Random Forest Classification of Proterozoic and Paleozoic rock types of Tsagaan-uul area, Mongolia

The Tsagaan-uul area of the Khatanbulag ancient massif in the Central Asian Orogenic Belt is located in the southern part of Mongolia, which belongs to the Gobi Desert. It has a low vegetation cover, and because of this, remotely sensed data can be used without difficulty for geological investigations. Factors such as sparse population and underdeveloped infrastructure in the region further create a need for combining traditional geological mapping with remote sensing technologies. In existing geology maps of the area, the formations are lithologically very diverse and their boundaries were mapped variously, so a need for a more precise lithology-based map arouse.

This study investigated combinations of fieldwork, multispectral data, and petrography for the rock type classification. A random forest classification method using multispectral Sentinel-2A data was employed in order to distinguish different rocks within Proterozoic Khulstai (NP₁hl) metamorphic complex, which is dominated by gneiss, andesite, sandstone, limestone, amphibolite, as well as the Silurian terrigenous-carbonate Khukh morit (S₁hm) formation, Tsagaan-uul area. Based on the ground samples collected from field surveys, ten kinds of rock units plus Quaternary sediments were chosen as training areas. In addition, morphometric parameters derived from SRTM data and band ratios used for iron-bearing minerals from Sentinel 2 bands are selected as variables in the accuracy of classification. The result showed that gneisses were recognized with the highest accuracy in the Khulstai complex, and limestones and Quaternary sediments were also well predicted. Moreover, the tectonic pattern was also well recognized from the results and compared to the existing maps provided a more detailed geological image of the area. This study emphasized the need for samples as baseline data to improve the machine learning methods, and the method provides an appropriate basis for fieldwork.