Optimized AI-MPM: Application of PSO for tuning the hyperparameters of SVM and RF algorithms

Modern computational techniques, particularly Support Vector Machines (SVM) and Random Forest (RF) models, are revolutionizing predictive mineral prospectivity mapping. These advanced systems excel at identifying prime resource locations but require meticulous fine-tuning of their internal settings to achieve peak performance. Careful calibration of these configurations during the learning phase significantly enhances their ability to detect promising deposits. The main goal of this study is to introduce a hybrid model called PSO -SVM and PSO-RF, which aim to combine particle swarm optimization (PSO) with SVM (with RBF kernel) and RF models. This hybrid model automatically adjusts the optimized hyperparameters of SVM and RF, resulting in highly accurate predictions and a wide range of applicability. The PSO algorithm has been applied to fine-tune two main parameters (C and λ) for SVM-RBF and three main parameters (N_T, N_S, and d) for RF, creating efficient models for both. The proposed hybrid model as well as the traditional versions of SVM and RF models, were tested using a geo-spatial dataset related to Cu mineralization in Kerman belt, SE Iran. Forecasting algorithms were developed by integrating diverse datasets: multi-element concentrations from stream samples, bedrock and fault line maps, indicators of hot fluid interaction, aeromagnetic survey results, coordinates of previously identified copper-rich igneous intrusions, and verified ore body positions as reference points. The models' performance was evaluated using four validation methods: Multi-round data partitioning (K-fold), error classification tables (confusion matrix), true-positive vs. false-positive graphical analysis ((ROC) curve), and P-A plot were used to assess algorithms and models performance. Tests revealed that the PSO-SVM surpassed all competitors. Impressively, this fine-tuned classifier identified prime target zones in merely one-seventh of the region (14%), yet these areas encompassed nearly all verified resource sites (97%).