Self‐Supervised Seismic Data Reconstruction With Blind‐Trace Mask Mapper and Attention

ABSTRACT High‐quality seismic data are crucial for interpreting subsurface structures. Due to economic costs, geological obstacles and other factors, the acquired seismic data are often sampled regularly or irregularly along spatial coordinates. Such incomplete data pose significant challenges to subsequent seismic data processing workflows, making seismic data reconstruction one of the most cost‐effective and critical techniques to address this issue in seismic exploration. While deep learning has yielded promising results in data reconstruction, supervised methods rely heavily on large volumes of labelled seismic data – a resource that is scarce in practical applications. To tackle this limitation, this study proposes a self‐supervised learning method for seismic data reconstruction, incorporating a global‐aware blind‐trace mask mapper and squeeze‐and‐excitation (SE) attention blocks. First, we introduce a global‐aware blind‐trace mask mapper, which mitigates the identity mapping problem in self‐supervised learning frameworks and leverages global information to enable efficient reconstruction. This mapper samples all traces at the blind‐trace positions on the interpolated volume and maps them to the same channel, allowing the loss function to optimize all blind traces simultaneously. Second, we adopt a nested U‐Net (UNet++) architecture integrated with SE attention blocks as the interpolation network, enhancing the model's capacity to capture key features. Third, we guide network training using a hybrid loss function combining mean absolute error and structural similarity index, which captures both fine‐grained details and global structural information to ensure reconstruction quality. Experiments on synthetic and field datasets demonstrate that our method outperforms existing approaches in interpolation accuracy.