In the year 2023, China successfully built the world’s most extensive earthquake Early Warning Station Network. This network is characterized by its dense distribution of stations, which serve as a robust data foundation for various critical applications, including earthquake localization, seismic monitoring, earthquake detection, and numerous other related fields. In recent years, the application of deep learning methods in seismic phase picking has grown significantly, leading to the development of highly effective neural network models. Some of the most representative and widely recognized models in this domain include PhaseNet, LPPN, and CSESnet, all of which are based on convolutional neural network （CNN） architectures. Additionally, there is EQTransformer, which is a more advanced neural network model that combines convolutional neural network, recurrent neural network （RNN）, and Transformer-based architectures to achieve superior performance. These deep learning methods have consistently demonstrated their ability to outperform traditional phase picking approaches, particularly when trained on large-scale labeled datasets containing high-quality earthquake data. Despite the success of these deep learning-based models, they are predominantly trained using datasets that mainly consist of data recorded by velocity seismometer, which poses a limitation.

The earthquake Early Warning Network stations records data from three types of seismic instruments; velocity seismometer, accelerator, and seismic intensity meter （MEMS） instruments. The differences in data characteristics between these instrument types lead to a decline in the accuracy of current deep learning models when applied to data recorded by accelerometers or MEMS instruments. Furthermore, there is no large-scale, high-quality dataset recorded by early warning stations to train a neural network model for performing phase picking in Early Warning Network station data. To address the issue of declining accuracy in neural network models when processing data from Early Warning Station data, we leveraged transfer learning, a machine learning technique that has proven to be highly effective in scenarios where labeled data is limited. Transfer learning enables pre-trained model to transfer knowledge acquired in one domain to a different but related domain, thereby enhancing model performance with minimal additional data and training costs.

In our study, we utilized the Bidirectional Recurrent Neural Network （BRNN） model published by Yu et al （2023） as the pre-trained model. Then, we perform transfer learning using 26,739 pieces of manually labeled data from Sichuan region. The data included in these samples are from velocity seismometer, accelerator, and MEMS instruments. To explore the effect of training iterations on model performance, we trained three separate transfer learning models with iteration counts of 500, 1,000, and 10,000, respectively.

The evaluation results, based on standard performance metrics such as precision, recall, F1 score, and the mean and standard deviation of time residuals between model-predicted and manually labeled phase arrivals, demonstrated that transfer learning significantly improved the detection accuracy of both Pg and Sg phases recorded by early warning station. The transfer learning model achieved an F1 score of 0.899 for the Pg phase and 0.862 for the Sg phase, corresponding to improvements of 0.02 and 0.067, respectively, compared to the pre-trained model. An analysis of different training iteration counts revealed that model performance stabilized after approximately 500 iterations, with further increases to 1,000 and 10,000 iterations yielding only marginal performance improvements. To assess the robustness of the transfer learning model, we tested its performance across various settings, including different instrument types, signal-to-noise ratio （SNR） conditions, and epicentral distances. The results indicated that transfer learning enhanced the model’s generalization ability. For instance, under low SNR conditions ranging from 0 to 5 dB, the F1 scores for Pg and Sg phases improved significantly, increasing from 0.697 and 0.479 to 0.817 and 0.676, respectively. The transfer learning model achieved an F1 score of 0.891 for detecting Pg in accelerator data, an improvement of 0.019 compared to the pre-trained model. The detection accuracy for Sg showed the greatest improvement, with an F1 score of 0.852, an increase of 0.102 compared to the pre-trained model. Additionally, detection results for individual earthquake events revealed that the pre-trained model exhibited significant failures to detect Sg phase at epicentral distances exceeding 80 km. In contrast, the transfer learning model effectively reduced these misses, achieving marked improvements in Sg phase detection for such cases. Furthermore, we evaluated the performance of the transfer learning model on the pre-trained dataset to verify its generalization ability. The results showed that the transfer learning model improved detection accuracy for the pre-trained dataset. Specifically, the F1 scores for Pg and Sg phases increased to 0.842 and 0.866, representing improvements of 0.024 and 0.045 compared to the pre-trained model, respectively.

In summary, this study demonstrates that transfer learning significantly reduces the reliance on large volumes of manually labeled data, thereby lowering the associated training costs and resource requirements. The transfer learning method in this study included data from different instrument types, enriching the training samples and enabling the model to learn a broader and more diverse range of waveform characteristics, thereby improving model’s performance. The transfer learning model developed in this research can better utilize the data recorded by early warning stations for related research, such as velocity structure inversion and seismic activity analysis.