Remote sensing estimation of aboveground biomass of different forest types in Xinjiang based on machine learning

Forest aboveground biomass (AGB) is a key indicator reflecting the function and quality of forest ecosystems, and accurate large-scale estimations of forest AGB are essential for effective forest ecosystem management. However, owing to limitations in forest AGB mapping methods and the availability of ground-based survey data, Xinjiang still lacks provincial-level forest AGB distribution maps. In this study, we focused on four major forest types in Xinjiang: Evergreen Needleleaf Forest (ENF), Deciduous Needleleaf Forest (DNF), Deciduous Broadleaf Forest (DBF), and Mixed Forest (MF). Using topographic and meteorological data and satellite imagery from Landsat and MODIS as the main data source, we applied the Boruta algorithm for feature variable screening. We then combined these features with Xinjiang forest inventory data to construct support vector machine (SVM), extreme gradient boosting (XGBoost), and random forest (RF) models for estimating the biomass of different forest types for AGB mapping of forests at the provincial scale in Xinjiang. Based on this, the regional distribution patterns of forest biomass were analysed. The findings indicated that climate, topography, and texture factors significantly influenced the selection of characteristic variables in the development of large-scale biomass inversion models. The RF model, combined with different forest types, significantly improved the estimation accuracy of forest AGB. Among the three machine learning estimation models, RF demonstrated the highest estimation accuracy, with R² values greater than 0.65, root mean square error (RMSE) between 24.42 and 41.75 Mg/hm2 and mean absolute error (MAE) between 30.59 and 60.46 Mg/hm2 for the four forest types. These results were more accurate than the estimated results for all forest sample plots. (3) The geographical distribution map of forest biomass, calculated using the optimum model, revealed that regions with high AGB values were mostly located in the three major mountain rims, followed by the Ili Valley and areas near the Tarim River Basin, with significant spatial heterogeneity. The findings of this study enhance the accuracy of forest AGB estimation in Xinjiang, providing a basis for variable selection in forest biomass estimation models and offering theoretical references and technological assistance for the remote sensing estimation of forest AGB across large areas.