We developed a fine-tuned deep learning model using SegPhase to pick seismic wave travel times from data observed by MeSO-net. The SegPhase model was fine-tuned using MeSO-net data to adapt its pre-trained capabilities to the limited dataset of 522 events and 13 observation points. Evaluation showed that fine-tuning significantly reduced evaluation function value compared to the pre-trained model. All P-waves were detected at a threshold of 0.1, with similarly high performance for S-waves.

Recent advancements have introduced deep learning techniques into seismic wave detection, particularly for identifying P-waves and S-waves and estimating their arrival times. However, most CNN-based models currently employed—such as the simple U-Net structures exemplified by PhaseNet — rely heavily on Cross Entropy loss functions, leaving considerable room for improvement. In this study, we propose a novel model architecture based on PhaseNet that is better suited to seismic waveforms, alongside a loss function specifically tailored for seismic wave detection. We tested our proposed method on a subset of the metropolitan seismic waveform dataset provided by GeoSciAI2024, achieving an average residual sum of squares of 0.5613. These results demonstrate the potential of our approach to enhance the accuracy of seismic wave detection.

We propose a neural network for identifying P and S phase arrivals from seismic waveforms using a U-shaped architecture enhanced with multi-scale residual connections and attention mechanisms. Our experiments identify that challenges arise from a wide range of signal-to-noise ratios (S/N) and earthquake magnitudes. This is further exacerbated by a small training set, and related over-fitting issues. Our final metric is 9.383 sec2, which corresponds to the RMS error of 0.3468 sec.

This paper presents a deep learning approach for P- and S-wave picking from seismic waveforms by applying spatial feature extraction inspired by full-waveform LiDAR data, where observation points and recorded waveforms resemble point clouds and LiDAR signals. Using FWNet++, a model for full-waveform LiDAR data, we extract spatial features from seismic waveforms at each observation point for wave picking, supported by a discriminator to verify spatial consistency against ground truth. The method incorporates a refinement strategy for rule-based results and shows lower residual errors than existing libraries. However, high residuals persist when P- and S-waves are visually indistinct, suggesting a need for seismological rules, such as epicenter distance, to enhance performance. Future research should integrate seismological insights with AI to develop more effective automated picking methods.

Solar flare prediction is essential for mitigating the adverse effects of space weather on satellites and ground-based infrastructure. This study proposes a novel method for solar flare prediction using low-resolution (512 × 512) line-of-sight magnetogram data provided by the Solar Dynamics Observatory (SDO). By utilizing low-resolution data, the proposed method aims to reduce computational costs while maintaining practical predictive accuracy. The method involves extracting active regions from the SDO's line-of-sight magnetogram, computing features such as magnetic complexity and polarity distribution, and applying a Long Short-Term Memory (LSTM) model for time-series analysis. The prediction focuses on M-class or larger flares. Using data from 2011 to 2014, the method achieved a True Skill Statistic (TSS) of 0.763, demonstrating its effectiveness. This result highlights the feasibility of using low-resolution data for efficient and accurate solar flare forecasting.

We develop LSTM-based models to predict the maximum wind speed of tropical cyclones 24 hours in advance using global atmospheric simulation data. Models with forecast intervals of 3, 6, 12, and 24 hours are compared in terms of accuracy. Results show that a single 24-hour prediction model achieves the highest accuracy, while shorter iterative forecasts accumulate errors due to nonlinear atmospheric dynamics.

In the GeoSciAI2024 meteorology challenge, a regression model was proposed to predict maximum wind speed for 24 hours later of simulated typhoons using the provided track data and two-dimensional atmospheric data. The variables derived from current data and data from 12 hours earlier were used as explanatory variables. First, typhoons were stratified into 21 regions based on their location. Predictive models were made for each stratum and generalized additive model (GAM) were used for regions with sufficient training data, while LASSO regression was applied for regions with fewer than 300 training samples. Despite the stratification into 22 regions, it was observed in many areas that including latitude and longitude information in the generalized linear models tended to reduce the training mean squared error. These findings suggest that more detailed grouping based on location may enhance the accuracy of the model.

I propose multiple models for typhoon maximum wind speed forecasting using Neural Network. Based on RMSE evaluation for Test data set, the best score is 5.16 achieved with multi variate LSTM. In this model, the typhoons are expressed in multiple 1D variables by averaging 2D data over concentric region from center. The RMSE on the land regions are comparatively poor than that on the sea, which implies the model needs more improvement on dealing the typhoons especially at their weakening stages.