Geophysical Journal International

AbstractAccurate 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 second 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 multi-scale time-frequency fusion for local magnitude (ML) 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 multi-scale convolutions; the frequency-domain branch combines hierarchical 2D 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 ML 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 dataset demonstrate substantial improvements over baseline models: mean absolute error decreased to 0.28, coefficient of determination R2 reached 0.872, with 82.5% of predictions achieving acceptable precision (error≤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.

SummaryIn marine seismic exploration, various types of seismic sources are employed to visualize geological structures beneath the seafloor, depending on survey objectives. Airgun sources, which generate large amounts of energy by releasing compressed air underwater, are typically used for imaging deep area; however, they have limited vertical resolution due to their low peak frequencies. In contrast, sparker sources generate wavelets with high peak frequencies using bubbles produced by discharging electrical energy to vaporize water, resulting in high vertical resolution. Sparker sources are useful for the detailed imaging of shallow strata but have a shallow penetration depth due to their low source energy. This paper proposes a method to integrate airgun and sparker data to broaden the frequency bandwidth and thus achieve more accurate geological interpretations. The study used small-scale airgun data and sparker data acquired in Yeongil Bay, Pohang, South Korea. A machine-learning-based shaping filter model was developed along with synthetic training data representing the airgun and sparker source wavelet characteristics, and the trained models were applied to regularize these source wavelets. Subsequently, time-variant spectral whitening (TVSW) and weighted integration were performed to yield the flattened broadband frequency spectrum. The integrated data have enhanced penetration depth and vertical resolution compared with the original single-source datasets, thus overcoming the interpretational limitations imposed by their limited frequency bandwidth and penetration depth and enhancing the reliability of associated geological interpretations.

SummarySite amplification in the Kumamoto area, Japan, is analyzed using 985 high-quality horizontal strong-motion records from 45 aftershocks (Mj = 2.7–4.9) recorded within 24 hours following the 2016 Kumamoto Mj 7.3 earthquake, as observed by 51 K-NET and KiK-net stations. For the generalized inversion technique (GIT), a reference station is required as a standard. In the GIT process, the number of events available for analysis is limited to those recorded by the reference station, and the stations whose site effects can be estimated are restricted to those that record common events with the reference station. To overcome the limitation of the GIT, the ‘transfer-station generalized inversion method (TSGI),’ a modified GIT, is introduced to increase the number of analyzed events and stations. The site responses obtained from GIT and TSGI for the same stations exhibit a high degree of consistency, thereby demonstrating the effectiveness of the TSGI. The discrepancies between the ${{Q}_S}$ estimates of GIT and TSGI can be attributed to the gradual expansion of the region represented by ${{Q}_S}$ as more events and stations are included in the inversion. However, the results of GIT and TSGI are relative to the reference station that may itself exhibit site effects. Thus, a reference-independent technique, i.e. genetic algorithm (GA), is also introduced to obtain the absolute site amplifications. The results show that at frequencies greater than about 1 Hz, the site response of the reference station is significantly lower than the theoretical amplification factor of 2, resulting in an overestimation of the site responses at other stations. When the results of GIT are corrected with the site response of the reference station obtained from GA, these two results agree very well for most of the stations. This indicates that the results of GIT are reliable if the reference station is an ideal surface rock station, and that the GA produces accurate absolute site amplification factors for the stations investigated in this study. In addition, we analyze the high-frequency attenuation characteristics of S-waves in the Kumamoto area, and establish $\kappa $ models for different site conditions and an empirical ${{\kappa }_0}$-${{V}_{S30}}$ relationship.