Pruned U-net with multi-scale feature fusion and attention for real-time UAV remote sensing of levee defects

Abstract

Abstract The long-term performance of levee infrastructure is increasingly threatened by environmental exposure and material degradation, underscoring the need for efficient, accurate inspection. Unmanned Aerial Vehicle (UAV)-based remote sensing offers a cost-effective solution, enabling rapid acquisition of high-resolution imagery over large surfaces; however, stains, occlusions, and illumination variability frequently degrade automated detection. To address these challenges, we propose a real-time semantic segmentation framework built on an optimized U-Net. The model integrates structured pruning to accelerate inference, a residual convolutional block attention module (ResCBAM) to suppress background interference and enhance defect saliency, and a multi-scale feature-fusion strategy with online feature distillation to strengthen fine-grained representations across resolutions. We evaluate the approach on UAV imagery collected from an aged levee section. The proposed method attains 90.05% accuracy, 88.94% recall, 89.22% precision, and 88.67% IoU, outperforming state-of-the-art baselines, while achieving a real-time processing rate of 57.74 FPS. These results demonstrate that the framework delivers a favorable speed–accuracy trade-off and is suitable for large-scale UAV-based levee monitoring. Overall, the experiments indicate strong potential for timely defect identification and proactive risk management in levee systems.