OmniDisaster: Multimodal Fusion for Multilingual Disaster Classification

Natural disasters are increasing in frequency and continue to cause significant human and economic losses. Meanwhile, social media provides a rich source of multimodal information, particularly images and text, that can support the timely identification of disaster-related events. In this work, we evaluate multimodal architectures for disaster-related social media classification using the CrisisMMD and DamageDataset benchmarks and introduce a cross-attention-based fusion strategy. We also present OmniDisaster, a new dataset composed of social media posts from South America. Our results demonstrate the benefits of combining visual and textual information, with the proposed cross-attention approach consistently achieving the best performance when paired with BGE-M3 text embeddings and PE-Core visual representations, yielding the strongest results across the evaluated configurations.