ChangeMamba Meets BRIGHT: Benchmarking Multimodal Damage Mapping and Cross-Event Transfer to a Japanese Wildfire

Rapid, reliable building damage mapping (BDM) is essential for effective humanitarian response and disaster management. Although Earth Observation (EO) data availability and AI model design have advanced rapidly, systematic and standardized comparisons of methods for multimodal BDM remain scarce. As new architectures emerge at a fast pace, understanding their relative strengths and limitations on common benchmarks is crucial for operational deployment.

In this work, we leverage the BRIGHT dataset, a recent large-scale benchmark for multimodal BDM, to conduct a comprehensive evaluation of representative strategies spanning traditional machine learning, Convolutional Neural Networks (CNNs), Transformers, Mamba, and emerging foundation models. Our benchmarking shows that, despite their scale, general-purpose foundation models are still outperformed by specialized architectures in complex multimodal BDM settings. In particular, ChangeMamba, a state-of-the-art Mamba-based model, achieves the strongest overall performance on BRIGHT. 

To further assess robustness and transferability beyond the benchmark, we perform a cross-event transfer evaluation on a recent wildfire in Oita, Japan. The results demonstrate ChangeMamba’s superior generalization in real-world conditions compared with other baselines. Finally, our analysis reveals a key sensitivity in multimodal fusion: the choice of pre-event optical imagery substantially affects performance when transferring to unseen events, highlighting an important practical consideration for operational damage mapping.