Irregularly Spatial Seismic Missing Data Reconstruction Using Transformer With Periodic Skip Connection.

Seismic exploration is one of the main methods used in geophysical exploration. However, due to the limitations related to high exploration costs and the environmental conditions, the obtained seismic traces may not be uniformly sampled along the spatial direction, which significantly impacts the processing and interpretation of seismic data. The conventional interpolation methods are not sufficiently accurate and robust, while some of them are inefficient. Furthermore, several authors have proposed deep learning (DL) methods based on convolutional neural networks (CNNs), which improve the efficiency of interpolation but also result in poor interpolation effects and poor robustness. To achieve enhanced robustness and ensure the efficiency of the interpolation method, we used transformer as the backbone network. Through patch segmentation and skip connections, the operational efficiency and effect of the model are guaranteed. Furthermore, we implemented self-supervised training strategy to mitigate the impact caused by limited dataset generalization. Experiments conducted on irregularly sampled spatial field ocean and land seismic data demonstrate that our proposed method has better robustness and accuracy than competing approaches. In addition, the efficiency of our method is much higher than that of competing conventional approaches. Thus, our method provides improvements that have important implications for geophysical exploration and the application of transformers in the field of seismic data interpretation.