Modelling forest vertical structure by SAR-LiDAR data fusion in the arctic-boreal zone
Abstract
The synergies of SAR and LiDAR have been widely employed in mapping vegetation structure, such as canopy cover, height, and biomass. LiDAR has been considered as the gold standard for remote sensing of vegetation as it can directly measure canopy and subcanopy height and density, whereas SAR backscatter contains mixed information about aboveground vegetation and subsurface soil. In this work, we aim to extend the limited spatial coverage of LiDAR data to the one of SAR, and explore how to decouple the SAR backscatter signals coming from the vegetation and soil. We first trained a Random Forests (RF) regression model using LiDAR-derived vegetation metrics, Tree Canopy Cover (TCC) and Canopy Height Model (CHM), along with SAR and DEM data. The boreal forest site in Delta Junction, Alaska was investigated first as there are multiple LiDAR and SAR datasets available in the area: discrete LiDAR of G-LiHT, waveform LiDAR of LVIS, and L-band SAR of airborne UAVSAR and spaceborne ALOS-2 data. The trained RF model can reveal the spatial distributions of different land cover classes (deciduous/evergreen/shrub), as well as wildfire burn scars. Based on the variable sensitivity analysis of the RF model, HV backscatter and local incidence angle provide most influence in the model, which is expected as they carry the information about radar volume scattering and viewing geometry respectively. DEM elevation is also important as it can help distinguish trees and shrubs for tree lines at high elevations. The validation against G-LiHT data shows RMSE values of 13% and 1.7 meters for TCC and CHM, respectively, and both have a R2 value about 0.8. The SAR TCC underestimates when LiDAR TCC is over 80% and is capped at ~90%, showing the signal saturation of large tree volumes. Similar patterns are also observed in Landsat TCC, where its TCC estimates are capped at 80% in the taiga-tundra ecotone. To test our model transferability over different ecoregions and soil moisture seasonality, we use the model to map TCC and CHM over various sites in the arctic-boreal zone using time-series of ALOS-2 and UAVSAR data acquired during the ABoVE airborne campaigns. Comprehensive analyses on both data and models are made to help understand and distinguish the contributions of vegetation structure and water content, soil moisture, and topography to total SAR backscatter.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2020
- Bibcode:
- 2020AGUFMB100...06C
- Keywords:
-
- 0414 Biogeochemical cycles;
- processes;
- and modeling;
- BIOGEOSCIENCES;
- 0428 Carbon cycling;
- BIOGEOSCIENCES;
- 0439 Ecosystems;
- structure and dynamics;
- BIOGEOSCIENCES;
- 0475 Permafrost;
- cryosphere;
- and high-latitude processes;
- BIOGEOSCIENCES