A Global-Local collaborative approach to quantifying spatial non-stationarity in three-dimensional mineral prospectivity modeling

Abstract

The three-dimensional (3D) modeling of mineral prospectivity plays a pivotal role in the mineral exploration process, enabling explorers to identify, evaluate, and prioritize deep-seated targets more effectively.Accurately predicting the 3D distribution of mineral prospectivity relies on effectively quantifying correlations to mineralization. The ore formation process is extremely intricate, marked by spatial non-stationary correlations. While state-of-the-art 3D mineral prospectivity modeling methods use machine learning (including deep learning) to capture correlation between geological evidence and mineralization hidden in the geosciences data, limited attention has been given to quantifying the local non-stationary nature in ore formation. This paper proposes a novel method that adapts to local spatial non-stationarity in 3D mineral prospectivity modeling. The approach integrates a prospectively model with a local regression model, creating a global–local collaborative model to quantify spatial non-stationary associations with mineralization. Random Forest (RF) and Reset are employed to construct mineral prospectivity models that quantify the global associations with mineralization. Simultaneously, the geographically weighted regression (GWR) model serves as a local model to capture the local non-stationarity within mineralization association. An Expectation Maximization algorithm is tailored to bridge the gap between global and local models, resulting in an iterative framework that progressively fuses global spatial associations and local non-stationarity related to mineralization. The iterative process generates a mineral perspectivity model that adapts to local non-stationarity in predictive areas. A case study on the Dingjiashan Pb-Zn deposit validates the effectiveness of the proposed method. The results demonstrate that the global–local collaborative mineral perspectivity model significantly boosts prediction accuracy in predictive areas. Moreover, the regression parameters of GWR derived from the model construction process effectively reveal the spatial non-stationary impacts of different ore-controlling factors, thereby improving the interpretability and reliability of predicted mineral prospectivity.