ABSTRACT Complex heterogeneities in the Earth's shallow subsurface can generate significant scattering noise. This noise obscures and distorts seismic reflections, reducing the signal‐to‐noise ratio (SNR) and further complicating subsequent processing steps such as static correction, velocity analysis and multiple attenuation. The geological assumption of small‐scale, low‐velocity bodies in the weathered layer provides a basis for constructing models that represent much of the land data. Using Gaussian random fields, we constructed models containing randomly distributed small‐scale heterogeneities in the near‐surface layer, which provided data similar to those we encounter in land, with low SNR. To reconstruct the distorted reflections from shallow scattering‐based noisy data, we propose an unsupervised scheme based on the Gabor Dictionary Learning Network (GDLNet). GDLNet is an unrolled optimization model for natural image denoising that offers strong performance and structural interpretability. Its design enables flexible control over the physical characteristics of the convolution kernels through parameter tuning, enhancing adaptability to diverse noise patterns. We analyse the challenges of applying GDLNet to scattering‐based noisy seismic data, including anisotropic resolution causing signal leakage and limited lateral continuity. To address these issues, we enhance the network with resampled Gabor kernels and directional convolution layers, and adopt a decaying learning rate strategy to prevent overfitting to noise. The improved network directly predicts the reflection signals from noisy input, and the scattering noise is subsequently obtained by subtracting the predicted reflections from the original data. For each shot gather, the proposed method requires only a few seconds of training, achieving computational efficiency comparable to conventional approaches. The proposed method is tested on both synthetic and field data, demonstrating the scheme's effectiveness in reducing scattering noise.
Zhou et al. (Fri,) studied this question.