A2ANet: Real-time detection of floating marine debris using atrous convolution and channel attention

Floating marine debris such as plastics, cans, and discarded packaging has become one of the most persistent threats to aquatic ecosystems and coastal sustainability. Detecting and tracking this debris in real time is vital for protecting biodiversity and guiding cleanup and policy actions. In this study, we introduce A2ANet, a lightweight deep learning framework that combines multi-scale atrous convolutions and Enhanced Channel Attention (ECA) to detect small, submerged, and visually ambiguous debris under challenging aquatic conditions. These mechanisms-expanding the receptive field and highlighting salient cues-reduce errors from reflections, glare, and clutter in aquatic scenes. A2ANet was evaluated on two datasets: a newly developed six-class dataset (D_six) representing real-world river conditions, and the publicly available FloW-Img benchmark. The model achieved mAP@0.5 of 0.841 and 0.892 on D_six and FloW-Img datasets, respectively, with inference time as low as 39 ms/image. Beyond detection performance, by enabling automated and frequent monitoring, A2ANet provides actionable insights for mapping pollution, tracking trends, and supporting ecosystem management. The framework offers a practical pathway toward intelligent aquatic observation systems aligned with Sustainable Development Goal 14: Life Below Water. All code and datasets are openly available (see Data Availability).