Abstract: Estimating forest aboveground biomass (AGB) is critical to achieving global carbon neutral goals. In this study, Howland forest in Maine, USA is taken as the research area. With the ground measured sample site data, different data sources (airborne LiDAR and hyperspectral remote sensing data) and machine learning algorithms (random forest, support vector machine, gradient boosting decision tree and K-nearest neighbor) are compared and analyzed. It is to improve the estimation accuracy of Howland forest AGB. The results show that the optimal accuracy of the model with airborne LiDAR and hyperspectral vegetation index variables is 0.874 and 0.868 respectively. The accuracy of the regression model with the combination of airborne LiDAR and hyperspectral vegetation index variables and gradient boosting decision tree is 0.927, that is, multi-source remote sensing data is better than a single data source. The synergistic use of LiDAR and hyperspectral data has applicability and application prospects for improving the accuracy of biomass estimates in areas such as Howland and beyond.

Key words: LiDAR, hyperspectral remote sensing, forest aboveground biomass, machine learning, GBDT