A Knowledge‐Guided and Feature‐Constrained Method for Accurate 3D Geometric Morphology Diagnosis of Damaged Buildings

Abstract

Buildings are the primary disaster‐bearing structures of natural disasters such as earthquakes, and accurately diagnosing the 3D geometric morphology of damaged buildings is crucial for emergency response and post‐disaster reconstruction. However, existing methods rely solely on pure point cloud geometric alignment and do not incorporate multidimensional domain knowledge, such as disaster, construction, and diagnostic standards, resulting in significant shortcomings in accuracy and efficiency when applied to complex damage scenes. To address these challenges, we propose a knowledge‐guided and feature‐constrained method for accurate 3D geometric morphology diagnosis of damaged buildings. First, we analyze the characteristics and interrelationships of damaged building scenes, constructing a multi‐domain associated scene knowledge graph (KG) based on “event‐object‐morphology‐rule”. Second, geometric features of buildings are extracted from multi‐temporal point clouds, and a precise comparison algorithm for 3D geometric morphology of damaged buildings constrained by linear and surface features is designed. Then, a scene knowledge‐guided building damage assessment method is proposed, and building geometric features along with associated rules are analyzed. Finally, a damaged building is selected as a case area for experimental analysis. Experiments show that the proposed method maintains a diagnostic accuracy above 98% for 3D geometric morphology, which is at least 10% higher than conventional methods, achieving precise diagnosis of 3D geometric morphological changes in damaged buildings. Furthermore, compared to the ICP and GeoTransformer, our approach improves diagnostic efficiency by over 80%, enabling timely and effective information support for post‐disaster emergency response and reconstruction, while also establishing a new paradigm for 3D geometric morphology diagnosis of damaged buildings.