Building facade debonding damage detection method based on explainable artificial intelligence and UAV-thermography in cross scale field of view

Abstract

Debonding damage detection in building façades using UAV-thermography is essential for ensuring structural safety. However, three major challenges remain: detection is difficult in the cross scale field of view, model is insufficient under small sample data-set, model predictions lack explicability. To address these issues, this paper proposes a building facade debonding damage detection method based on explainable artificial intelligence. Firstly, a double attention enhanced C2f module (DA-C2f) is designed to effectively capture multi-scale information through simultaneous spatial and channel feature calibration; Then, an Omni-Dimensional Dynamic Convolution (ODConv) is employed to achieve multidimensional adaptive adjustment of convolution kernels; Simultaneously, the gradient-weighted class activation mapping (Grad-CAM) is introduced to generate heatmaps, constructing a visual explanation system. The results show that the proposed method achieves an accuracy of 85%, a recall of 84%, and a mean average precision (mAP) of 93.1% with 455 samples, significantly outperforming traditional state-of-the-art detection models.