基于集成学习的成矿预测——以鲁西苗山−新泰地区铁金多金属矽卡岩型矿床为例OA

[Objective]Traditional mineral prospectivity models show limited capacity for fusing multi-source geoscience datasets under complex geological settings.Current ensemble-learning-based mineral prospectivity studies still have two major drawbacks:Insufficient analysis of complementary effects among base learners in identifying geological ore-controlling variables,and inadequate integration of prior geological constraints within data-driven modelling workflows.To address these gaps,this study constructs a two-layer Stacking ensemble learning framework for skarn-type iron-gold polymetallic mineral prospectivity prediction.[Methods]The Stacking architecture consisted of a base-learner layer and a meta-learner layer.Random forest(RF),extreme gradient boosting(XGBoost),and categorical boosting(CatBoost)were selected as base learners,while logistic regression was adopted as the meta-learner to synthesize prediction outputs from individual base models.A perturbation-based feature importance scoring evaluator was developed to quantitatively interpret geological predictive variables under nonlinear model conditions.Taking the skarn-type iron-gold polymetallic deposits in the Miaoshan-Xintai area of western Shandong as the research case,this study integrated multi-source geological,geophysical,and geochemical datasets.Redundant variables were eliminated through combined Pearson linear-correlation and Spearman rank-correlation analyses,and 17 core predictive factors were finally retained.The synthetic minority oversampling technique(SMOTE)was applied to mitigate severe sample imbalance between mineralized positive samples and barren negative samples.A five-fold stratified cross-validation strategy was implemented for model training and comprehensive performance assessment.[Results]The Stacking ensemble model outperformed three standalone base-learner models(RF,XGBoost,and CatBoost)in multiple quantitative indicators including accuracy,precision,F1-score,and AUC value.The spatial distribution of predicted mineral prospectivity probabilities showed high consistency with the locations of known skarn-type deposits in the study area.Three base learners exhibited obvious complementary capabilities in identifying ore-controlling geological factors.Model interpretation results further revealed that fault structures and Late Yanshanian intrusive rock bodies represented the dominant ore-controlling factors for skarn-type iron-gold polymetallic mineralization in this study area.[Conclusion]By integrating model-derived mineral prospectivity maps with regional metallogenic geological constraints,three prospecting target areas with exploration potential are delineated for further field investigation.The proposed perturbation-driven feature-importance evaluation tool provides a useful methodological reference for mineral prospectivity modelling.The technical workflow combining multi-source geodata fusion and heterogeneous-model collaborative optimization substantially improves the reliability of mineral prospectivity prediction and can provide technical support for the new-round national mineral exploration breakthrough strategy.Nevertheless,further improvements are still required in intelligent hyper-parameter optimization and multi-scale sample balancing strategies,thereby enhancing the generalizability of the mineral prospectivity approach based on ensemble learning.