Abstract:
Due to factors such as the performance of acquisition equipment and stratigraphic absorption in marine seismic exploration, the frequency spectrum of acquired data is often concentrated within a specific frequency band, with missing high‑ or low‑frequency components. This leads to poor event continuity and reduced resolution in seismic records, thereby compromising the accurate characterization of subsurface geological structures. Bandwidth extension techniques based on deep learning have been widely studied because they overcome many limitations of conventional methods, such as sensitivity to noise, difficulty in obtaining Q values, and reliance on sparsity assumptions. However, most existing supervised learning approaches rely on external broadband data for training. To address this issue, propose a self‑supervised learning method for seismic data bandwidth extension. In this method, the measured data are first filtered to generate narrowband seismic data with different bandwidths. These narrowband data are then mapped to relatively wider frequency bands, and multiple band‑to‑band mapping relationships are constructed within the dataset, enabling full extraction of features from data with varying bandwidths. After the model is trained to an optimal state, bandwidth extension is achieved by directly feeding the original acquired seismic data into the trained model. Experimental results on both synthetic and field data demonstrate that the model trained with the proposed self‑supervised learning strategy can effectively broaden the frequency band of the target data.