Reconstruction of Consecutively Missing Seismic Data Using a Partial Convolution Network

High‐quality, regular sampling is a fundamental prerequisite for seismic exploration. However, environmental constraints and other factors often impede the uniform deployment of sources and receivers, leading to substantial gaps in seismic records, particularly consecutively missing traces. To address this challenge, we propose a deep learning method based on the U‐Net framework specifically designed to reconstruct consecutively missing seismic data. Our approach replaces standard convolutional layers with partial convolutions (PConvs), which adaptively assign weights to masked and unmasked regions. This mechanism enables the network to effectively capture the distribution characteristics of consecutive traces and the textural features of the missing data. Furthermore, we integrate attention modules into the skip connections between the encoder and decoder to enhance feature representation. We validated the method using synthetic seismic data and applied a transfer learning strategy to field marine seismic data. The results demonstrate that our knowledge‐consistent attention PConv U‐Net accurately reconstructs consecutively missing traces even with limited field training samples, underscoring the practical utility and generalization potential of the proposed approach.