Abstract
Haloxylon ammodendron is an essential species within the Central Asian desert ecosystem, with its aboveground biomass (AGB) serving as a crucial marker of ecosystem health and desertification levels. Precise and effective methods for predicting AGB are vital for understanding the spatial distributions and ecological roles of desert regions. However, the low vegetation cover in these areas poses significant challenges for satellite-based research. In this study, aboveground biomass training and validation datasets were gathered using UAV LiDAR, covering an area of 50 square kilometers. These datasets were integrated with high-resolution, multi-temporal satellite images from Sentinel-1 (S1) and Sentinel-2 (S2). This study applied a spatial cross-validation method to develop a quantile regression forest (QRF) prediction model. This model was used to predict the AGB of Haloxylon ammodendron forest across a study area of 14,000 square kilometers. The findings suggest that, when supported by ground data, multi-source remote sensing technology can estimate the AGB distribution of Haloxylon ammodendron over large areas. Significant uncertainty exists within the model due to the low vegetation cover characteristic of arid regions and the uneven distribution of sampling points. This uncertainty can be reduced by using area of applicability (AOA) and uncertainty maps, which identify the regions where the model's predictions are most accurate and guide further data collection to enhance precision. This study provides improved insight into the spatial distribution and extent of Haloxylon ammodendron AGB in the research area and offers essential geospatial information for ecosystem conservation strategies. The results also contribute to the understanding of how desert vegetation growth and carbon cycling respond to environmental changes, and for forecasting future vegetation dynamics in arid regions.