A Multitask Deep Learning for Simultaneous Denoising and Inversion of 3-D Gravity Data

Noise present in real gravity data can lead to inaccurate inversion results. Multitask strategy in deep learning provides a promising method to solve this problem. In this study, a multitask framework is proposed for simultaneous inversion and denoising of noisy gravity data, in which the denoising task can constrain the inversion task. To extract multiscale field information for high-precision inversion, a novel backbone network, known as the cross-dimensional UNet (CDUNet), is proposed. CDUNet employs cross-dimensional skip connections to transfer different scale field features, wherein transformation modules are used to convert 2-D features extracted from gravity data to 3-D features for density model reconstruction. A noisy dataset was synthesized to train the network, which comprised diverse density models having highly random geometric and physical characteristics. The test set evaluations showed that CDUNet and the multitask framework could integrate well, and the inversion accuracy of the network could reach 70.8% over the density perturbation area and 97% over the entire area. The synthetic examples showed that the inverted models are characterized by distinct boundaries and relatively accurate values. Finally, the method was validated using real data from the Vinton salt dome in Texas and Louisiana, USA.