Structural Classification of Forest by means of L-band Tomographic SAR

Abstract

Synthetic Aperture Radar (SAR) signals at lower frequencies can penetrate into forest. This capability allows, through advanced imaging techniques such as SAR tomography, the extraction of relevant information related to the three dimensional structural parameters of the observed scene. Recently, several airborne tomographic SAR experiments at low frequencies have already proven the potential of this technique in order to provide an insight on forest structure. And, in the next decade, future space missions such as BIOMASS and Tandem-L, will allow the monitoring of 3D forest structure, at global scale. However, research is still ongoing in order to translate the three dimensional reflectivity information provided by the tomographic SAR measurements into geophysical information that practitioners are familiar with and can assimilate in their analyses. The major challenge arises from the fact that the measured signal in a given resolution cell results from the combination of the contributions of an undefined number of tree parts (canopy, trunk, leaves …). Therefore, the parameters commonly employed in ecology relying on individual based measures (tree basal area, diameter of the tree crown, diameter of the tree stem, density of trees among others) cannot be retrieved in a direct way. Indirect correspondences between radar measures and forest characterization based on terrestrial observations have to be explored instead. This paper suggests a correspondence between the quantification of forest structure from terrestrial data (both horizontal and vertical) and analogous measures obtained from tomographic SAR data. More specifically, two complementary descriptors, estimated from the tomographic SAR reflectivity profiles, are defined. They account for the horizontal and vertical variability of the profiles and it is proven that they are correlated with the information reflected in the traditional measures of forest structure from terrestrial data. The extent of this correlation is tested on both simulated and real data.