UAV-based individual shrub aboveground biomass estimation calibrated against terrestrial LiDAR in a shrub-encroached grassland

Abstract

Shrub encroachment is an important ecological issue that is increasingly receiving global attention in arid and semiarid grasslands. Monitoring the spatial distribution of encroached shrub aboveground biomass (AGB) is critical for ecological conservation and adaptive ecosystem management. However, the low stature and fine spatial heterogeneity of encroached shrub communities increase difficulties for coarse spatial-resolution satellite images to adequately capture detailed characteristics of individual shrubs. Unmanned aerial vehicle (UAV) can acquire centimeter-level optical images or high-density LiDAR point cloud data, providing an effective means to map encroached shrub AGB spatially explicitly, even at the individual scale. In this study, we first extracted the individual shrubs based on thresholds in normalized difference vegetation index (NDVI) and canopy height model (CHM) using UAV-based multispectral and LiDAR data. For each shrub, we then derived and determined the dominant geometric, spectral, and textural features from the high-resolution multispectral image and the volumetric features from the LiDAR data as predictors of shrub AGB. Finally, we compared the capability of different data sources (UAV-based multispectral image, LiDAR, and their combination) and regression methods (multiple linear, random forest, and support vector regression) to estimate and map the individual shrub AGB in the study area. The volume-based approaches to individual shrub AGB, including global convex hull method, voxel method, and surface differencing method, were also employed using terrestrial laser scanning (TLS) to further calibrate the UAV-based estimation. Our results show that individual shrubs can be accurately extracted based on the threshold method with an overall classification accuracy of 91.8%. The UAV-based AGB estimation suggests that the textural feature, the sum of contrast metric within the individual shrub canopy, is the most important predictor of individual shrub AGB, followed by volumetric, geometric and spectral features. Moreover, the high-resolution multispectral image shows greater potential (R2 = 0.83, RMSE = 106.46 g) than LiDAR (R2 = 0.77, RMSE = 123.33 g) in the estimation of individual shrub AGB, and their combination can only slightly improve the estimation accuracy (R2 = 0.86, RMSE = 101.97 g). Our results also show that TLS-derived volume based on the surface differencing method obtained the best prediction accuracy of individual shrub AGB (R2 = 0.91, RMSE = 79.98 g), and can be used as an alternative of destructive harvesting. This study provides a new insight for quantifying and mapping individual shrub AGB using UAV-based optical sensors and TLS without destructive harvesting in arid and semiarid grasslands.