An analytical study on the effect of inverse prediction based on two types of convolutional neural networks NestU-Net and ResU-Net++

Exploring new methods for gravity anomaly inversion to quickly and accurately understand the distribution and physical properties of subsurface space is the focus of current research. In this paper, we analyse the significant differences in the prediction results of different network structures for the same subsurface information from machine learning, and highlight the differences in the theoretical model prediction recovery effects of four U-Net networks (AttU-Net, NestU-Net, R2U-Net, ResU-Net++). Two of the more effective networks, NestU-Net and ResU-Net++, were selected and applied to 3D density imaging of the Mobrun sulphide ore body in Noranda, Quebec, Canada. By combining the previous inversion results and comparing the prediction effect of the two inversions, it is concluded that ResU-Net++ is more effective than NestU-Net network application in the work zone of sulphide ore body, and it also provides a new theoretical and methodological support for the prediction and restoration of the geophysical actual subsurface geological information.