GEOPHYSICS GUIDED DEEP IMPUTATION NETWORKS FOR PREDICTING 3D GEOSPATIAL DATA

Imputing 3D drill-hole data plays an important role for resource estimation and mine planning. Many interpolation models have been proposed including kriging and deep neural networks. Compared to these existing models, a locality-adaptive approach is preferable when the mineralization exhibits complex pattern and mineral distribution as an imputation target can be related to independent information such as the geophysical data. For example, the mineralization types such as volcanogenic massive sulfides are characterized by pronounced electrical conductivity, chargeability, and magnetic susceptibility responses.

We propose a novel interpolation model using the Convolutional Generative Adversarial Imputation Network (CGAIN) with low-rank and spatial smoothness regularizations. In addition to using the generator and discriminator of CGAIN to produce imputation from incomplete input data, the hinter component extracts the information from the auxiliary geophysical data and provides it to discriminator. As a result, the discriminator can better guide the generator to produce estimations to missing data by utilizing the geophysical information. The imputed 3D data is filtered by plausibility score thresholds, which are post-processed by the low-rank spatial consistency tensor completion method to discover the spatial continuity of mineral distribution. The proposed CGAIN is fully flexible for incorporating auxiliary information collected from the multiple sources, thereby expanding its applicability to the various application scenarios in mineral exploration. We have applied the proposed CGAIN on imputing drill-hole data collected from a volcanogenic massive sulfide deposit in British Columbia, Canada. We utilize the conductivity, chargeability, and magnetic susceptibility model as the hinting information to indicate possible areas of mineralization. The three geophysical models are derived respectively from 3D inversions of DC resistivity, induced polarization, and full tensor magnetic gradiometer data. In our experiments, we observe that the imputation performance is improved by integrating geophysical data.