Detecting the distribution of grass aboveground biomass on a rangeland using Sentinel-2 MSI vegetation indices

Abstract

Quantifying the grass biomass content of rangelands is potentially useful to rangeland managers. This study assessed the potential of image-derived vegetation indices to assess in-situ spatial variations in grass aboveground biomass (AGB). Sampling was conducted at the peak growth stage, at widely distributed sampling sites whose grass cover homogeneity was wider than 20 m. Up to three AGB samples per site were collected, in 1 m quadrats. Dry biomass weights were determined and averaged per site. Rainy season Sentinel-2 MSI images of the rangeland that corresponded with the fieldwork dates were obtained. Atmospheric correction of the images was performed using the Sen2Cor algorithm, which yielded 20 m resolution visible (vis), red edge (RE), near infrared (NIR), and short wave infrared (SWIR) bands. Seven biomass-sensitive vegetation indices that utilised Sentinel-2 MSI vis-RE-NIR-SWIR spectral ranges were then tested using the images, to determine which correlated best with the field-derived AGB. They were the Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), Soil Adjusted Vegetation Index (SAVI), Normalized Difference Red Edge index (NDRE), Red Edge Inflection Point (REIP), Aerosol Free Vegetation Index (AFRI 2.1μm ), and Normalized Difference Water Index (NDWI). In decreasing order of magnitude, the EVI, SAVI, NDVI, AFRI 2.1μm , and NDWI had statistically significant correlations with AGB ( p ≤ 0.05). The discontinuous canopies but high biomass of tall-grass species weakened the correlations. The EVI model of AGB was then used for depicting rangeland-scale, location-context spatial variations in AGB. Pixels with woody vegetation cover ≥ 50% were excluded using a binary mask that was developed through sub-pixel classification of a grass-senescence period Sentinel-2 MSI image. A statistically significant linear relationship ( F = 21.192, p = 0.000) between EVI-predicted and actual AGB was established ( r = 0.765, p < 0.001). The results from the studied savannah rangeland suggest higher vegetation index prediction accuracy at short-grass than tall-grass sites.