Comparison of Gaussian process regression and least squares linear regression to estimate above-ground biomass using Sentinel-2 data (Case study: Kheyrud Forest)

The aim of this research was to investigate the performance of the non-parametric Gaussian process regression (GPR) and the parametric linear least squares regression (LMSR) in estimating above-ground biomass (AGB) in a heterogeneous mountain forest using Sentinel-2 data. To model and validate the above-ground biomass, 102 square-shaped sample plots with dimensions of 45 × 45 meters were collected using a selective method in pure Fagus orientalis and Carpinus betulus L. stands in the Kheyrud forest. Tree volume, and subsequently AGB were estimated using a local volume table and average wood density for each species. Atmospheric correction was applied to the Sentinel-2 image. The main spectral bands, vegetation indices, Tasseled Cap transformation, and principal component analysis veriables were used to model AGB. Seventy percent of the field sample plots were used for modeling with three datasets (main spectral bands, vegetation indices, and the combination of main bands and vegetation indices). To validate the models thirty percent of field sample plots were used. Based on the coefficient of determination and relative root mean squared error (RRMSE), the GPR achieved the best result, with R² = 0.56 and RRMSE = 21.14%. The results of above-ground biomass modeling using the main bands for LMSR produced an R² = 0.43 and RRMSE = 23.32%. A combination of vegetation indices (VIs) and main spectral bands did not improve the model accuracy for both GPR and LMSR. Overall, our results indicated that combining GPR with Sentinel-2 data reasonably improved forest AGB estimation.