Convolutional autoencoder with local wavefield characteristic constraint for suppressing near-surface seismic scattered waves

In mountainous seismic exploration, complex near-surface environments cause strong wave scattering, reducing the signal-to-noise ratio and complicating data processing. Therefore, suppressing scattered waves is crucial. To address this issue, this study proposes a convolutional autoencoder constrained by local scattered wave characteristics to suppress near-surface scattered waves (NSWs) in full-wavefield seismic data. This method uses the scattered waves predicted by seismic interferometry as the network input, and the original seismic records containing true scattered waves as the label. Since there are differences in amplitude and phase between the predicted scattered waves and the true scattered waves, the network introduces local wavefield features of the scattered waves for constraint, and adds a smoothness regularization term in the loss function to ensure the continuity of the output waveform. After training, the network maps the energy of predicted scattered waves into the actual seismic records, thereby accurately extracting scattered waves. Finally, by subtracting the network output from the original records, clean data with scattered waves removed can be obtained. This method is a self-supervised learning strategy and does not require additional clean signal samples. During training, the weights of each item in the loss function can be dynamically adjusted to guide the network to focus on local scattered-wave features, avoid learning effective wave information, and ensure that only scattered-wave components are retained in the output. Practical application results show that this method can effectively suppress NSWs and improve the signal-to-noise ratio of seismic data.