How can integrated Space–Air–Ground observation contribute in aboveground biomass of shrub plants estimation in shrub-encroached Grasslands?

Abstract

Shrub encroachment in grassland has become an ecological issue of mounting concern. Accordingly, an accurate estimation of aboveground biomass (AGB) of shrub vegetation is the basis for a sound assessment and in-depth understanding of carbon cycling in shrub-encroached grassland ecosystems. Yet the relatively low stature of plants in the shrub community, coupled with the high spatial heterogeneity of their distribution, contributes substantially to greater uncertainty in remote sensing estimation of shrub vegetation’s AGB. This study proposes a space–air-ground integrated approach to accurately estimate the AGB of shrub vegetation in shrub-encroached grassland ecosystems. The results showed that, at the UAV scale, the estimation of AGB for a monoculture shrub was highly dependent on planar geometric features. Based on the orthorectified images obtained from unmanned aerial vehicles (UAVs), four planar geometric features of shrub plants, namely crown area (S), crown perimeter (C), long-to-short crown dimension ratio (A1, A2), were retained as the most crucial predictors for AGB estimation. Among the 102 features related to vertical structure extracted via Light Detection and Ranging (LiDAR), only the crown height variation and the first layer’s density variable were retained. Utilizing the mentioned features and a random forest regression, the AGB prediction model for the shrub Caragana microphylla performed remarkably well, in having an R2 value of 0.84 and an RMSE of 310.14 g/plant. At the satellite scale, there was significant nonlinear relationship between the AGB of the shrubs and the band, texture, and index features extracted from GF-6 imagery. The derived AGB estimation model based on the Random Forest method demonstrates higher accuracy (R2 = 0.81, RMSE = 14.61 g/m2, MAE = 11.26 g/m2) than the linear stepwise regression (SR) and partial least squares regression (PLSR) models. Notably, the green band reflectance was retained in all three modeling approaches despite pronounced differences in their selected features uses. Yet both NDVIre1 and NDREI indices with red-edge bands were more important, suggesting the red-edge bands of GF-6 can serve as an ideal tool for remote sensing investigations of the AGB of shrub in shrub-encroached grasslands. This study provides technical and scientific support for quantitative assessments of shrub AGB in arid and semi-arid grassland regions.

Keywords

• Shrub-encroached grasslands
• Space-Air-Ground Integrated Earth Observation
• Aboveground Biomass
• Shrub plants
• GF-6
• Machine learning