Integrating Multi-Source Geoscientific Data via Geologically Constrained Feature Engineering for Gold Prospectivity Mapping: A Case Study of Jiaoxibei, China

The Jiaoxibei gold cluster is one of the most significant gold-producing regions in China and retains substantial regional prospecting potential. However, the superposition of multiple mineralization events has resulted in strong spatial coupling, multi-scale variability, and substantial redundancy among structural, alteration, geophysical, and geochemical information, limiting the effective extraction of key ore-controlling features. This study developed a geologically constrained feature-engineering framework for regional mineral potential evaluation. An initial indicator system comprising 32 geologically meaningful factors was constructed from structural geometry, remote-sensing alteration, gravity, magnetic, and geochemical information. The previously developed SOMML method was extended by introducing borehole-derived geological constraints to construct G-SOMML and generate the comprehensive geochemical anomaly factor Chem_F. Correlation-based redundancy reduction, Random Forest importance evaluation, SHAP interpretation, and geological screening were then combined to identify nine core factors. The selected factors were subsequently transformed according to their mineralization-response directions and integrated through category-balanced fuzzy synthetic evaluation. By linking geology-guided feature construction, borehole-constrained geochemical anomaly extraction, data-driven feature diagnosis, and balanced evidence integration, the framework provides a reproducible and interpretable workflow for organizing heterogeneous geoscientific information in regional mineral potential evaluation. The resulting high-potential zones captured 21 of the 22 known gold occurrences in the Sanshandao and Jiaojia areas and 17 of the 23 occurrences in the other parts of the study area, yielding an overall deposit capture rate of 84.4%. Compared with the all-feature fuzzy synthetic evaluation model (AT-Fuzzy), feature engineering reduced the high-potential area ratio from 50.951% to 25.356% while increasing the deposit capture rate from 51.1% to 84.4%. At the map level, the proposed framework delineated a substantially smaller high-potential area than the Random Forest model while achieving higher deposit capture rates than both the Random Forest and Weights of Evidence models under the common evaluation domain and statistical thresholding criterion. These results support the applicability of the framework for interpretable regional mineral potential evaluation and target prioritization in complex metallogenic districts.