Geochemical Anomaly Detection Cooperatively Driven by Multiple Models Based on Stacking Framework

In response to the systematic uncertainty of geochemical anomaly detection related to Fe-mineralization in the Hunjiang area in Jilin Province, China, this paper synthesizes nine geochemical elements (Fe2O3, Al2O3, etc.) and a geological constraint to construct a characteristic dataset and then proposes a method based on the Stacking framework for using Gaussian Naive Bayes to integrate three anomaly detectors of Support Vector Classification, Random Forest, and K-Means-SMOTE-Boost model. The results indicate that: (1) The single model has limitations and high uncertainty: the high anomaly areas detected by Support Vector Classification, Random Forest, and K-Means-SMOTE-Boost model only captured 33.3, 33.3, and 28.6% of known iron deposits. This shows that the accuracy of anomaly detection is not high, presenting a high degree of uncertainty. (2) Multi models averaging effectively reduces uncertainty: by integrating Support Vector Classification, Random Forest, and K-Means-SMOTE-Boost with Gaussian Naive Bayes, the spatial correlation between the high anomaly areas and known Fe-mineralization deposits is significantly enhanced, capturing 85.7% of known iron deposits. This shows that the accuracy has been improved and uncertainty has been reduced. (3) Stacking framework demonstrated clear advantages: compared with a single model, the model designed based on the Stacking framework converges faster and is close to the level of an ideal anomaly detector, effectively improving the accuracy and generalization ability. In summary, the method based on the Stacking framework for using Gaussian Naive Bayes to integrate and average Support Vector Classification, Random Forest, and K-Means-SMOTE-Boost model provides new ideas for reducing the uncertainty of geochemical anomaly detection, and lays a better support foundation for geochemical exploration work and mineral resource prospecting.