PREDICTING 3D GEOSPATIAL DATA USING MACHINE LEARNING-BASED IMPUTATION METHODS

Interpolating and estimating geoscientific data in 3D is fundamental in geoscience, especially in resource exploration and mining. Since its development in 1960s, the kriging method has been a ubiquitous tool for this purpose. However, recent advances in machine learning (ML) techniques such as matrix completion and artificial neural network-based imputation methods may provide a new and more data-adaptive alternative. Instead of relying on the implicit Gaussian assumption expressed through a prescribed variogram by kriging, ML imputation allows us to discover the spatial continuity from the available sparse data such as precious and base metal assay data from drill holes and use that information for intelligent interpolation and estimation. In addition, the ML imputation approaches are fully capable of incorporating auxiliary information as an additional input in the neural network, which could be advantageous compared to the traditional co-kriging approach.

In this paper, we will present the Convolutional Generative Adversarial Imputation Nets (CGAIN) for missing data imputation. Generative Adversarial Imputation Nets (GAIN) generalizes the well-known generative adversarial nets (GAN) for accurately imputing missing data. CGAIN adds convolutional layers to GAIN, such that it can estimate unobserved data without knowing the truth of the complete data distributions. Compared to traditional data interpolation models, such as kriging, CGAIN can significantly improve the flexibility of the learning model, thereby expanding its applicability to different application scenarios. Moreover, by further training a hinter in the neural network, auxiliary geophysical information and additional drill-hole data collected from multiple sources can be utilized to improve the accuracy of imputed data. We have applied the CGAIN with hinters on both simulated data and real-world data for grade estimation. Clear performance gains have been observed in estimating the true data distributions from partially observed input data.