UWAD-Net: attention-enhanced underwater waste detection in freshwater pond environments

Introduction Detection of underwater waste in freshwater environments remains a challenging task due to the scarcity of domain-specific annotated datasets and adverse visual conditions caused by poor visibility, inconsistent illumination, color degradation, and suspended particles. Unlike existing underwater debris datasets that primarily focus on marine environments, freshwater pond ecosystems remain relatively underrepresented. Methods We introduce PondDebris, a publicly available underwater waste detection dataset comprising 1,526 images and 3,224 bounding box annotations across two ecologically relevant categories: biodegradable and non-biodegradable waste. The dataset was collected under controlled aquarium conditions designed to emulate several visual characteristics of shallow freshwater ponds. Furthermore, we propose UWAD-Net (UnderWater Attention Debris Network), an attention-enhanced object detection framework that integrates the Convolutional Block Attention Module (CBAM) into YOLOv8s to improve feature representation and localization of underwater waste. Results YOLOv8s provides a lightweight and efficient detection backbone, while CBAM enhances channel and spatial feature attention for improved feature discrimination in challenging underwater scenes. Experimental results demonstrate that UWAD-Net achieves an mAP50 of 0.7294, mAP50-95 of 0.4517, precision of 0.8206, and recall of 0.6421. Cross-dataset evaluation on an external underwater waste dataset yielded an mAP50 of 0.6629 and a precision of 0.7738, demonstrating its generalization capability. Conclusion The proposed dataset and detection framework provide an effective solution for freshwater underwater waste monitoring and support the development of intelligent environmental management systems.