Assessing the performance of indicators resulting from three-component Freeman–Durden polarimetric SAR interferometry decomposition at P-and L-band in estimating tropical forest aboveground biomass

Abstract

ABSTRACT Aboveground Biomass (AGB) estimation performance in a dense tropical forest using Polarimetric Interferometric Synthetic Aperture Radar (PolInSAR) data, at P-and L-band, is evaluated. At high levels of biomass, the backscatter saturation effect leads to a low sensitivity of the backscattered intensity for biomass. This study tries to overcome this problem based on decomposition of PolInSAR data into ground, ground-trunk, and volume components to retrieve the vertical structure information of the forest and polarimetric characteristics of layers. Some sensitive parameters, which extracted from Polarimetric Synthetic Aperture Radar (PolSAR) and PolInSAR data are chosen. Then, many sets of these features are used for assessing biomass estimation using Linear Regression (LR) and Support Vector Machine (SVM) regression models. The data analyzed in this paper are from the P-and L-band airborne dataset acquired by Office National d’Études et de Recherches Aérospatiales (ONERA) over French Guiana in 2009, in the frame of the European Space Agency (ESA) campaign, TropiSAR. The average forest biomass is 374.54 t ha−1 and goes up to 503 t ha−1 for the in-situ plots. The results display that the indicator of double-bounce (C DB) contribution matrix, extracted from PolInSAR data decomposition, is the highest correlated feature to AGB, at both P-and L-band. The combination of PolInSAR indicators improved the Root-Mean-Squared Error (RMSE) values up to 24.76 t ha−1 at P-band and 18.02 t ha−1 at L-band. At the best state, AGB estimation RMSE reduces to 40.95 t ha−1 (12.73%) in simultaneous use of the features derived from PolSAR and PolInSAR data, at P-band. SVM produces a more realistic biomass value than LR. However, cross-validated AGB RMSE of two regression models is close to each other.