SODNet: a scale-oriented detection network for efficient UAV-based sewage outfall detection

Abstract

Abstract Accurate identification of river sewage outfalls is crucial for effective water pollution control. Unmanned Aerial Vehicles (UAVs), with their high mobility and wide coverage, have become a vital tool for this monitoring task. However, this application is hampered by the dual challenges of robust multi-scale object detection and lightweight model deployment on computationally limited platforms. To address the trade-off between accuracy and efficiency, this study proposes an efficient deep learning-based detection method, termed Scale-Oriented Detection Network (SODNet). Specifically, we propose an Efficient Context Feature Pyramid Network (ECFPN) to enhance multi-scale feature representation. Additionally, a shared decoupled head with a Multi-Scale Grouped Fusion (MSGF) module strengthens feature extraction while reducing computational costs. Furthermore, a channel pruning strategy is employed to compress the model, notably improving inference speed. Experimental results demonstrate that SODNet achieves an AP@50 of 89.9% and a precision of 91.1%, representing improvements of 1.2% and 2.7% over the baseline model, respectively. Meanwhile, parameters and GFLOPs are reduced by 77.5% and 73.6%. On a deployed edge device, SODNet achieves 40.3 FPS. These findings indicate that SODNet gains substantial computational efficiency while maintaining excellent detection performance, making it ideal for resource-constrained UAV scenarios and offering a feasible solution for intelligent environmental supervision.