Multi-CD mini: a dataset for 3D urban semantic change detection with high spatial resolution imagery and nDSM

Abstract

Urban semantic change detection (SCD) using remote sensing imagery is critical for monitoring urban dynamics, yet traditional two-dimensional (2D) methods suffer from occlusion, shadow interference, geometric distortion and an inability to quantify vertical variations. Integrating normalized Digital Surface Models (nDSM) with optical imagery provides height information that enables three-dimensional (3D) SCD, but the lack of standardized benchmark datasets has hindered progress. To address this gap, we present Multi-CD mini, a 3D urban SCD dataset integrating multi-temporal high-resolution optical imagery and LiDAR-derived nDSM. It includes 745 pairs of 1024 × 1024 patches at 0.1 m spatial resolution, with pixel-level annotations for nine urban land‑cover classes covering both building and non-building changes. We benchmark four representative 2D SCD models with early and mid-level fusion schemes and analyze multi-modal information impacts. Experimental results show elaborate multi-modal fusion improve 3D SCD performance. Further analyses reveal current fusion methods struggle with inherent modal conflicts: spectrum-dominant changes with height ambiguity and height-dominant changes with spectral ambiguity. Multi-CD mini provides a challenging and reliable benchmark for evaluating multi-modal fusion algorithms. This study also highlights the need for conflict-aware adaptive fusion strategies for future robust 3D urban SCD research.