SRGO-YOLO: Super-Resolution and Granularity-Optimized Detection of Extremely Small Beach-Litter Objects in UAV Imagery

Beach litter poses persistent ecological and socioeconomic risks to coastal zones, yet reliably detecting extremely small objects (ESOs) in unmanned aerial vehicle (UAV) imagery remains challenging: these targets occupy only a very small image area on the order of 10-5of a typical RGB frame and are easily confused with textured sand backgrounds. We propose SRGO-YOLO, a super-resolution and granularity-optimized detector for UAV-based beach-litter monitoring. The network integrates a Scale-Adaptive Spatial Recalibration with Channel Attention module, a lightweight C3K2-MGLite neck, and a Spatial-Optimization-Aware Head with an additional P2 tiny-object branch to enhance shallow high-resolution features, perform mixed-granularity multi-scale fusion, and refine predictions for low-contrast ESOs. To evaluate SRGO-YOLO, we construct the Beach-Litter-UAV dataset by screening two public UAV garbage datasets and combining them with a newly collected Zhuhai Coastal Beach Litter Dataset, in which more than 80% of annotated instances have normalized sizes below 0.02. Experiments show that SRGO-YOLO achieves 81.4% mAP50and 48.2% mAP50:95, surpassing YOLOv11s by 21.8 and 13.5 percentage points, respectively, while maintaining 127 FPS on an RTX 3090 GPU. On a Jetson AGX Orin 64GB platform with TensorRT FP16 deployment, SRGO-YOLO further sustains 31.21 FPS in a 10-min continuous inference test with an average end-to-end latency of 32.04 ms. Per-class results further show consistent gains across all five categories: compared with YOLOv11s, SRGO-YOLO improves mAP50by 28.8%, 31.2%, 5.4%, 29.5%, and 14.5% for plastic, flexible pouch, metal, glass, and food container, respectively, with particularly strong benefits for visually ambiguous small items such as flexible pouch and food container. Qualitative comparisons further indicate that SRGO-YOLO recovers more complete instances under partial occlusion, tight adjacency, and overlapping litter patches. These results indicate that SRGO-YOLO not only improves ESO detection accuracy but also shows strong potential for real-time UAV-based beach-litter detection, supporting timely shoreline litter mapping and targeted coastal cleanup planning.