Abstract: To improve the precision and robustness of P-wave picking in microseismic monitoring, this study develops an automatic picking method based on a dual-branch fusion and lightweight deep regression network. The model employs parallel time- and frequency-domain branches to extract temporal dependencies and spectral features, respectively. Specifically, the time-domain branch leverages convolutional neural network–bidirectional long short-term memory network（CNN-BiLSTM）to model waveform dynamics, while the frequency-domain branch employs Mel-frequency cepstral coefficients (MFCC) with two-dimensional convolutional neural network (2D CNN) for robust spectral representation. Both branches integrate squeeze-and-excitation (SE) channel attention channel attention and are adaptively fused through a gated module. The fused features are fed into a lightweight multilayer perceptron (MLP) regression layer to produce a probability curve for precise P-wave onset localization. Experiments on 1 028 real records demonstrate that the proposed method attains a precision of 79.9% within a 5 ms tolerance range, outperforming conventional AIC and STA/LTA methods and maintaining stable performance under low signal-to-noise ratio (SNR) conditions. This method integrates a lightweight design with high precision, thereby offering a reliable tool for phase identification and event localization in microseismic monitoring.

Key words: microseismic monitoring, P-wave picking, MFCC features, BiLSTM, regression network, deep learning