An 1km global forest above-ground biomass map generated from multiple satellite products and ancillary information

Abstract: Global forest biomass plays a key role in quantifying carbon stock and assessing climate change, and accurate estimation of forest aboveground biomass (AGB) distribution globally is significant in understanding the carbon cycle and the roles in global climate change. A new global AGB map in 2005 has been generated based on machine learning algorithms. The LiDAR-derived biomass estimates with the limited spatial and temporal coverage and a suite of optical satellite products(LAI, GPP, VCF and land cover type), field measurement data, and auxiliary dataset (precipitation, temperature and topographic data) were used to train the machine learning models for five forest types (Evergreen Needleaf Forest, Evergreen Broadleaf Forest, Deciduous Needleaf Forest, Deciduous Broadleaved Forest and Mixed Forests), respectively. Three machine learning methods, including Artificial Neural Network (ANN), Multivariate Adaptive Regression Splines (MARS) and Gradient Boosting Regression Tree (GBRT), were explored. By comparing these models, the GBRT was found to be the optimal algorithm and a global forest AGB map in 2005 was generated by this algorithm at a spatial resolution of 1 km. The independent validation of the global forest AGB map showed good accuracy with R² value of 0.49 and root-mean-square error (RMSE) value of 72.09 Mg/ha. Moreover, the generated global forest AGB map was also compared with other published forest AGB maps and they are highly consistent. One of the potential ability of our new method is it can easily produce global forest AGB maps annually.