Joint Modeling of Spaceborne Radar and LiDAR Data With Ensemble Learning For Forest Aboveground Biomass Estimation

Accurate estimation of forest above-ground biomass (AGB) is fundamental to natural resource management and ecosystem monitoring, which can improve understanding of energy exchange and carbon cycling. Synthetic Aperture Radar (SAR), which is weather-independent and can operate for all-weather, all-day ground observation, is uniquely suited for forest AGB estimation at high elevations in harsh environments. In the study, Sentinel-1 and ALOS-2 data were used as independent data sources for forest AGB modeling, and ICESat-2 was considered as an additional source of data used in conjunction to further improve forest AGB estimation. The XGBoost model in ensemble learning with multiple parametric and nonparametric models was built to estimate natural forest AGB in the eastern Tibetan Plateau. The results showed that the forest AGB models constructed by ALOS-2 were overall better than Sentinel-1, and the model errors were further greatly reduced after joining ICESat-2. The XGBoost achieved the best prediction accuracy among all models with a coefficient of determination of 0.75 (RMSE=18.57 Mg/ha).