Automatic lithological mapping from potential field data using machine learning: a case study from Mundiyawas-Khera Cu deposit, Rajasthan, India

Potential field data play a vital role in mineral resource mapping, especially in deriving the lithological information of poorly mapped terrains. The Mundiyawas-Khera area of the Alwar basin in Rajasthan, India, is known for Cu mineralization hosted within the felsic volcanic rocks. However, much of the area is covered with soil and needs detailed lithological mapping. In this study, different machine learning (ML) algorithms have been employed to integrate the digital elevation, drilling wells, gravity, and magnetic data, together with their derivatives, for obtaining accurate lithology information of the area. Initially, five different ML algorithms, random forest (RF), K -nearest neighbor, support vector machine, multi-layer perceptron (MLP), and gradient boosting (GB) were employed using 540 samples from six lithological units to obtain the refined lithologic map of the area. Subsequently, a stacking classifier was built, considering the best-performing ML models in the base learner. Comparison of evaluation matrices (precision, recall, and confusion matrix) of these ML algorithms suggests that RF, GB, and stack model (RF + GB + MLP with RF meta-classifier) provide the highest accuracy score (RF: 74.69%, GB: 74.69%, and stack: 75.31%) and class membership probabilities in predicting the lithology. Adding derivatives and analytic signal information to the input data improves the classification accuracy of ML models by ~ 5–8%. Overall the study results demonstrate that ensemble ML algorithms can aid in creating the first-pass lithology map of areas with limited outcrops, drilling, and geochemical data.