Comparison of hyperspectral HJ-1A/HSI and multispectral Landsat 8 and Sentinel-2A imagery for estimating alpine grassland coverage in the Three-River Headwaters region

Abstract

Accurate estimation of grassland coverage is important in both monitoring grassland growth and promoting grassland management. This study, for the first time, aims to evaluate the alpine grassland coverage estimation performances of four widely used methods [i.e., random forest classification (RFC), regression analysis (RA), multiple endmember spectral mixture analysis (MESMA), and support vector machine regression (SVMR)] from three typical remote sensing images [i.e., hyperspectral (HJ-1A/HSI) image and multispectral (Landsat 8 and Sentinel-2A) images] in the Three-River Headwaters region, China. The total grassland coverage and three levels of grassland subcoverage (i.e., low, moderate, and high coverage) are estimated from the three images by each method. Meanwhile, the overall accuracy (OA) and root-mean-square error (RMSE) of each coverage result are assessed. The experimental results show that (1) for the total grassland coverage estimation, RFC method combined with Sentinel-2A data and RA method used both Landsat 8 and Sentinel-2A data generated the highest OA of 79.4%, whereas the highest OAs of SVMR and MESMA methods are 3% to 14.7% lower than those of RFC and RA methods, (2) for three grassland subcoverage estimation, Landsat 8 data combined with RFC method generated the greatest OAs (100%) and the lowest RMSEs (7.34%) for low grassland coverage and Sentinel-2A data combined with SVMR method obtained the highest OAs (≥83.3 % ) for moderate and high grassland coverage, (3) Sentinel-2A and Landsat 8 data generated higher OAs than HJ-1A/HSI data in estimating the total coverage while Sentinel-2A produced better performance than Landsat 8 in estimating the grassland subcoverage, and (4) SVMR method performed more stably than other methods in estimating alpine grassland coverage.