关键词: 荒漠梭梭林, 地上生物量, 无人机多光谱, 无人机激光雷达, 机器学习

Abstract:

Objective: To address the challenges of remote sensing estimation caused by the sparse, short, and structurally complex characteristics of Haloxylon ammodendron in desert regions, this study explores a high accuracy method for aboveground biomass (AGB) estimation in arid desert shrublands, providing technical support for carbon stock assessment. Method: The H. ammodendron stands along the southern margin of the Gurbantunggut Desert in Xinjiang were used as the research object. Individual shrub segmentation was performed using UAV-LiDAR point cloud data, and AGB of the coverage area was estimated for expanded plots based on allometric equations. On this basis, spectral, textural, and structural features were extracted separately from UAV-MSI and UAV-LiDAR data, and random forest (RF) importance ranking was used for feature selection. Three machine learning algorithms RF, support vector machine (SVM), and extreme gradient boosting (XGBoost) were applied to develop regional-scale AGB models. Model performance was evaluated using leave-one-out cross-validation, and modeling results based on MSI features alone, LiDAR features alone, and their combination were compared. The optimal model was then used to map the spatial distribution of AGB in the H. ammodendron sites. Result: 1) Feature selection revealed that vegetation indices such as normalized difference vegetation index (NDVI), ratio vegetation index (RVI), and maximum point cloud height (Hmax) contributed significantly to AGB estimation. The combined MSI and LiDAR features exhibited a more balanced importance distribution, demonstrating strong complementarity. 2) Among all modeling methods, the AGB models based solely on UAV-MSI features outperformed those based solely on LiDAR features. RF achieved an R² of 0.82 and RMSE of 0.66 t?hm^–2, SVM achieved an R² of 0.79 and RMSE of 0.75 t?hm^–2, while XGBoost performed best with an R² of 0.84 and RMSE of 0.63 t?hm^–2, indicating that spectral features had greater predictive power. 3) The fusion of UAV-MSI and UAV-LiDAR features further improved model accuracy. The XGBoost model combining both feature sets achieved the highest accuracy, with an R² of 0.89 and RMSE of 0.53 t?hm^–2, confirming the complementary value of spectral and structural information. 4) Among the four sampling sites, Site 1 exhibited the highest average AGB at 2.50 t?hm^–2. Sites 2, 3, and 4 showed progressively lower mean AGB values (0.90, 0.84, and 0.64 t?hm^–2, respectively), with over 70% of the area having AGB values below 1 t?hm^–2. AGB spatial distribution varied significantly across sites, showing a decreasing trend with increasing distance from the oasis. Conclusion: This study has established a site-level AGB estimation workflow tailored to desert shrubs in arid regions and demonstrated the synergistic potential of combining UAV-MSI and UAV-LiDAR data in desert shrub AGB estimation. Compared to conventional field-based methods, the proposed approach offers advantages such as non-destructiveness, high resolution, and low cost, making it suitable for biomass estimation of desert shrublands in arid ecosystems.

Key words: desert Haloxylon ammodendron shrubland, aboveground biomass, UAV multispectral, UAV-LiDAR, machine learning