Task-Aware Multimodal Large Language Models for Distributed Acoustic Sensing–Based Seismic Monitoring

Abstract

Seismic monitoring detects earthquake activity to investigate fundamental geophysical processes, advance understanding of earthquake mechanisms, and support hazard mitigation. Distributed acoustic sensing (DAS) uses existing fiber-optic cables to provide dense, long-range, high-resolution measurements over wide areas at low cost, showing strong potential for seismic monitoring. However, manually labeled DAS data are scarce and large annotated datasets are hard to obtain, which limits supervised deep-learning methods. In this paper, we present DAS-MLLM, a training-free framework that leverages multimodal large language models (MLLMs) for seismic monitoring. DAS-MLLM comprises three components: (1) a task-aware Visual Generator that adaptively selects informative spectro-temporal visualizations of DAS signals; (2) a composite-similarity Retriever that uses time- and frequency-domain features to retrieve relevant exemplars for in-context learning; and (3) a feedback-guided Self-Corrector that integrates unselected visualizations to refine predictions. We evaluate DAS-MLLM on the Ridgecrest dataset and on the Xinfengjiang DAS dataset from Guangdong, China (558 events, with a high proportion of microearthquakes); the latter also includes seismometer recordings for some of the same events, enabling direct comparison with seismometer-based methods. Results show that our method improves most metrics in earthquake detection, magnitude estimation, and hypocenter localization tasks. DAS-MLLM reduces localization error by up to 28.8% and improves detection F1-score by 19.4% compared with state-of-the-art methods.