Mapping Shrub Coverage in Xilin Gol Grassland with Multi-Temporal Sentinel-2 Imagery

Abstract

In recent decades, shrubs dominated by the genus Caragana have expanded in a large area in Xilin Gol grassland, Inner Mongolia, China. This study comprehensively evaluated the performances of multiple factors for mapping shrub coverage across the Xilin Gol grassland based on the spectral and temporal signatures of Sentinel-2 imagery, and for the first time produced a large-scale shrub coverage mapping result in this region. Considering the regional differences and gradients in the types and sizes of shrub in the study area, the study area was divided into three subregions based on precipitation data, i.e., west, middle and east regions. The shrub coverage estimation accuracy from dry- and wet-year data, different types of vegetation indices (VIs) and multiple regression methods were compared in each subregion, and the key phenological periods were selected. We also compared the accuracy of four mapping strategies, which were pairwise combinations of zoning (i.e., subregions divided by precipitation) and non-zoning, and full time series of VIs and key phenological period. Results show that the mapping accuracy in a dry year (2017) is higher than that in a wet year (2018). The optimal VIs and key phenological periods show high spatial variability. In terms of mapping strategies, the accuracy of zoning is higher than that of non-zoning. The root mean square error (RMSE), overall accuracy (OA) and recall for ‘zoning + full time series (or + key phenological period)’ strategy were 0.052 (0.055), 76.4% (79.7%) and 91.7% (94.6%), respectively, while for ‘non-zoning + full time series (or + key phenological period)’ strategy were 0.057 (0.060), 75.5% (74.6%) and 91.7% (88.6%), respectively. The mapping using VIs in key phenological periods is better than that of using full time series in the low-value prediction of shrub cover. Based on the strategy of ‘zoning + key phenological period’, the shrub coverage map of the whole region was generated with a RMSE of 0.055, OA of 80% and recall of 95%. This study not only provides the first large-scale mapping data of shrub coverage, but also provides reference for shrub dynamic monitoring in this area.