An explainable visual structural health monitoring framework for corrosion diagnosis of steel components using high-resolution UAV images

Reliable corrosion monitoring is essential for maintaining the safety, serviceability, and durability of steel structures. However, most vision-based corrosion monitoring methods have been developed for bridges or exterior steel surfaces, and their applicability to indoor industrial environments remains limited because of adverse illumination, cluttered backgrounds, partial occlusion, and uneven corrosion distributions. Moreover, segmentation-based corrosion monitoring is often treated as a pixel-level recognition task, while its connection with condition assessment and diagnostic interpretability remains insufficiently explored. To address these limitations, this study proposes an explainable UAV-based visual health monitoring framework for corrosion localization and condition assessment of steel components in industrial buildings. A dataset comprising 195 high-resolution UAV images was established for this challenging indoor scenario. The proposed framework integrates foreground-guided cropping, sliding-window inference, and an improved semantic segmentation model to preserve fine-grained corrosion information and achieve full-image prediction. A region of interest (RoI)-level image-based apparent corrosion index is introduced to convert segmentation outputs into local condition indicators, while a Kernel SHAP-based module provides diagnostic explanations for RoI-level predictions. Experimental results show that the proposed model achieves an F1-score of 0.71 and outperforms representative comparison models in corrosion segmentation. Additional cross-device validation using smartphone-captured images further showed that limited target-domain samples substantially improved segmentation performance under different acquisition conditions. The framework provides preliminary support for visual screening and inspection prioritization for subsequent detailed assessment.