Hybrid CNBLA architecture for accurate earthquake magnitude forecasting

Abstract

Earthquake prediction in seismology is challenging due to sudden events and lack of warnings, requiring rapid detection and accurate parameter estimation for real-time applications. This study proposed a novel automatic earthquake detection model to enhance the processing and analysis of seismic data. The hybrid model comprises convolutional layers, normalization techniques, bidirectional long short-term memory (Bi-LSTM) networks, and attention mechanisms, collectively referred to as the hybrid convolutional–normalization–BiLSTM–attention (CNBLA) model. The attention mechanism allows the model to focus on critical segments of seismic sequences, while layer normalization stabilizes training by normalizing activations, thus reducing the effects of input scale variations. This dual approach mitigates the impact of input scale variations and enhances the model’s ability to effectively decode complex temporal patterns. The hybrid CNBLA model optimizes the extraction and processing of temporal features from raw waveforms recorded at single stations, thereby improving the accuracy and efficiency of seismic magnitude estimation. The proposed model is evaluated using two datasets: the STEAD and USGS achieving a mean square error (MSE) values 0.054 and 0.0843 and a mean absolute error (MAE) 0.15 and 0.2526 respectively. The hybrid CNBLA model outperforms two baseline models and five state-of-the-art approaches in earthquake magnitude estimation, improving seismic monitoring and early warning systems.