Confidence-aware adaptive multimodal fusion for crisis event analysis from social media

During crisis events, social media provides timely textual and visual information, but automated classification is challenged by class imbalance, missing modalities, and variation in the informativeness of textual and visual content. This study introduces CAGMF, a Confidence-Aware Adaptive Gated Multimodal Fusion framework that combines bidirectional cross-attention, adaptive gates, and dynamic \\(\\alpha \\) -routing to adaptively fuse textual and visual evidence for each paired sample. CAGMF was evaluated on the CrisisMMD informativeness detection task (Task 1) and humanitarian category classification task (Task 2), with results reported as mean performance over three fixed seed values. Under identical backbone encoders and optimization settings, CAGMF achieved Macro F1 scores of 90.68% for Task 1 and 83.45% for Task 2, compared with 90.10% and 80.86%, respectively, for Concatenation Fusion. In comparison with representative CrisisMMD studies, CAGMF achieved competitive Task 1 performance and the strongest Task 2 Accuracy and Weighted F1 among the main reported results. Ablation results identified bidirectional cross-attention as the largest contributor, with adaptive gating, dynamic \\(\\alpha \\) -routing, and training strategies also contributing. Under random modality removal, CAGMF with modality dropout achieved a Macro F1 of 76.97%, compared with 74.03% without modality dropout and 72.95% for Concatenation Fusion. These findings support confidence-aware adaptive fusion for aligned multimodal crisis classification.