Abstract
<jats:p>Natural-hazard early warning relies on heterogeneous geophysical observations whose reliability varies sharply over time, yet most multi-modal geoscience pipelines fuse them with fixed weights or static attention that cannot adapt when one stream degrades. We propose GeoDMB, a contribution-aware dynamic multi-modal balance framework that fuses seismic waveforms, InSAR deformation fields, and infrasound spectrograms for joint event detection and epicenter localization. GeoDMB introduces a Contribution-aware Dynamic Balancing module that estimates a per-modality reliability score and a cross-modal consistency term at each time step, and combines them through a temperature-softmax to produce time-varying balance weights; a gating mechanism with a learned prior further ensures graceful degradation when all modalities are corrupted. The fused representation is decoded by a Conv-Transformer head that jointly predicts event probability and source location. To evaluate GeoDMB, we build GeoHazard-MM, a multi-modal dataset aligning IRIS seismic, Sentinel-1 InSAR, and IMS infrasound recordings for 3,200 global events from 2018 to 2024. On the test set, GeoDMB achieves a detection F1 of 88.9 percent and a localization error of 8.6 kilometers, outperforming fixed-weight, attention, and late-decision fusion baselines. The advantage widens under simulated modality dropout, where GeoDMB retains up to 84.1 percent F1 while attention fusion drops to 79.3 percent, and ablations confirm that the dynamic balancing module, the reliability estimator, and the gating mechanism each contribute to robustness. Per-event-type and cross-region analyses show consistent gains, and human evaluation by seismologists rates GeoDMB predictions as most credible. The results validate that input-conditioned contribution-aware balancing is essential for robust multi-modal fusion in the Earth sciences.</jats:p>