Graph-Based Relational Learning for Building Change Detection

Abstract

Recent graph-based change detection (CD) methods either encode relationships implicitly or rely on post hoc matching, limiting the explicit modeling of object-level spatial and topological relationships under geometric distortions. To address this issue, we propose a CD framework based on a rasterized graph feature map (GFM) that encodes node- and edge-level relationships and is integrated with RGB imagery. The framework incorporates relational feature modulation and anisotropic relational modeling to capture orientation-dependent interactions between buildings, and a graph-aware composite loss is designed to guide relationally consistent predictions. Experiments on the BANDON dataset demonstrate that the proposed method improves the overall CD performance, achieving higher $F1$ -score and mIoU with a reduced false positive rate under challenging off-nadir conditions. Ablation studies further show that node- and edge-level cues are most effective when jointly represented as a full GFM and that graph-aware losses contribute to improving prediction reliability by suppressing spurious responses.