Deep learning-based time-evolving magnitude estimation for strong earthquakes using high-rate GNSS data

Waveform clipping, sensor tilting, and saturation in traditional seismic data often lead to systematic underestimation in rapid magnitude estimation for strong earthquakes. High-rate Global Navigation Satellite Systems (GNSS) offer an effective alternative by capturing coseismic displacements without being affected by these issues. However, in tectonically complex regions such as the Qinghai-Xizang Plateau, the scarcity of observational data limits the application of data-driven methods. In this study, we constructed 4200 rupture scenarios ranging from M_W5.37 to M_W8.06 and generated 617,400 three-component waveforms with realistic noise to simulate magnitude variations in northeastern Xizang. Based on this dataset, we developed a deep learning network framework combining dense layers and Long Short-Term Memory (LSTM) for time-evolving magnitude estimation. Validation on the synthetic test set showed that the percentages of predictions within the ±0.3 magnitude tolerance reached 95.7%, 96.6%, and 97.4% at 32 s, 64 s, and 128 s, respectively, with a mean absolute error of 0.103 at 128 s. Sensitivity tests based on station distribution further indicate that the model can reliably estimate magnitudes even with only four stations or uneven station distributions. Ablation experiments confirmed that the LSTM captures temporal dependencies, whereas the dropout layer enhances model stability. Compared with empirical PGD scaling laws and the CNN-based model, the proposed framework provides faster convergence, improved temporal stability, and reduced early-stage underestimation. Retrospective tests using the strike-slip (2022 M_W6.6 Menyuan and 2021 M_W7.4 Maduo) and thrust (2013 M_W6.6 Lushan) earthquakes further confirm the robustness and generalization capability of the proposed framework under different tectonic environments and faulting styles. These results demonstrate that the proposed framework provides a practical and robust approach for high-rate GNSS-based earthquake early warning in tectonically complex active continental regions.