Comparative analysis of machine learning algorithms and scales for estimating boreal forest above ground biomass in the Qilian Mountains, China

Accurate and up-to-date maps of the aboveground biomass (AGB) of boreal forests have become the key to global climate change research in recent years. Although Geoscience Laser Altimeter System (GLAS) data is a powerful tool for forest AGB estimation, the data coverage dispersion often means they need to be combined with auxiliary data to map wall-to-wall forest AGB distribution. Exploring the proper estimation approach and mapping scale for estimating boreal forest AGB distribution is critical. In this study, four machine learning algorithms (XGBoost, LightGBM, SVR and RF) and four mapping scales (15 × 15, 30 × 30, 60 × 60 and 90 × 90 m) are compared regarding their ability to estimate the AGB of boreal forest over the Qilian Mountains in northwestern China based on field measurements, spaceborne LiDAR and optical and environmental data. The results showed that the two GBDT AGB models (XGBoost and LightGBM) had better performance in AGB estimation and the LightGBM AGB model yielded superior results (R 2  = 0.41, RMSE = 19.40Mg/ha). The results also indicated that 90 × 90 m was the best resolution for the AGB distribution map in this region. Moreover, an appropriate mapping scale can further improve the accuracy of AGB estimation.