Cascaded Attention-Based Multimodal Framework for Robust Disaster Tweet Classification

Disasters require timely and informed responses, and social media platforms like Twitter (now X) have become vital sources of real-time information. Unimodal methods rely on a single data source that leverages either text or visual modalities, resulting in incomplete situational awareness. However, tweets often contain noisy, brief, and embedded image information, making classification a challenging task. To effectively harness textual and visual modalities, we propose a cascaded attention-based multimodal framework that improves the precision and relevance of disaster-related insights. The model employs a dual-stream architecture that separately encodes textual and visual features, followed by a cascaded attention mechanism that dynamically refines and fuses modality-specific information. Our approach leverages both intra-modal attention to highlight informative features within each modality and inter-modal attention to align and fuse complementary semantics across text and visuals. Experiments conducted on the benchmark disaster tweet dataset demonstrate that the proposed framework significantly outperforms existing unimodal and multimodal baselines in terms of accuracy and robustness. The proposed framework primarily contributes to research in multimodal disaster analysis. It can be integrated into social media monitoring and decision-support systems used by humanitarian organizations, emergency management agencies, and public safety authorities to enhance situational awareness and support faster, more informed crisis response.