Use of UAVs and deep learning for beach litter monitoring

Abstract

Stranded beach litter is a ubiquitous issue. Manual monitoring and retrieval can be cost and labour intensive. Therefore, automatic litter monitoring and retrieval is an essential mitigation strategy. In this paper, we present important foundational blocks that can be expanded into an autonomous monitoring-and-retrieval pipeline based on drone surveys and object detection using deep learning. Drone footage collected on the islands of Malta and Gozo in Sicily (Italy) and the Red Sea coast was combined with publicly available litter datasets and used to train an object detection algorithm (YOLOv5) to detect litter objects in footage recorded during drone surveys. Across all classes of litter objects, the 50%–95% mean average precision (mAP50-95) was 0.252, with the performance on single well-represented classes reaching up to 0.674. We also present an approach to geolocate objects detected by the algorithm, assigning latitude and longitude coordinates to each detection. In combination with beach morphology information derived from digital elevation models (DEMs) for path finding and identifying inaccessible areas for an autonomous litter retrieval robot, this research provides important building blocks for an automated monitoring-and-retrieval pipeline.