deposits and Class 02: Absence of known mineralization. Model performance was evaluated using Precision, Recall, and F1-score metrics. The workflow for data preprocessing, rasterization, feature standardization, model training, and validation is illustrated in Figure 4. Figure 4 - Workflow diagram illustrating the application of the Random Forest and artificial neural network (multilayer perceptron, MLP) algorithms used in this study. Source: Marchena et al. (2025). To evaluate model robustness and reduce overfitting, k-fold cross-validation was implemented during the training stage. The dataset was randomly partitioned into k subsets, where each subset was iteratively used as validation data while the remaining k−1 subsets were used for training. Performance metrics, including Precision, Recall, and F1-score, were averaged across all validation iterations to assess model stability and generalization capacity. Because geological datasets commonly exhibit spatial autocorrelation, cross-validation results were interpreted cautiously. The adopted procedure evaluates statistical reproducibility within the available dataset but does not fully eliminate the possibility of spatial dependency between training and validation samples. Consequently, the prospectivity maps generated should be regarded as regional permissivity models rather than fully independent predictive forecasts. The resulting outputs include Classification performance metrics (Figure 5) and Continuous prospectivity maps derived from both RF and ANN models (Figure 6). These maps represent relative permissivity rather than deterministic predictions of mineral occurrence.
RkJQdWJsaXNoZXIy MTM0Mzk2