Optimized YOLO-FOD for Enhanced Water Surface Object Detection in Environmental Monitoring

Environmental monitoring using unmanned surface vehicles (USVs) faces significant challenges in detecting small floating objects due to complex aquatic environments characterized by dynamic lighting, cluttered backgrounds, and frequent occlusions. This paper presents YOLO-FOD, an advanced object detection network that substantially improves detection accuracy through three key innovations: (1) Varifocal Loss function for precise bounding box regression, (2) SegNeXt-inspired multi-scale attention mechanisms for enhanced feature representation, and (3) a dual-module architecture combining Detail-Enhanced Convolution (DEConv) with Selective Boundary Aggregation (SBA) for improved small object detection. Comprehensive evaluations demonstrate that our approach achieves state-of-the-art performance with 97.3% mAP on water surface datasets, representing a 3.7-point improvement over baseline models while maintaining real-time processing at 156 FPS. The proposed system exhibits exceptional robustness across diverse aquatic conditions, effectively addressing critical challenges in environmental monitoring applications. These advancements in both accuracy and computational efficiency make YOLO-FOD particularly suitable for deployment in practical USV-based monitoring systems where reliable detection of floating debris is essential for water quality management and ecosystem preservation.