MP-Net: an end-to-end approach based on time–frequency fusion for earthquake magnitude prediction
Year: 2026
Type: article
Abstract: SUMMARY Accurate and rapid magnitude prediction is critical for earthquake early warning systems, directly affecting emergency response decisions and public safety. With global seismic monitoring networks expanding to over 15 000 stations and the emergence of crowdsourcing-based IoT device monitoring systems, daily seismic data has reached petabyte scales, posing enormous challenges for real-time processing under the typical 3–10 s warning window constraint. Existing deep learning methods predominantly adopt single-modal information processing strategies, focusing either solely on temporal features of time-domain waveforms or spectral information after frequency-domain transformation, failing to fully exploit the joint evolution patterns and complementary information of seismic signals in the time–frequency domain, thereby limiting prediction accuracy and generalization performance. This paper proposes MP-Net, an end-to-end deep learning framework based on multiscale time–frequency fusion for local magnitude ($M_{\rm L}$) prediction. The method employs a dual-branch architecture that simultaneously processes raw three-component waveforms and spectrograms: the time-domain branch captures features from microscopic waveform details to macroscopic energy evolution through parallel multiscale convolutions; the frequency-domain branch combines hierarchical 2-D convolutional networks with adaptive spectral attention mechanisms to automatically identify magnitude-related frequency components while suppressing noise; a cross-attention based fusion module achieves deep integration of complementary information from both modalities. To preserve the absolute amplitude information physically consistent with the $M_{\rm L}$ definition, logarithmic amplitude features are extracted prior to waveform normalization and provided as auxiliary inputs to the fusion layer. Comprehensive experiments on the large-scale STEAD data set demonstrate substantial improvements over baseline models: mean absolute error decreased to 0.28, coefficient of determination R$^2$ reached 0.872, with 82.5 per cent of predictions achieving acceptable precision (error$\le$0.5). The proposed approach provides an efficient and accurate solution for real-time single-station magnitude prediction, applicable to earthquake early warning systems operating in both centralized and distributed computing environments.
DOI: 10.1093/gji/ggag204
Authors: Hanling Wang, Qiang Ma, Quancai Xie, Huiyu Zhu, Dongwang Tao, Jiang Wang
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $3,416
OpenAlex ID: W7162184306