First arrival traveltimes picking of real data using MODWT-BiLSTM

In this work, we propose to combine the Maximum Overlap Discrete Wavelet Transform (MODWT) and Bidirectional Long Short-Term Memory (BiLSTM) networks for first-arrival picking. MODWT overcomes the limitations of traditional discrete wavelet transform (DWT) by preserving the signal length and eliminating aliasing artifacts, it also captures the high- and low-frequency features critical for first-arrival detection. The extracted multi-scale features are then input into a BiLSTM network, a specialized recurrent architecture that processes data in both forward and backward temporal directions, which enhances the robustness in noisy conditions and the ability to adapt to diverse seismic datasets. This approach provides an efficient and accurate solution for first-arrival picking, particularly in noisy environments and with zero-phase shot records. The proposed approach is tested using synthetic and field examples, where the results show accurate first-arrival traveltime picking values.