Abstract: In order to solve the discontinuous problems in the seismic first-break breaking using the U-Net network, a variantone based on level set method is proposed. Firstly, the feature map of seismic image is extracted based on U-Net network, and the loss is calculated pixel by pixel. Then the level set loss is calculated by using the level set method. Finally, the weighted sum of the two is taken as the final loss function. The improved network uses the level set method to ensure the continuity of first breaks between the adjacent seismic channels, and retains its ability of "end-to-end" fine classification. The Focal Loss function is used to alleviate the imbalance of sample categories of the training set. The mixed training set combined with the synthetic data and the actual field data is used to overcome the training set insufficiency problem. The results tested by the theoretical data and actual seismic data show that the improved U-Net network can improve the picking accuracy and robustness in low signal-to-noise ratio seismic data.