings demonstrate that deep learning, when combined with remote sensing and ecological understanding, can form the basis of scalable and automated wildlife monitoring systems. The research aligns with the Remote Sensing-enabled Essential Biodiversity Variables (RS-EBVs) framework and supports global biodiversity initiatives such as the Convention on Biological Diversity. Future directions include enhancing model generalization across ecosystems, integrating temporal dynamics, and developing real-time, multimodal monitoring systems. Overall, this thesis contributes to the advancement of AI-enabled biodiversity monitoring and offers practical tools for more inclusive, efficient, and scientifically grounded conservation practices.
