Improving above ground biomass estimates of Southern Africa dryland forests by combining Sentinel-1 SAR and Sentinel-2 multispectral imagery

Abstract

Having the ability to make accurate assessments of above ground biomass (AGB) at high spatial resolution is invaluable for the management of dryland forest resources in areas at risk from deforestation, forest degradation pressure and climate change impacts. This study reports on the use of satellite-based synthetic-aperture radar (SAR) and multispectral imagery for estimating AGB by correlating satellite observations with ground truth data collected on forest plots from dryland forests in the Chobe National Park, Botswana. We undertook nineteen quantitative experiments with Sentinel-1 (S1), and Sentinel-2 (S2) and tested simple and multivariate regression including parametric (linear) and non-parametric (random forests) algorithms, to explore the optimal approaches for AGB estimation. The largest AGB value of 145 Mg/ha was found in northern Chobe while a large part of the study area (85%) is characterized by low AGB values (< 80 Mg/ha), with an average estimated at 51 Mg/ha. The results show that the AGB estimated using SAR backscatter values from vertical transmit receive (VV) polarization is more accurate than that based on horizontal receive (VH) polarization, accounting for 58% of the variance compared to 32%. Nevertheless, the combination of S1 SAR and S2 multispectral image data produced the best fit to the ground observations for dryland forests explaining 83% of the variance with an accuracy of 89%. Furthermore, the optimal AGB model performance was achieved with a random forest (RF) regression trees algorithm using S1 (SAR) and S2 (multispectral) image data (R2 = 0.95; RMSE = 0.25 Mg/ha). From the 11 vegetation indices tested, GNDVI, Normalized Difference Red Edge (NDRE1), and NDVI obtained the highest linear relationship with AGB (R2 = 0.71 and R2 = 0.56, p < 0.001), however, GNDVI and NDRE1 improved the AGB estimation at medium to high-density forests compared to NDVI. The GRVI and EVI were the least correlated with AGB (R2 = 0.09 and R2 = 0.31) at a significance level of p < 0.001, respectively. We show that NDVI saturates in areas with >80 Mg/ha AGB, whereas the inclusion of SAR backscatter and optical red edge bands (B5) significantly reduces saturation effects in areas of high biomass. GNDVI and red edge (B5) derived vegetation indices have more potential for estimating AGB in dryland forests than NDVI. Our results demonstrate that dryland AGB can be estimated with a reasonable level of precision from open access Earth observation data using multivariate random forest regression.