Building Damage Assessment Captioning: A Vision-Language Dataset Leveraging State Space Models

Assessing building damage after disasters is critical for effective response and recovery efforts. While Remote Sensing Image Change Captioning (RSICC) has shown promise in automating damage descriptions, existing datasets often lack the scale and diversity needed to train robust models for realworld scenarios. To address this gap, we introduce the Building Damage Assessment Captioning Dataset (xBD-C), an enriched extension of the widely recognized xBD dataset. By leveraging large language models (LLMs), we generate detailed, contextaware captions for bitemporal remote sensing images, capturing nuanced disaster impacts across diverse scenarios. This work stands apart from prior studies by combining real-world disaster data with comprehensive captions, enabling the development and evaluation of advanced RSICC models. To validate the utility of xBD-C, we employ the state-of-the-art RSCaMa model, which integrates joint spatial-temporal modeling through differenceaware and temporal-traversing mechanisms. The evaluation results demonstrate the effectiveness of the dataset, achieving BLEU scores of 90.51, 84.91, 80.76, and 77.44 for BLEU-1, BLEU-2, BLEU-3, and BLEU-4 respectively. Additional metrics include a METEOR score of 50.65, a$\mathbf{ROUGE}_{L}$score of 82.14, and a CIDEr-D score of 211.78, reflecting strong semantic alignment and descriptive accuracy. These findings confirm that xBD-C provides a valuable benchmark for advancing RSICC research, facilitating the development of models capable of generating precise and context-rich descriptions for disaster scenarios. This contribution empowers the remote sensing community with a robust resource for improving disaster impact analysis and decision-making.