A review of deep learning techniques for detecting animals in aerial and satellite images

Abstract

Deep learning is an effective machine learning method that in recent years has been successfully applied to detect and monitor species population in remotely sensed data. This study aims to provide a systematic literature review of current applications of deep learning methods for animal detection in aerial and satellite images. We categorized methods in collated publications into image level, point level, bounding-box level, instance segmentation level, and specific information level. The statistical results show that YOLO, Faster R-CNN, U-Net and ResNet are the most used neural network structures. The main challenges associated with the use of these deep learning methods are imbalanced datasets, small samples, small objects, image annotation methods, image background, animal counting, model accuracy assessment, and uncertainty estimation. We explored possible solutions include the selection of sample annotation methods, optimizing positive or negative samples, using weakly and self-supervised learning methods, selecting or developing more suitable network structures. Future research trends we identified are video-based detection, very high-resolution satellite image-based detection, multiple species detection, new annotation methods, and the development of specialized network structures and large foundation models. We discussed existing research attempts as well as personal perspectives on these possible solutions and future trends.

Keywords

• Biodiversity
• Wildlife
• Livestock
• Remote sensing
• Artificial intelligence
• Object detection