A Seismic Phase-Picking Model Transfer Learning Approach Based on Maximum Mean Discrepancy

Thanks to the abundance of earthquake data and the advancement of deep neural networks, many seismic phase-picking models have been proposed to detect the arrival time of P-waves and S-waves. Currently, the accuracy of these models exceeds that of traditional methods, such as short-term average/long-term average (STA/LTA). While many studies concentrate on phase picks on a single seismic dataset, there is limited research regarding the generalization of phase-picking models. Specifically, applying a model trained on one seismic dataset to an unseen dataset remains underexplored. A phase-picking model is designed in this paper, which consists of an encoder and a decoder. The self-attention is applied in the encoder module to focus seismic features. Then a seismic phase-picking model transfer learning approach based on maximum mean discrepancy (SPPTL-MMD) is proposed to train a universal seismic feature extraction module by utilizing diverse seismic datasets. The module will be transferred and a significantly improved phase-picking model is obtained through fine-tuning techniques. Experiments show that the model after fine-tuning can be generalized to other datasets.