A Benchmark Evaluation of Intelligent Identification Models for Marine Outfalls

Monitoring marine outfalls is crucial for mitigating coastal pollution and protecting marine environments. Current methods rely mainly on manual inspection and satellite remote sensing interpretation, which are inefficient, inaccurate, and inadequate for large-scale real-time monitoring. Although UAV visible-light imagery has been introduced for marine outfall detection, challenges remain, including insufficient and diverse target features, small multi-scale target detection difficulties, and complex background interference. To address these limitations, this study systematically benchmarks mainstream object detection models (YOLOv8n, YOLOv9t, YOLOv10n, YOLOv11n, and RTDETR-light) on a dedicated multi-source remote sensing fusion dataset that we constructed for marine outfalls along Zhanjiang’s southern coast, incorporating NDWIs. Our comparative experiments evaluate the models’ effectiveness in this challenging scenario. Experimental results indicate that YOLOv8n is the most balanced model for marine outfall detection, achieving 84.1% precision, 68.6% recall, 77% mAP50, and an F1 score of 0.75. This benchmark provides empirical evidence and practical model selection criteria for intelligent marine outfall monitoring, thereby offering a reference framework for researchers and engineers in related fields.