Modeling forest growing stock volume in a north subtropical region using the hierarchical Bayesian approach based on multi-source data

Accurate estimation of Forest Growing Stock Volume (FGSV) is needed to achieve the goal of carbon neutral. Quantitative inversion of FGSV using remote sensing technologies is still a research challenge. Optical remote sensing technology is one of the most important means for FGSV estimation, but cannot provide sufficiently accurate estimates due to lack of canopy structure features and data saturation problem. Although airborne Lidar can overcome the shortcoming of optical sensor data, its high cost in data collection and limited observation area constrain its extensive application. This research employs integration of Sentinel-2, ZY-3 stereo, and airborne Lidar data to explore the performance of FGSV estimation in north subtropical regions, and examines the advantages of using the hierarchical Bayesian approach to develop FGSV estimation models under the condition of small population of sample plots. The objective is to solve low modeling accuracy caused by the single sensor data and insufficient number of sample plots. The results indicate that the hierarchical Bayesian approach based on combination of Sentinel-2 and Canopy Height Model (CHM) data (subtraction of Lidar-derived digital elevation model data from ZY-3 stereo-derived digital surface model data) provides the best estimation results with relative Root Mean Square Error (rRMSE) of 27.6%. The Root Mean Square Error (RMSE) using this approach reduced by 13.6 m3/ha comparing with the RMSE based on Sentinel-2 data alone, and reduced by 7.4 m3/ha based on CHM data alone. The research shows that use of multi-source data can effectively improve the problems of overestimation when FGSV is small and of underestimation when FGSV is relatively high, that is, use of multi-source data can reduce the overestimation by one forth and the underestimation by one third comparing with use of single data source alone. Comparing with traditional modeling approaches such as linear regression and random forest, the hierarchical Bayesian approach can effectively reduce the requirement of number of samples due to use of stratification strategy and reduce the impacts of forest types and terrain differences on FGSV estimation accuracy. This research provides new insights of using integration of different data sources to develop FGSV estimation models to achieve accurate estimates, and provides key technology for FGSV mapping in subtropical regions. © 2022, Science Press. All right reserved.