Monocular depth prior-guided scale-aware reliability fusion for UAV floating plastic detection

Floating plastic pollution has become an environmental issue that threatens aquatic ecosystem health. However, large-scale aquatic monitoring remains challenging due to complex backgrounds, illumination variations, and the small size, irregular shapes, and low visual contrast of floating plastic targets. UAV-based object detection provides an efficient approach for rapid aquatic pollution assessment. Existing methods mainly rely on RGB appearance information, which limits detection performance when targets exhibit weak textures, small scales, and low saliency. To address these limitations, this paper proposes SARF-YOLO, a UAV-based floating plastic detection framework with monocular depth prior-guided scale-aware structure alignment and reliability-aware fusion. Different from methods using physical depth sensors, SARF-YOLO generates pseudo-depth structural representations from RGB images using a monocular depth estimation model. The generated Monocular Depth Prior (MDP) does not provide physical distance measurements. Instead, it encodes object boundaries, spatial layouts, and relative geometric relationships to provide complementary structural cues for RGB features. Based on this representation, the Scale-aware Reliability Interaction Module (SRIM) enhances the cross-representation interaction between RGB appearance features and depth structural priors. The Alignment-guided Reliability Calibration Module (ARCM) dynamically calibrates feature contributions according to their reliability under complex aquatic conditions. This design improves detection stability for floating plastic targets with weak visual evidence. A UAV-based floating plastic dataset, AerialPlastic-5, is constructed in this study. The dataset contains 4487 real-world UAV images collected from aquatic environments and covers five typical floating plastic categories. Experimental results show that SARF-YOLO achieves 97.3% mAP50 on the test set, improving YOLOv11n (Baseline) by 3.3 percentage points while maintaining competitive computational efficiency. Ablation studies, small-object detection analysis, and image degradation experiments further verify the effectiveness of the proposed structure alignment and reliability-aware fusion strategies. The proposed framework provides an automated solution for UAV-assisted aquatic monitoring and supports large-scale floating plastic pollution assessment and ecological management.