K-fold cross-validation of categorical variable predictions generated by 3D convolutional neural networks (3D-CNN): a case study of a gold deposit.

چکیده مقاله

For evaluating the quality and validity of geological and ore deposit models, as a crucial step before using these models, various methods have been proposed, each having its advantages and shortcomings. Cross-Validation (CV) techniques are among the common methods for evaluating models, especially in geostatistical techniques of continuous variables. The case is somehow different for categorical variables, considering their natures, scales, etc. In this study, the rock-type model of the Nohor-Miardan gold deposit in northwestern Iran was constructed using a deep learning neural network method called “3D Convolutional Neural Network” (3D-CNN). This is particularly essential, considering the key role of the lithological models in accurately delineating the gold mineralisation zones and estimating their corresponding assays. First, the corresponding 3D-CNN codes applied in this study were programmed in a MATLAB environment. After preprocessing the input dataset, the 3D-CNN models were constructed and evaluated using K-fold CV to ensure the reliability of the generated models. The method was implemented on a lithological dataset from 20 boreholes in the Nohor-Miardan epithermal gold deposit. The deposit which occupies an area of about 100,000 m2 is hosted within a volcanic system composed of Late Cretaceous bedrock units and intermittently overlying Eocene volcanics. Moreover, the plutonic rocks in the area are believed to be formed during the acidic Oligocene intrusions. The gold mineralisation was also created within these mainly andesitic rocks by the hydrothermal fluid injections into the Eocene volcanic rocks. The mineralisation occurred as quartz-sulphide veins of 20 cm to 2 m thickness (with Au grades reaching >10 ppm) and in highly fractured silicified zones (with about 1.2 ppm gold grade). The inputs of 3D-CNN included the rock types at the spatial locations of (x, y, z), and the outputs were the predicted rock type codes at the model grid cells. The appropriate prediction models were achieved in the training number of 18,000 in this study. The K-fold CV results reflected a high accuracy of 3D-CNN in this case and can be proposed for an appropriate evaluation of the categorical variable models in similar cases. The current study gives new insights into applying CNN and machine learning techniques in accurately predicting lithological domains in mineral deposits.