Image-Supervised Vegetation Segmentation and Conservation-Oriented Mapping for Historic Masonry Point Clouds

Spontaneous vegetation on historic masonry structures may accelerate material deterioration, obscure masonry joints, and interfere with point-cloud-based documentation and conservation assessment. This study proposes a conservation-oriented image-supervised workflow for vegetation segmentation and spatial mapping in historic masonry point clouds. The Daqin Temple Pagoda in Xi’an, China, was selected as a case study. Field inspection and deterioration mapping indicated that vegetation was concentrated around the upper eaves and spire regions. An Excess Green (ExG) index was used as an auxiliary indicator for identifying vegetation-prone regions and characterizing height- and orientation-related vegetation responses. To reduce the burden of three-dimensional annotation, the point cloud was projected into two-dimensional views, where pixel-level annotations were used to supervise vegetation–masonry classification. An HSV-based multi-scale attention-enhanced multilayer perceptron was introduced to improve feature discrimination under uneven illumination. The optimal image-based configuration achieved an overall accuracy of 92.1%. Within the proposed image-supervised 2D-to-3D framework, confidence-guided spatial refinement improved segmentation consistency, achieving projected-point-based IoU values of 0.9354 for vegetation and 0.8859 for masonry. The ExG-based spatial analysis revealed higher vegetation-related responses around the spire and upper eaves, providing spatial information for vegetation inspection and documentation. The proposed workflow offers a practical approach for vegetation-aware point-cloud preprocessing and HBIM-oriented documentation.