Remote sensing of grassland communities in Mongolian Steppe combining multi-source data and machine learning classification algorithms

In this study, we investigate the performance of machine learning classification approaches and different remotely sensed data sources for identifying and mapping three types of grassland communities found in the Mongolian Steppe region (Artemisia, Caragana and grass dominated steppes). The Mongolian steppe is intensively used as pasture and provides the economic basis for approximately 1 million herders. The grassland types differ in their forage values, which is why a spatially-explicit estimation of their occurrence is of high importance. We compared different sensors, classifiers, and training-sample strategies to identify the most effective approaches for mapping these communities. Ten datasets were used: Landsat 8 OLI (30 m), pan-sharpened Landsat 8 (15 m), Landsat 8 Surface Reflectance (30 m), Sentinel 2 (10 m), Sentinel 2 (20 m), Worldview 3 (0.5 m and 1.2 m), integrated Landsat 8 and Sentinel 2 (30 m), temporal Landsat 8, and temporal Sentinel 2. The two foremost classifiers at producing high accuracy of land cover classification, SVM and RF, were applied with the same training datasets. The training samples were collected in a manner so that they could be used for different spatial resolutions (i.e., ranging from 0.5 to 30 m) with the least effect from mixed training samples and spatial autocorrelation. The results of this study indicate that remote sensing is a viable method for the identification of different grassland communities in the Mongolian Steppe region.