Deep learning-based public crisis event identification for multimodal data contexts

Public crisis events surface across text, sensors, imagery, and logs, yet single-source detectors miss early weak cues.To address fragmented evidence, this study presents a multimodal bidirectional transformer for crisis recognition.First, source-aware tokens preserve time, space, provenance, and quality while synchronisation gates align asynchronous streams.Then, cross-source attention separates corroboration from dissent and memory tokens retain long-range hints.Finally, self-supervised pretraining and calibrated classification deliver auditable alerts.On composite streams, the method reaches AUPRC 0.612, AUROC 0.915, F1 0.672, ECE 0.038, and average lead time 31.6 minutes, exceeding the best baseline by 7.9 AUPRC points, 3.2 AUROC points, 6.9 F1 points, and 8.5 minutes.These gains provide earlier, more reliable, and well-calibrated alerts for public response.