GLoDA-Net: A lightweight global–local directional aggregation network for floating waste detection in inland waters using unmanned surface vehicles (USVs)

Floating bottle waste detection in inland water environments remains challenging because targets are often small, low-contrast, elongated, partially submerged, and easily confused with water-surface reflections, ripples, vegetation, and bank-side textures. To address these challenges, this study proposes GLoDA-Net, a lightweight YOLOv12n-based detector for USV-oriented floating debris monitoring. The network combines compact global–local feature extraction, contextual interaction, and direction-aware feature fusion through RepViT, C2PSAMILA, and APC2f. Experiments on the FLOW-img benchmark show that GLoDA-Net achieves 88.0% Precision, 80.9% Recall, and 87.7% mAP@50, outperforming the original YOLOv12n by 5.1, 1.7, and 4.8 percentage points, respectively. The model maintains a compact structure with 2.59M parameters and a size of 5.6 MB, while reaching 117 FPS on an RTX 4090 GPU. Ablation, robustness, cross-scenario, multi-platform inference-speed, and cross-dataset evaluations further demonstrate improvements in detection accuracy, robustness, transferability, and computational efficiency. The measured speeds support real-time inference on the evaluated GPU platforms and indicate deployment potential for USV-based floating bottle monitoring, while resource-constrained onboard systems may require further optimization.