Mapping Forest Biomass using Multi-source Remote Sensing and Deep Learning Methods

Keywords: Terrestrial Carbon Cycle, Forest Aboveground Biomass, Integrated Remote Sensing Technology, Machine Learning
Abstract Type: Paper Abstract

Abstract

Aboveground biomass (AGB) of forests, is an important indicator for evaluating the terrestrial carbon cycle. Remote sensing technologies, especially optical, synthetic aperture radar (SAR), and Light Detection and Ranging (LiDAR) technology, have been recognized as effective methods for large-scale quantification of AGB. Optical satellite cannot estimate the biomass with high accuracy in tropical forests due to cloud cover and signal saturation. At the same time, when using SAR satellite data in low biomass areas, it is often difficult to obtain satisfactory accuracy, affected by dwarf trees and irregular canopies. For overcoming these limitations, we developed a general integrated remote sensing approach for large-scale quantification of AGB and continuous monitoring with high accuracy, and evaluated the carbon stocks and fluxes in the Guangdong-Hong Kong-Macao Greater Bay Area between 2018 and 2022. To improve the accuracy of AGB maps, we extracted 53 optical indices from Sentinel-2 images that can reflect the growth and nutritional status of plants. Using polarimetric SAR (PolSAR) and interferometric SAR (InSAR) techniques, we extracted 45 SAR indices from Sentinel-1 images that reflect plant geometry and canopy density. We calibrated the AGB model using The Global Ecosystem Dynamics Investigation (GEDI) lidar data. Our model can explain account for 72% of variance in forest AGB, which predicts better than models using only optical indices or SAR indices. The results of this study demonstrate the great potential of integrated remote sensing techniques in regularly monitoring the forest AGB to evaluate the carbon stocks and carbon fluxes for a large-scale region.