SD2GC-F: Enhancing Building Change Detection With Sequential Detection to Graph Comparison Framework

Abstract

Accurately detecting building changes based on high-resolution remote sensing imagery remains technically challenging owing to positional inconsistencies and geometric distortions. To address these limitations, this study proposes a novel framework that combines deep learning-based building detection with graph-based structure comparison. Detectron2, an object detection model, is employed to extract building instances and derive accurate node positions by computing the center points from rotated bounding boxes. The minimum spanning tree algorithm is then applied to create building graphs from these nodes based on the connectivity between adjacent buildings. Subsequent analysis of structural variations within this graph enables change detection and identifies which building changes will concomitantly alter their links to neighboring buildings. Experimental results across synthetic and real-world datasets (including off-nadir imagery) confirm that the proposed method effectively captures building changes in complex urban environments. Notably, it achieved high change-detection accuracy, particularly in scenarios involving relief displacement and perspective distortion, wherein conventional methods often yield high false positive rates. This approach offers practical utility for large-scale urban monitoring and addresses the key challenges posed by complex positional discrepancies and environmental variations in remote sensing imagery.