PROPOSED CONVBILSTM-NET MODEL FOR ENHANCING EARTHQUAKE PREDICTION PERFORMANCE USING SPATIOTEMPORAL FEATURES

Abstract

Accurate earthquake prediction remains a significant challenge due to the complex spatiotemporal dependencies inherent in seismic events. To address this issue, the present study proposes ConvBiLSTM-Net. This hybrid deep learning model combines Convolutional Neural Networks (CNNs) for spatial feature extraction with Bidirectional Long Short-Term Memory (BiLSTM) networks for temporal sequence modeling. The model integrates historical earthquake data with spatial information in the form of fault density (FD), derived using Kernel Density Estimation (KDE). The KDE bandwidth is optimized using the Bivariate Local Indicator of Spatial Association (LISA) method to enhance spatial adaptivity. The dataset comprises earthquake records from the USGS catalog (1974–2023) and active fault data compiled in the 2017 Indonesian Earthquake Source and Hazard Map, published by the National Earthquake Study Center (PuSGeN). ConvBiLSTM-Net is evaluated under short-term and medium-term prediction scenarios, targeting earthquake magnitude, depth, and epicenter coordinates (latitude and longitude), using standard performance metrics such as accuracy, F1 score, root mean square error (RMSE), mean absolute error (MAE), and the coefficient of determination (R2). In the short-term scenario, the model achieves average improvements of 9.31% in R2, 3.41% in accuracy, and 6.06% in F1 score, while reducing RMSE by 10.63% and MAE by 12.40% across magnitude, depth, and latitude predictions. For longitude, R2, accuracy, and F1 score also improve by 10.88%, 11.76%, and 17.54%, respectively, although RMSE and MAE increase by 13.09% and 20.74%, indicating a tradeoff between enhanced pattern recognition and higher absolute error. Under the medium-term scenario, the model demonstrates average improvements of 7.49% in R2, 3.39% in accuracy, and 7.06% in F1 score, while reducing RMSE and MAE by 6.22% and 17.72%, respectively, for magnitude, depth, and latitude predictions. For longitude, R2, accuracy, and F1 score improve by 12.50%, 2.48%, and 1.60%, respectively, though RMSE and MAE increase by 37.31% and 37.01%, again highlighting a trade-off between better pattern recognition and increased absolute error in this dimension. These findings demonstrate that ConvBiLSTM-Net, engineered to integrate spatial and temporal features, is a robust and adaptive architecture for enhancing earthquake prediction performance. Its spatiotemporal modeling approach yields consistently high accuracy and stability across forecasting horizons, particularly in predicting earthquake epicenters. Despite minor trade-offs in absolute error for longitude predictions, the overall performance improvements affirm its potential as a reliable tool for seismic hazard assessment and disaster risk mitigation.