Improving grassland classification accuracy using optimal spectral-phenological-topographic features in combination with machine learning algorithm

Abstract

Accurate mapping of large-scale grassland types is important for grassland and water resources management. The similarity of spectral characteristics between grassland types lowers the classification accuracy of different grassland types. To improve the classification accuracy of large-scale grasslands, this study proposed a new framework which integrates Sentinel-2 images with DEM and climate zones data. In this framework, optimal spectral-phenological-topographic features are fed into Random Forest (RF) model based on Google Earth Engine (GEE) platform. The proposed framework was applied in Inner Mongolia, China. A grassland map of the region was obtained with an overall accuracy (OA) exceeding 80 %, which is higher than the OA (60 %-70 %) of current large-scale grassland type classification studies. In WIM (Western Inner Mongolia) and NEIM (Northeast Inner Mongolia), the OA reaches 96.97 % and 95.85 %, respectively. SWIR2 band and elevation have a clear advantage in distinguishing different grassland types. Compared to 1980s, the area of temperate meadow steppe (TMS) and temperate desert steppe (TDS) have increased by 111.94 % and 126.00 %, respectively. The area of temperate typical steppe (TTS) decreased by 7.38 %. The framework proposed in this study has a great potential to be applied to improve large-scale grassland classification in other regions in future.