Interpreting regional characteristics of heritage houses based on 3D digital landscape and point cloud deep learning – Take Tibetan houses in the northeastern region of Aba prefecture as an example

This study proposes an application framework for automatically interpreting regional characteristics of heritage houses with the support of 3D digital landscape (3D-DLS) and point cloud deep learning (PC-DL) models, taking the northeastern region of Aba prefecture as an example. The significant contribution of this framework is introducing (1) the self-developed MSBFI model to decode the quantitative discriminative logic of the PC-DL model in categorizing the regional characteristics of heritage house point clouds, and (2) how it combines the generated saliency point clouds and the semantic segmentation point clouds to automatically filter the feature elements of each sub-region and analyze the regional characteristics of heritage houses. The proposed framework can adaptively calibrate multi-scale sensitivity weights, demonstrating strong compatibility with built heritage studies through its focused analysis of composite building scales, while maintaining superior computational efficiency. The conclusions of this framework on heritage houses are more quantitative, objective, and validated by the relevant field survey studies. In addition, this study fills the gap in interpreting the regional characteristics of Tibetan houses on a whole-area scale in the northeastern region of Aba prefecture.