Improving generalization performance of deep learning–based seismic data interpolation

Abstract

Seismic data interpolation techniques are vital for preprocessing, as spatial undersampling in seismic data presents processing challenges. Recently, multiple deep learning–based interpolation techniques have emerged, each catering to distinct missing data scenarios, including regular, irregular or large gaps. However, this standardized approach can induce a creeping overfitting issue in terms of various missing types, notably undermining the generalization capability of trained deep learning models. It is worthy of serious consideration for performance generalization of deep learning–based trace interpolation in terms of various missing patterns. This study introduces an innovative approach, redefining deep learning–based seismic data interpolation to focus on enhancing generalized performance be treating unseen data. We highlight how data biases in the training dataset substantially impair interpolation performance on target data with varying features. Then we offer some guidelines to counter these biases during training dataset construction. Furthermore, we propose a versatile, single deep learning model applicable to any case of missing data in real-field situations, utilizing U-Net3+ as the backbone. Experiments using field data considering various missing scenarios reveal that our method excels in interpolating unseen target data; it does this by using an unbiased dataset, bolstering general interpolation performance. This study emphasizes the importance of a systematically designed training dataset to augment generalization in deep learning–based interpolation and indicates the need for more comprehensive research to create a universally applicable deep learning–based seismic data interpolation network for practical use.