Delineation of Complex Basement Relief Simultaneously With Smooth and Rifted Features Using Gravity Data Based on CNN

Gravity inversion for basement relief is an important method to depict the structure of the sedimentary basin. At present, researches on this topic are mainly based on the assumption of minimum structure using different norms. However, in practical applications, it is often difficult for these methods to determine appropriate norms. Hence, errors are easily introduced, especially for complex relief. To solve this problem, this study presents a CNN-based gravity inversion method for complex basement relief with both rifted and smooth features. A multitask learning neural network has been designed, which can extract rifted and smooth features through different subnetworks. Then, a converged subnetwork is constructed to merge different features for getting undulations of the basement. The presented method is tested on a synthetic model. The results show that, compared with traditional L2-norm-based methods, this method exhibits improvement in accuracy and ease of use in recovering complex basement relief. Finally, the method is applied to gravity data collected in the Los Angeles Basin in California. The obtained structure shows good consistency with known structure.