Self-supervised suppression of vehicle noise in seismic data using a recursive residual U-Net

Urbanization has increasingly exposed field seismic acquisition to environmental noise, among which vehicle noise is one of the dominant sources. Investigating vehicle noise suppression methods for seismic data is important not only for improving data quality but also for extending the applicability of seismic acquisition in complex urban environments. Based on the U-Net deep learning framework, this study proposes a self-supervised vehicle noise suppression framework by incorporating the temporal waveform and amplitude characteristics of vehicle noise. This framework is developed from four aspects: training strategy, network architecture, loss function, and evaluation metric. First, a mask-based training set construction strategy is adopted, enabling the network to recover missing signals using contextual information. Then, a recursive residual U-Net, RecResU-Net, is introduced to enhance the feature extraction capability of the standard U-Net. In addition, the Huber loss and total variation loss are employed to guide the network to learn seismic signal features while suppressing vehicle noise. Finally, an evaluation metric based on local standard deviation statistics is proposed to assess the vehicle noise suppression performance and visually reflect the changes in noise components before and after denoising. The results on synthetic and field data demonstrate that the proposed method can effectively suppress vehicle noise to a certain extent.