YOLO26m-CW: A Water-Augmented Spatial Attention Framework for Robust Riverine Waste Monitoring

Abstract

This thesis presents YOLO26m-CW, a lightweight object detection framework for identifying floating plastic waste in rivers using Unmanned Surface Vehicles (USVs). The framework integrates Coordinate Attention (CoordAtt) and Water-Specific Augmentations (WaterAugs) into the YOLO26 architecture to improve detection accuracy under challenging aquatic conditions, including glare, reflections, and waves. Evaluated on the FloW-Img benchmark, the model achieved an F1 score of 0.859 and mAP50 of 0.866 on the 60–40 data split, outperforming Faster R-CNN and Cascade R-CNN. The study highlights the model’s potential for real-time plastic waste monitoring and automated environmental management through edge deployment and cloud-based GIS integration.