Undercutting Damage Volume Quantification in Earthen Sites: Coupling Photogrammetry with Semantic Segmentation

Undercutting significantly affects the mechanical stability of earthen archaeological sites and represents a critical parameter for heritage conservation decision-making. To quantify undercutting volume, this study developed a methodology coupling close-range photogrammetry with semantic segmentation. We selected seven earthen sites spanning diverse climatic and historical contexts, acquired 3D models, and generated RGB-DSM fused images. A dataset of 576 samples was established with manual annotations of undercutting and co-occurring deterioration (holes, fissures, and gullies). For deterioration identification, we employed an enhanced U-Net architecture featuring reduced down-sampling, a dilated convolution context module in the bottleneck, an FPN-style decoder, and an optimized loss function. The model achieved per-class IoU values of 0.485–0.510 for co-occurring deterioration and 0.668 IoU (0.801 F1-score) for undercutting. Volume was computed by fitting a reference surface above the undercutting zone and integrating pixel-wise elevation differences. Analysis showed that DOM/DSM should be acquired as perpendicular to the wall as possible, because angular deviations introduce geometric underestimation of the area and reduce recognition performance. The observed volume CV of 8–25% far exceeds the theoretical 4.2% from geometry alone, confirming recognition variability as the dominant uncertainty source. By simplifying 3D point cloud recognition to a 2D image task, this approach avoids prohibitive annotation and computational costs, offering a practical solution for linearly distributed earthen sites and providing essential data for structural safety assessment and deterioration mechanism studies.