Machine Learning-Based Spatial Modeling of Stratified Mineral Deposits: Verification by Mine Survey Measurements

Abstract

This paper addresses the problem of improving the reliability of spatial prediction of mining and geological conditions for coal seam extraction. The aim of the study is to develop an automated methodology for modeling coal seam properties using machine learning techniques and to evaluate its predictive accuracy based on actual mine survey measurements. Using exploration borehole data, digital models of seam thickness, ash content, and sulfur content were developed with the GradientBoostingRegressor algorithm combined with Gaussian filtering. A comprehensive Mining Attractiveness Index (MAI) was also calculated. Validation against mine survey measurements from four development roadways showed high predictive accuracy for intrinsic seam properties, with seam thickness errors ranging from −0.40 to +0.35 m. However, systematic discrepancies of up to 5 m were identified in the spatial positioning of the predicted seam surfaces. These deviations were attributed to exploration borehole deviation and inconsistencies between exploration and underground survey coordinate systems rather than limitations of the machine learning algorithm. The results demonstrate the need for dynamic data fusion to continuously update geological models using mine survey measurements acquired during roadway development.