A machine learning scheme for estimating fine-resolution grassland aboveground biomass over China with Sentinel-1/2 satellite images

Estimates of grassland aboveground biomass (AGB) upscaled from the plot level to broad scales are lacking extrapolation robustness across grassland biomes. This is because the coupling relationships between diverse driving variables and model forms are poorly understood and/or are not robustly defined. Here, using 17,421 in situ ground-truth AGB samples and 31 driving variables generated from Sentinel-1/2 satellite images and climatic-topographical-soil data, we explored the coupling relationships among these multifaceted driving variables and 6 machine learning (ML) algorithms in AGB upscaling estimates across 18 grassland types in China. Monte Carlo simulations indicated that these 31 multifaceted driving variables did not enhance the performance of the ML algorithms in AGB modeling and that prioritizing climatic factors at a 10 km resolution and the satellite enhanced vegetation index (EVI) at a 10 m resolution have scale effects; the former dominates the variation in grassland AGB among 18 grassland types, and the latter reflects local heterogeneity. ML algorithms prevalently suffer from saturation convergence, and the underrepresentation of training samples leads to instability in ML establishment when constrained to a specific grassland type. Among the 6 ML algorithms, the random forest (RF) algorithm outperformed the other methods in terms of the mean absolute error (MAE) during iterative training; however, the best performance strongly relies on saturated training, with an R2 of 0.68 over all grasslands in China. Finally, based on the optimal estimate of grassland AGB using RF (0.25 Pg C) and both belowground biomass (1.52 Pg C) inferred from the ratio of roots to shoots (R/S) of the 18 grassland types and the approximate intake of grazing animals (0.02 Pg C), the plant-based carbon stock over Chinese grasslands is estimated to be 1.79 Pg C. This study emphasizes that to untangle the limitations of using ML approaches, the robustness of model training and the complementary effect between climatic factors and satellite metrics should be considered when upscaling estimates of grassland AGB.