3-D Gravity Intelligent Inversion by U-Net Network With Data Augmentation

Abstract

Gravity inversion has become the main method to obtain the 3D density distribution in the exploration of petroleum and mineral resources as well as the study of geological structure. In recent years, with rapid development of deep learning algorithms, gravity intelligent inversion based on convolutional neural networks have also achieved certain good results. However, in the previous study on 3D gravity intelligent inversion, some issues in "data-driven" inversion still need to be faced, that is, insufficient number and lack of the diversity of data sets. This paper selects a suitable U-net network with four-layer down-sampling structure, designs typical block and inclined plate and their combined models to construct the data sets. In order to reduce the overfitting of the network, the data augmentation is introduced to improve the training process for datasets. The data augmentation method greatly alleviates the limited number and types of samples, allowing the network to learn more situations and improve the generalization ability. Theoretical model data tests show that the data augmentation method of Mixup and flip can effectively improve the generalization ability of U-net network, and the noise-adding augmentation has good inversion result for noisy data. The tests on the measured gravity data also verify the effectiveness of the proposed method, which indicates that data augmentation does play a significant role in the improvement of gravity intelligent inversion.