Abstract
Floating debris threatens aquatic ecosystems, yet detecting it in close-range aquatic remote sensing imagery remains difficult because small, irregular, or partially submerged objects resemble specular highlights and ripples. This study proposes FLNet, a YOLO11-L-based detector for water-surface monitoring from unmanned surface vehicles (USVs). FLNet redesigns the backbone, neck, and head with three components: a Dynamic Polarity-Aware Robustness Block that calibrates deep activations and models similar and opposing feature relations; an Adaptive Multi-Receptive Pyramid that repeatedly fuses three adjacent scales through multi-branch depthwise convolutions; and a Detail-Preserving Shared Head that applies one boundary-sensitive prediction trunk to all scales. On a custom five-category dataset of 12,237 images, FLNet achieved 85.93% macro-Precision, 85.08% macro-F1, and 88.77% mean average Precision at an intersection-over-union threshold of 0.5 (mAP@0.5), the highest single-run values among the evaluated detectors, with 24M parameters and 83 GFLOPs. Across three random seeds, its macro-Precision and macro-F1 advantages over the strongest competitor, Hyper-YOLO, exceeded seed-to-seed variability, whereas its mAP@0.5 was statistically comparable to that of Hyper-YOLO and Hyper-YOLO retained a higher mAP@0.5:0.95. In separately trained benchmarks, FLNet achieved the highest F1-score on FloW-Img (80.26%) and the highest macro-F1 on MarineDebris-yolo (79.82%). With TensorRT half-precision (FP16) acceleration, detector-side inference reached 29.8, 111.2, and 56.4 frames per second on Jetson Xavier NX, AGX Orin, and Orin Nano Super, respectively. FLNet therefore offers a favorable accuracy–efficiency balance for close-range floating-debris detection.