Tungsten prospectivity mapping using multi-source geo-information and deep forest algorithm

• The deep forest (DF) model achieves superior performance when comprehensively considering classification accuracy, generalization ability, and predictive efficiency. • The superior performance of the DF model benefits from its inherent ensemble learning strategy, characteristic cascade architecture and well interpretability. • The well interpretability of the DF model links modeling results and the understanding of the target mineral system, facilitating valid guidance for future exploration. Mineral prospectivity mapping (MPM) using multi-source geo-information and artificial intelligence (AI) algorithms is an effective and increasingly accepted tool for delineating and prioritizing potential targets for further mineral exploration. In this study, the extraction and integration of multi-source data, including geological, geochemical, geophysical, and remote sensing data, are conducted to yield fifteen evidential layers, based on which the deep forest (DF) model, an ensemble learning framework with deep architecture suitable for addressing complex classification tasks, is trained together with benchmarked random forest (RF), support vector machine (SVM), artificial neural network (ANN), convolutional neural networks (CNN), and recurrent neural networks (RNN) models, to map the tungsten prospectivity in southern Jiangxi Province (SJP). The results indicate that the DF model is the optimal predictor based on its comprehensively superior performance on classification precision, generalization ability and predictive efficiency. The DF model achieves the second-best classification performance with high values of accuracy (0.8648), sensitivity (0.8314), specificity (0.8972), and Kappa value (0.7293), and showcases the sub-optimal generalization performance indicated by its highest AUC value of the test set (mean AUC of 0.9460) and the low measured overfitting degree of 0.0511. In addition, the DF model exhibits high predictive efficiency regarding a higher success-rate within a smaller target area, which is a primary concern in practical mineral exploration. The prospectivity map was generated by the DF model combined with consideration of uncertainty measurement. The delineated low-risk and high-potential targets occupy only 3.98% of the study area, yet contain 48.31% of the known deposit sites. The DF model benefits from the characteristic cascade architecture, ensemble learning strategy and well interpretability. Specifically, the introduction of the derived features from the second layer of the cascade structure enhances the capability of the DF model in capturing complex patterns in high-dimensional and multi-source geo-information datasets. The interpretability analyses highlight the significant contributions of geochemical anomalies, proximity to Yanshanian intrusions and density of fault intersections on model output, which can be linked to the ore-forming processes and specific geological setting of the tungsten mineral system in SJP, thus providing interpretable guidance for future exploration in the study area.