Estimating the grassland aboveground biomass in the Three-River Headwater Region of China using machine learning and Bayesian model averaging

Spatially and temporally explicit information on the biomass in terrestrial ecosystems is essential to better understand the carbon cycle and achieve vegetation resource conservation. As a climate-sensitive critical ecological function area, accurate monitoring of the spatiotemporal variation in the grassland aboveground biomass (AGB) is important in the Three-River Headwater Region (TRHR) of China. In this study, based on field observation, remote sensing, meteorological and topographical data, we estimated the grassland AGB in the TRHR and analyzed its spatiotemporal change and response to climatic factors. Four machine learning (ML) models (random forest (RF), cubist, artificial neural network and support vector machine models) were constructed and compared for AGB simulation purposes. The AGB results estimated with the four ML models were then applied in integrated analysis via Bayesian model averaging (BMA) to obtain more accurate and stable estimates. Our results demonstrated that the RF model performed better among the four ML models (testing dataset: correlation coefficient (r) = 0.84; root mean squared error = 76.99 g m(-2)), and BMA improved grassland AGB prediction based on the multimodel results. The spatial distribution of the grassland AGB in the TRHR was heterogeneous, with higher values in the southeast and lower values in the northwest. The interannual variation in the grassland AGB in most areas of the TRHR exhibited nonsignificant increasing trends from 2000 to 2018, and the sensitivity of the AGB to the annual precipitation was obviously modulated by regional water conditions. This study provides a more precise method for grassland AGB estimation, and these findings are expected to enable improved assessments to obtain a greater grassland AGB understanding.