Random Forest classifier for lithological mapping of the Mundiyawas-Khera mineralized belt of the Alwar basin, India, from remote sensing and potential field data

Lithological interpretation of remote sensing and geophysical data plays a vital role in mineral resource mapping, especially in areas of the limited outcrop. This study applied a Random Forest (RF) classifier to obtain the refined lithological map of the Mundiyawas-Khera mineralized belt of the Alwar basin, India, from remote sensing and potential field data. A total of 540 samples covering the major lithologies were fed to RF for training (80%) and testing (20%), and its performance was evaluated using precision, recall, and accuracy. The degree of uncertainty associated with RF was also computed using the information entropy technique to pinpoint the regions where the refined lithology map is incorrectly classified. The results indicate that RF yields an overall accuracy of 73.15% in classifying all the major lithological units in the region, such as felsic volcanic, carbon phyllite, mica schist, quartzite, and tremolite-bearing dolomite. Among all the five lithologies, RF showed the best precision (84.62%) and recall (90.91%.) for quartzite and M-mica schist respectively and comparable precision/recall values for the felsic volcanic rocks that host Cu mineralization. Whereas other lithologies, dolomite and carbon phyllite, were not accurately predicted by RF, which might be due to the limited number of samples. The results of the class membership probabilities indicate that not all the litho-units predicted by the model are absolute. The study illustrates that RF can be used as a viable alternative in regions with limited outcrops and geochemical information to prepare the new lithology map or refine the existing geological maps. 

Keywords: Machine Learning, Lithology Classification, Gravity and Magnetic Data