LFTNet: A Lightweight Multi-Scale Attention Network for Real-Time Seismic Event Detection and Phase Picking

Abstract

Although deep learning has advanced seismic analysis, accurately identifying seismic phases in noisy or complex waveform environments remains challenging. Many existing models struggle with high computational cost, reduced accuracy, and poor robustness under low signal-to-noise ratio (SNR) conditions, limiting their use in real-time applications. To address these issues, we propose LFTNet, a lightweight fully convolutional temporal network employing a multi-task learning framework to simultaneously perform seismic event detection and P- and S-phase picking. This joint optimization approach leverages shared contextual information across tasks, improving accuracy, reducing redundancy, and enhancing robustness under diverse seismic conditions. LFTNet features two novel modules: (a) the Residual Separable Depthwise Block (RSDB), a lightweight module for efficient local feature extraction; (b) the Multi-Scale Squeeze-Excitation Temporal Convolutional Network (MSSE-TCN), a multi-scale attention mechanism designed to accurately detect seismic phases by capturing long-range temporal dependencies. Experiments on the STEAD and INSTANCE data sets demonstrate that LFTNet achieves state-of-the-art performance, with F1 scores up to 98.6%/98.1% (P-phase/S-phase) on STEAD and 88.1%/84.4% on INSTANCE, while reducing model parameters by approximately 31% and increasing inference speed by about 24% compared to EQTransformer. LFTNet also maintains high accuracy and robustness under noisy conditions, providing a reliable solution for real-time earthquake detection and early warning.