关键词: 柠条锦鸡儿, 无人机, 生物量, 极端梯度提升决策树, 激光雷达数据

Abstract:

Objective: With unmanned aerial vehicle data, an object-oriented method was used to identify individual Caragana korshinskii in Ordos City. RF, SVR, and XGBoost machine learning algorithms were compared to achieve high-precision extraction of individual C. korshinskii and accurate estimation of the biomass, providing a reference for environmental protection and carbon storage research in arid areas. Method: By comprehensively utilizing UAV-borne multispectral and lidar data, and integrating spectral and vertical structure information, an object-oriented method was used to conduct high-precision extraction of individual C. korshinskii. On this basis, three machine learning algorithms of random forest (RF), support vector regression (SVR) and extreme gradient boosting decision tree (XGBoost) were compared to conduct remote sensing accurate estimation of biomass. Result: 1) The ultra-high-resolution image data was obtained by UAV, and the LSMS segmentation algorithm and SVM classifier were able to achieve high-precision identification of individual C. korshinskii. The segmentation accuracy of C. korshinskii in each sample plot was above 86%, the accuracy of the total sample plot was above 90%, the under-segmentation and over-segmentation errors were below 6%, and the overall classification accuracy reached 91.51%. 2) The Recursive Feature Elimination (SVM-RFE) method based on support vector machines identified 17 variables with high contributions to biomass modeling, including 2 planar features and 15 height variables. The cumulative contribution of height variables to biomass was significantly more than that of planar variables (8.7 vs. 1.39). 3) Compared to the RF and SVR models, the XGBoost model provided higher biomass estimation accuracy for C. korshinskii in the study area (R2 = 0.95, RMSE = 259.57 g, MAE = 157.51 g), especially when biomass was below 2 000 g. 4) The multiple vegetation vertical structure information extracted from UAV-LiDAR reflected the diversity and vertical complexity of internal vegetation growth, which was beneficial for improving biomass estimation accuracy. Additionally, integrating multidimensional height variables, such as mean absolute deviation, coefficient of variation, variance, and percentile height, for biomass prediction showed advantages over using a single maximum height variable. Conclusion: The combination of LSMS segmentation and SVM classification for individual shrub extraction offers a technical reference for identifying individual vegetation. The introduction of multi-dimensional point cloud height metrics for biomass estimation compensates for the lack of vertical structure information in C. korshinskii provided by single multispectral data, improving the accuracy of biomass estimation. The XGBoost model provides a new perspective and tool for small-scale shrub biomass estimation in arid regions. Additionally, the high-resolution imagery and point cloud data obtained from UAVs avoid damage to the local ecological environment, which is particularly important in the fragile sandy areas.

Key words: Caragana korshinskii, unmanned aerial vehicle, biomass, XGBoost, LiDAR