Abstract Using 8 779 event waveforms recorded at 55 seismic stations of Hubei seismic network from 2013-01 to 2021-07 as the dataset, the optimal model of convolutional neural network(CNN) is trained for seismic detection, and the recurrent neural network(RNN) method is used for seismic phase arrival time picking. The advantages and shortcomings of the RNN model, AR-Pick, STA/LTA algorithms are compared and analyzed, and the relationship between the picking results of RNN model and the signal-to-noise ratio, epicentral distance and magnitude are further analyzed. The results show that the convolutional neural network trained with a small number of datasets can also have excellent classification performance, and the RNN model is with lower accuracy in P-wave and S-wave phase arrival time picking compared with AR-Pick, STA/LTA algorithms, and its seismic phase picking results are more strongly correlated with signal-to-noise ratio of waveforms, and have little relationship with epicentral distance and magnitude of earthquakes. Therefore, for regions with fewer earthquakes, the optimal CNN model can be trained for seismic detection using fewer seismic waveform data, and the P-wave phase arrival is picked up by AR-Pick algorithm, and the S-wave phase arrival is picked up by STA/LTA algorithm, which can improve the accuracy of seismic detection and phase picking.