Abstract:  To fully utilize the revealed hydrogeological information for the generation of deep hydrogeological profiles, a generation method of deep hydrogeological profiles based on conditional generative adversarial network (CGAN) was proposed. Firstly, the shallow geological profiles were segmented, the revealed geological profile data were extracted as training samples, and the corresponding virtual boreholes were generated as conditional data to construct the sample library for the conditional generative adversarial network. Subsequently, the conditional generative adversarial network was constructed and trained. The U-Net architecture was adopted by the generator, a multi-scale convolutional network was employed by the discriminator, and a multi-scale discriminative loss function was designed. Then, the deep borehole data were input into the well-trained generator to effectively generate the deep hydrogeological profile of the target area. Finally, a comparative experiment was conducted with the XGBoost model based on iterative convolution.The results show that the CGAN method presents obvious advantages in the detail and structural representation of profiles, with more complete structure and higher accuracy of generated profiles. When inputting the same six boreholes, compared with the XGBoost method, the peak signal-to-noise ratio (PSNR) is averagely increased by 1.165 2 dB, the structural similarity index measure (SSIM) is averagely improved by 0.051 0, and the learned perceptual image patch similarity (LPIPS) is averagely reduced by 0.023 8. When six deep boreholes are input, the profile generated by the proposed model is highly consistent with the hydrogeological profile of the mining area compiled by integrating multiple types of data.

Key words: hydrogeology, drilling data, deep learning, conditional generative adversarial network (CGAN), profile generation