Abstract: Urbanization has led to increasing environmental noise contamination in field seismic data acquisition, among which vehicle noise is one of the major sources. Research on methods for suppressing vehicle noise in seismic data can not only improve the quality of seismic records but also expand the feasible areas for seismic data acquisition. Based on the U-Net deep learning framework, this study proposes a self-supervised vehicle noise suppression framework that specifically considers the temporal waveform and amplitude characteristics of vehicle noise. The framework is designed from four aspects: training strategy, network architecture, loss function, and evaluation metric. First, a training dataset preparation method based on a masking strategy is adopted, forcing the network to recover the missing signal information using contextual information. Then, to enhance the learning of seismic signal features, a RecResU-Net network that combines recursive and residual structures is employed to improve the feature extraction capability of the conventional U-Net. During this process, Huber loss and total variation (TV) loss are introduced to constrain the network in distinguishing vehicle noise components. Finally, a noise suppression evaluation metric based on local standard deviation is proposed to intuitively reflect the level of vehicle noise before and after denoising. Experiments on both synthetic and field seismic data demonstrate that the proposed method can effectively suppress vehicle noise to a certain extent.