Mineral prospectivity mapping susceptibility evaluation based on ensemble learning: A case study of Fe-Au polymetallic skarn-type deposits in the Miaoshan-xintai area, western Shandong

Abstract: [Objective] Aiming at the limitation of the traditional metallogenic prediction model in the lack of multi-source data fusion ability under complex geological conditions, a metallogenic prediction method based on ensemble learning is proposed. [Methods] A two-layer Stacking integration strategy was adopted. Three algorithms, Random Forest, XGboost and Catboost, were integrated in the base learning layer. The meta learning layer used logistic regression algorithm to integrate the output of the base learning. At the same time, a screening mechanism based on the importance score is constructed to quantitatively analyze the influence of variables on the nonlinear model, so as to provide a basis for the optimization of geological variables. Taking the iron gold polymetallic skarn deposit in Miaoshan-Xintai area of Western Shandong Province as an example, 17 variables were selected as evaluation factors for prediction based on multi-source geological, geophysical and geochemical data. [Results] The results indicate that the integrated model outperforms the single model significantly across four metrics: accuracy, precision, F1 score, and AUC value. Furthermore, the metallogenic probability predictions made by the integrated model align well with the spatial distribution of known deposits. The feature recognition capabilities of the three types of base learners are complementary, and the integrated learning mechanism enhances the multi-dimensional representation of geological features. [Conclusion] Combined with the prediction results and the analysis of metallogenic geological background, three prospecting target areas are delineated in the Miaoshan-Xintai region of western Shandong, guiding the direction for subsequent exploration efforts. The design of the feature selection scorer holds significant value for methodological promotion. The method proposed in this study, which involves multi-source data fusion and collaborative optimization of heterogeneous models, significantly enhances prediction reliability and offers new technical support for the new round of strategic action to achieve a breakthrough in prospecting.