Transfer Learning With U-Swin Transformer for Adaptive Ground Roll Attenuation in Seismic Records From Sabkha Environments

Ground roll is a dominant coherent noise in land seismic data, characterized by low frequency, low velocity, and high amplitude. often overlaps with reflection arrivals, thereby reducing the reliability of seismic imaging and interpretation. Existing deep learning approaches exhibit inherent limitations in ground-roll attenuation, since Convolutional Neural Network-based models are constrained by local receptive fields, while standard Transformer architectures require large training datasets and incur substantial computational and memory costs due to their global self-attention mechanism. In this study, we propose a U-shaped Swin Transformer (U-Swin Transformer) for ground-roll attenuation. The proposed model replaces global self-attention with window-based self-attention, achieving linear computational complexity with respect to the input size and thereby significantly improving efficiency and scalability. Two near-surface field datasets acquired from the inland sabkha of Jayb Uwayyid (SJUW) in the Saudi Kingdom and the coastal sabkha of Abu Dhabi (CSAD) are used to evaluate the method under realistic geological conditions. Experimental results demonstrate that the proposed U-Swin Transformer effectively handles aliased ground roll and signal-noise overlap, which remain challenging for the U-shaped neural network (U-Net) and F-K filtering. Although the denoising performance of the U-Swin Transformer is comparable to that of the Vision Transformer (ViT) with similar mean absolute error, peak signal-to-noise ratio, and structural similarity index measure, it achieves this performance with significantly lower computational cost and GPU memory consumption. Furthermore, incorporating a transfer learning strategy enables efficient adaptation of the U-Swin Transformer across different sabkha seismic environments, substantially reducing training time compared to training the model from scratch.