A Deep Learning Approach with Uncertainty Estimation to Assess Aboveground Biomass Mapping of Tropical Rainforest in Thailand

This study proposes an aboveground biomass (AGB) density model to overcome the limitation of sparse data coverage in Thailand and scale at national level. The model is based on a U-Net convolutional neural network and fuses spaceborne remote sensing data including Sentinel-1, Sentinel-2, and GEDI. Five models were developed using a deep ensemble technique to provide uncertainty estimation, where accuracy was validated using three types of forest inventory data. Tropical rainforests and other forested areas such as perennial crop regions with similar geographical characteristics were evaluated. Results show high uncertainty in the out-of-distribution data areas, with %RRMSE for Pará rubber, mixed deciduous forest, and dry evergreen forest of 17.05, 24.47, and 14.0, respectively.