Enhancing Landsat image based aboveground biomass estimation of black locust with scale bias-corrected LiDAR AGB map and stratified sampling

• Main Authors: Qin, Shuhong, Wang, Hong, Li, Xiuneng, Gao, Jay, Jin, Jiaxin, Li, Yongtao, Lu, Jinbo, Meng, Pengyu, Sun, Jing, Song, Zhenglin, Donev, Petar, Ma, Zhangfeng
• Other Authors: Asian School of the Environment
• Format: Article
• Language: English
• Published: 2023
• Subjects: Science::Geology; Aboveground Biomass; Stratified Sampling
• Online Access: https://hdl.handle.net/10356/171569
• Institution: Nanyang Technological University

There is a growing interest in leveraging LiDAR-generated forest Aboveground Biomass (LG-AGB) data as a reference to retrieve AGB from satellite observations. However, the biases arising from the upscaling process and the impact of the sampling strategy on model accuracy still need to be resolved. In this study, we first corrected the bias arising from upscaling the LG-AGB map to match the spatial resolution of Landsat observations. Subsequently, the stratified random sampling method was used to select training samples from the corrected LG-AGB map (cLG-AGB) for the Random Forest (RF) regression model. The RF model features were extracted from the Landsat observations and auxiliary data. The impact of strata numbers on model accuracy was explored during the sampling process. Finally, independent validation was conducted using in situ measurements. The results indicated that: (1) about 68% of the biases can be corrected in the up-scale transformation; (2) compared to no stratification, a three-strata model achieved a 6.5% improvement in AGB estimation accuracy while requiring a 37.8% reduction in sample size; (3) the black locust forest had a low saturation point at 60.52 ± 4.46 Mg/ha AGB and 72.4% AGB values were underestimated and the remaining were overestimated. In summary, our study provides a framework to harmonize near-surface LiDAR and satellite data for AGB estimation in plantation forest ecosystems with small patch sizes and fragmented distribution.