Improving Porphyry Copper Prospectivity Mapping with Stacked Ensemble Machine Learning and Feature Contribution Analysis: A Case Study of the Eastern Sirjan 1:250,000 quadrangle map, Iran

Abstract. Mineral prospectivity mapping (MPM) is an essential tool for reducing exploration costs and risks by identifying areas with high mineralization potential. This study employed Support Vector Regression (SVR) and a multilayer perceptron-based Artificial Neural Network (MLP-ANN) to predict porphyry copper prospectivity. To improve accuracy and minimize predictive uncertainty, a stacked ensemble learning strategy was applied, in which the predictions of the two base models served as inputs to a meta-learner. Gaussian Process Regression (GPR) was selected as the meta-learner due to its nonparametric nature and ability in quantifying uncertainty. In an extended version, the original input features were also incorporated together with the SVR and ANN outputs, thereby enhancing the diversity and richness of meta-model’s input space. The study area covers the eastern part of the Sirjan 1:250,000 quadrangle map within the Kerman Cenozoic Magmatic Arc (KCMA). Model performance was evaluated using correlation metrics (R and R²), error indicators (MAE, MSE and RMSE), and prediction–area (P–A) analysis. Results showed that the stacked ensemble improved predictive performance compared to individual models. Feature contribution analysis further revealed that GPR could identify negatively contributing features. Removing these features led to a refined GPR model under feature filtering (so-called FF-GPR), which increased the R value to 0.98 and the R² value to 0.96. This refinement also improved the accuracy of P–A plot and reduced uncertainty. These findings highlight the effectiveness of feature-augmented stacked ensemble learning, particularly when combined with GPR, in advancing MPM and supporting future exploration.