A Global Earthquake Prediction Model Based on Spherical Convolutional LSTM

Abstract

Utilizing deep learning techniques for earthquake prediction, within the context of defining it as a global-scale spatiotemporal forecasting issue, has led to enhanced outcomes in contrast to previous methods, but the challenges of spatial distortion have arisen and persisted. This article proposed a new model based on the spherical convolutional LSTM and the U-Net framework. The spherical convolutional LSTM is formed by incorporating spherical convolutional neural networks (CNNs) into the long short-term memory (LSTM) architecture, aiming to address spatial distortion challenges in global-scale earthquake prediction. To assess its effectiveness, we generate earthquake distribution maps using latitude and longitude to construct the dataset with the definition of global earthquake prediction as a spatiotemporal series problem. We conducted two experiments using datasets of map sizes 1920 and 3840, comparing our results with previous studies. Precision and recall along with other metrics are used to evaluate the model’s performance. Our findings demonstrate significant performance gains in Experiment 1 (map size = 1920) with (63.29%, 49.26%) for precision and recall than previous (57.18%, 50.65%). Moderate enhancements in Experiment 2 (map size = 3840) than previous studies are achieved with (64.86%, 51.85%) than previous (64.54%, 51.83%). Notably, the results highlight the potential of the proposed model and original spherical CNNs in mitigating spatial distortion issues in the global earthquake prediction problem.