|Paper title||Estimating the vertical distribution of above-ground biomass in tropical forests by combining spaceborne lidar and TomoSAR observations|
|Form of presentation||Poster|
Forests play a critical role in the global carbon cycle. However, estimates of forests carbon storage still have large uncertainties, especially in tropical forests. In addition, the distribution of above-ground biomass (AGB) at certain heights in forests (vertical AGB distribution) is completely underexplored at large scales with remote sensing. Synthetic aperture radar (SAR) and light detection and ranging (lidar) are common remote sensing tools used to estimate AGB. SAR has a large coverage imaging capability, and lidar can achieve high accuracy for measuring forest structure. The tomographic SAR mode (TomoSAR) of ESA’s upcoming P-band SAR satellite BIOMASS together with NASA’s Global Ecosystem Dynamics Investigation (GEDI) spaceborne lidar system will provide an unprecedented opportunity to estimate the vertical distribution of AGB at a regional or global scale. Our objective in this study was to develop and evaluate an approach to estimate the vertical distribution of AGB by combining observations from GEDI and a TomoSAR system (DLR’s airborne F-SAR) for the forest sites in Lopé and Mondah, Gabon, Africa.
According to the ESA WorldCover 10 m 2020 product, the research area in Lopé is covered by 79% trees and 19% grassland. The research area in Mondah is covered by 76% trees, 5% grassland, 15% permanent water bodies and 2% built-up. We used P-band TomoSAR data from the F-SAR system acquired during the ESA AfriSAR 2016 campaign, and GEDI level 2A (ground elevation, canopy top height, relative height metrics) and level 4A (footprint level above ground biomass) products. GEDI data were filtered based on the available quality flags and sensitivity metrics. There were 1,446 and 182 filtered GEDI footprints at 25 m resolution in Lopé and Mondah, respectively. Airborne lidar data from NASA Land, Vegetation, and Ice Sensor (LVIS) was used as reference.
Firstly, we applied the Capon method to reconstruct the reflectivity profiles from 10 tracks HV polarised P-band SAR images. We normalised the tomographic intensities into [0, 1] and used 0.1 as minimum threshold to cut profiles. The lowest peak of each profile was regarded as ground (relative height, RH0). The position above the highest peak and with intensity equal to 0.1 was selected as the RH100, considering the penetration capability of P-band microwave. The relative height (RH) retrieved from GEDI, TomoSAR and LVIS were compared at 25 m and 200 m spatial resolution, representing the resolution of LVIS and GEDI height products, and the resolution of future BIOMASS height products, respectively. Instead of using common height-AGB allometry relationships or power law models based on TomoSAR intensity at certain height level (e.g., 30 m), we attempted here to estimate total AGB from the TomoSAR profile directly. This approach also enables us to quantify the contribution of different TomoSAR height levels to the estimation of total AGB. Therefore, with GEDI AGB as response, random forest regression was then applied to estimate total AGB from TomoSAR profiles at 50 m (resolution of LVIS AGB product) and 200 m resolution (resolution of BIOMASS AGB product). The input features are TomoSAR intensities from 0 to 60 m in 5 m steps. Theses profiles were subset to start from R0 and the intensities above RH100 were set to zero for ensuring the fixed length (i.e., 13) of predictors. The samples were split into training set (80%) and testing set (20%). A five-fold cross-validation was carried out to test model’s transferability and the model with highest coefficient of determination (R²) from five cross-validation models was selected as the final model. In order to estimate the vertical distribution of AGB, we combined in-situ measurements and data from the Biomass And Allometry Database (BAAD) to describe the vertical AGB distribution of individual trees. Therefore, the crown was modelled as a sphere and the stem was modelled as a cone. These individual AGB profiles were then summed up to get the AGB profile at the plot scale. An optimal extinction factor for the P-band microwaves was estimated based on the root-mean-square-error (RMSE) between TomoSAR profiles and normalised field AGB profiles at grid level. Considering the discrepancy between TomoSAR RH100 and in-situ measured (or modelled) forest height, we subset the TomoSAR profiles corresponding to field plots using in-situ forest height rather TomoSAR RH100. By combining the estimate of total AGB with the optimised extinction factor and the TomoSAR profiles, we derived the vertically distributed AGB of the whole research area at 50 m resolution.
Our results show that the RH metrics from GEDI, TomoSAR and LVIS match well in the two study areas. For the cross-validation of the random forest model, models for Lopé (R² = 0.77) and Mondah (R² = 0.81) have similar performance at 50 m resolution, while model for Lopé (R² = 0.86) at 200 m performs better than that for Mondah (R² = 0.77). In both study sites, the R2 between predictions and reference data at 200 m resolution is around 0.2 higher than the R2 at 50 m resolution when these models are extended to the whole research area. The feature importance of random forest model in Lopé and Mondah show that the tomographic intensity between 20 m and 40 m contribute most to the total AGB. From the perspective of normalised root-mean-square (NRMSE), the forest height estimated from TomoSAR satisfies the requirement of the BIOMASS mission (BIOMASS: 30%, Lopé: 13%, Mondah: 11%), while the AGB from TomoSAR does not (BIOMASS: 20%, Lopé: 26%, Mondah: 28%). With an optimal extinction factor, the mean R² between reconstructed TomoSAR AGB profiles with their counterparts derived from field observations is 0.7. In summary, our results demonstrate the potential of combining spaceborne lidar measurements with future spaceborne TomoSAR measurements to get a more detailed insight in the vertical distribution of biomass in tropical forests and understand performance limitations of prospective BIOMASS products.