Spatial modelling of analysis results such as grade, geochemical properties and density is required for resource estimation in almost all commodities. For spatial modelling of most commodity properties, geostatistical methods are the most popular because they are usually more accurate than deterministic methods, can quantify uncertainty and can use auxiliary information to improve predictive accuracy. However, geostatistical methods have the primary disadvantages that they have onerous preprocessing steps such as variogram modelling, rely on rigid statistical assumptions, and incorporation of numerous auxiliary variables which have non-linear relationship with the target variable is difficult. To address these disadvantages, a machine learning method based on quantile regression forest algorithm is proposed as an alternative approach for spatial modelling. This newly proposed method (termed geographic quantile regression forest) does not require variogram modelling, can quantify uncertainty and can easily incorporate numerous auxiliary information of differing data type. To evaluate the performance of the new method, the accuracy of predictions of coal relative density is compared to inverse distance weighting and regression kriging. Data from an active mine site in the Bowen Basin, Queensland Australia is used for the comparison. Using evaluation metrics from leave-one-out cross-validation, this paper demonstrates that the geographic quantile regression forest method has higher accuracy than inverse distance weighting and similar or higher accuracy than regression kriging accuracy across all geological domains. The high accuracy, similar performance to regression kriging and stated advantages over geostatistical methods makes it a candidate for future inclusion in geological model packages.

Open-Access Online Publication: March 01, 2023