Estimating Spatio-Temporally Continuous GEDI Aboveground Biomass Density During 2015-2020 From Multi-Source Data Using Machine Learning and Deep Learning

Wang, Xia and Zhang, Yihang and Atkinson, Peter M. and Zhang, Kerong (2026) Estimating Spatio-Temporally Continuous GEDI Aboveground Biomass Density During 2015-2020 From Multi-Source Data Using Machine Learning and Deep Learning. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 19. pp. 3839-3857. ISSN 1939-1404

Text (Wang Xia et al IEEE JSTARS)
Wang_Xia_et_al_IEEE_JSTARS.docx - Accepted Version
Available under License Creative Commons Attribution.
Download (18MB)

Abstract

Accurate estimation of forest aboveground biomass density is essential for quantifying terrestrial carbon exchange. NASA's Global Ecosystem Dynamics Investigation (GEDI) LiDAR mission has ushered in a new era of high-quality aboveground biomass density (AGBD) estimation from space. Owing to GEDI's LiDAR data collection, one of the GEDI AGBD products are presented as 1 km spatial discontinuity grids, limiting practical application. This study proposes a method based on machine learning and deep learning to estimate spatio-temporally continuous GEDI AGBD and its associated uncertainty (standard error) during 2015-2020 using multi-source optical, SAR, LiDAR, topographic, and climatic data. MODIS vegetation continuous field (VCF) was used as an indicator to select unchanged GEDI AGBD samples in 2020 for each year from 2015 to 2020, resulting in 1,800–2,400 samples per year. The selected annual training samples were then used to estimate annual spatially continuous GEDI AGBD and uncertainty. Experiments in China's Han River Basin demonstrated that integrating all available datasets resulted in a more accurate spatial continuous GEDI AGBD map in 2020 compared to use any single dataset. Convolutional neural network (CNN) considering spatial neighboring information (CNN_spatial) outperformed 1-D CNN and four benchmark machine learning methods of extreme learning machines (ELM), generalized regression neural network (GRNN), support vector regression (SVR) and random forest (RF) with an R2 of 0.8232 and RMSE of 27.7557. Using the unchanged GEDI AGBD training samples resulted in more accurate GEDI AGBD maps (R2>0.82) during 2015-2020 than using the training samples from 2020 alone. From 2015 to 2020, 2.16% of AGBD pixels in Han River Basin experienced a relatively significant increase, while 1.24% showed a relatively significant decrease. Compared with the ESA_CCI and China's AGBD products, our estimates achieve better accuracy relative to the field plot data. The proposed method offers a solution to generate high-quality spatio-temporally continuous GEDI AGBD in large-scale complex forest landscapes.