Advanced machine learning based gold prospectivity mapping in the Dharwar Craton, India: A hybrid knowledge-data driven paradigm integrating ensemble and deep learning

Abstract

Developing more sophisticated as well as efficient exploration methods to identify the hidden ore bodies are necessary to meet the world wide increasing demand of mineral resources. In this regard, mineral prospectivity mapping (MPM) is crucial. This study undertakes a pioneering effort to apply and compare four machine learning (ML) models—random forest (RF), XGBoost (XGB), support vector classifier (SVC) and artificial neural network (ANN)—for gold prospectivity mapping within the Archean Dharwar Craton, India. The primary goals included the development and evaluation of these models, a systematic assessment of their comparative performance through cross-validation, feature important analysis and ultimately, production of prospectivity map. The culmination of this work is a high-resolution, combined prospectivity map, designed to produce a new prospectivity areas. Diverse geospatial data was meticulously integrated as per mineral system of the area, including geological maps, structural lineaments, geochemical, geophysical and ASTER remote sensing imagery. For model training, 79 known gold occurrences were carefully collected alongside an equal number of selected non-occurrence locations, framing the task as a supervised binary classification problem. Rigorous evaluation, employing 5-fold cross-validation and a 70:30 train-test split, confirmed the exceptional capabilities of these models. XGB and RF consistently emerged as top performers, with impressive AUC-ROC values of 0.9992 and 0.9965, respectively, coupled with high precision, recall and F1-scores and few false positives or negatives. While ANN also showed excellent performance, SVC, though strong, yielded comparatively lower metrics. A detailed feature importance analysis exhibits the positive role of Meta-Basalt, geochemical principal component 1 and Bouguer gravity anomaly and its derivative maps. Success-rate curves vividly illustrated the models' efficiency capturing over 76% of known occurrences within just 20% of the highest-ranked areas shows targeting precision. The generated combined prospectivity map, a robust synthesis from XGB, RF and ANN, based on a stringent consensus criterion validates existing knowledge and precisely delineates high-priority exploration targets, fundamentally reshaping the approach to future mineral exploration.