All ML models followed a common workflow: imputation of missing values (primarily using the median), stratified splitting of the data into 70% for training and 30% for validation and using the location of known deposits ("yac") as the target variable for prediction. Grid cells intersecting known deposits were labeled as positive (“yac” = 1), and remaining cells as negative (0). Although the response variable was binary, models generated continuous prospectivity scores, enabling evaluation with regression metrics (R², MAE, RMSE) and classification metrics (AUC, Precision, Recall, F1) using the threshold on the validation dataset. The selected algorithms represent diverse modeling paradigms, including ensemble tree methods, kernel-based methods, distance-based classification, and semi-parametric regression models. Algorithms, key hyperparameters, and performance metrics are detailed below: 2.3.1 Random Forest (RF): Random Forest was implemented as an ensemble tree-based method using 50 decision trees (ntree = 50). The model was trained using bootstrapped samples, and variable importance was internally assessed through the standard Random Forest framework (Figure 3). 2.3.2 Generalized Additive Model (GAM): The Generalized Additive Model (GAM) was implemented using smooth spline functions fitted via the Restricted Maximum Likelihood (REML) method, including a spatial smoothing term to account for potential spatial autocorrelation, following the framework described by Pedersen et al. (2019). The effective degrees of freedom estimated to control smoothing complexity (Figure 3). 2.3.3 Extreme Gradient Boosting (XGB): Extreme Gradient Boosting (XGB) was implemented using 100 boosting iterations (nrounds = 100), a maximum tree depth of 6 (max_depth = 6), and a learning rate of 0.1 (eta = 0.1). Gradient boosting was applied to optimize classification performance by minimizing the loss (Figure 4). 2.3.4 LightGBM (LGB): LightGBM (LGB) was configured with 100 boosting iterations, a maximum of 31 leaves per tree (num_leaves = 31), and sampling parameters (feature_fraction = 0.8; bagging_fraction = 0.8) to improve efficiency and reduce overfitting by using histogram-based learning to accelerate training (Figure 4). 2.3.5 K-Nearest Neighbors (KNN): The K-Nearest Neighbors (KNN) algorithm was implemented using k = 10 neighbors and Euclidean distance as the similarity metric. Classification was based on majority voting among the nearest samples in feature space (Figure 5). 2.3.6 Support Vector Regression (SVR): Support Vector Regression (SVR) was implemented using a radial basis function (RBF) kernel with parameters C = 1, gamma = 0.01, and epsilon = 0.15. To address class imbalance, a weighting factor of 3 was assigned to class (“yac”). The model produced continuous prediction scores later converted into binary classifications through threshold optimization (Figure 5). The dataset exhibited class imbalance due to limited number of mineralized relative to non-mineralized cells.
RkJQdWJsaXNoZXIy MTM0Mzk2