Visible-infrared imagery and deep learning for UAV-based automated levee hazard detection

Levee hazard inspection remains challenging because manual surveys are labor-intensive, prone to missed detections under dense vegetation and complex backgrounds, and often ineffective under low-light or nighttime conditions. Existing image-based inspection methods still suffer from sensitivity to illumination variations, image noise, weak detection of small-scale hazards, and insufficient exploitation of complementary visible and infrared information, which limits identification reliability and quantitative assessment accuracy. To address these issues, based on the complementary use of visible-light and infrared UAV imagery. A multimodal UAV acquisition strategy is developed to capture complementary texture and thermal information. A robust image enhancement pipeline is then constructed by integrating non-local means filtering, bilateral filtering, wavelet denoising, multi-scale Retinex, histogram equalization, and gamma correction to improve image quality under adverse environmental conditions. For hazard detection, an improved YOLOv8 model incorporating the Convolutional Block Attention Module and an optimized loss function is designed to strengthen the identification of small-scale hazards. Meanwhile, an enhanced ORB-based matching method and a seamline-optimized progressive fusion strategy are introduced to achieve reliable registration and visually natural fusion of homologous multi-view hazard images. Hazard regions are further segmented using DeepLabv3+ with morphological refinement, and a quantitative measurement procedure is established to estimate hazard dimensions. The proposed framework offers an effective and scalable solution for high-precision levee hazard detection and quantitative assessment.