6th International Workshop on Retrieval of Bio- & Geo-physical Parameters from SAR Data for Land Applications

Interest of Support Vector Regression (SVR) for Biomass retrieval from POLSAR data over Boreal Forests

19 Nov 2015, 11:10

20m

Harwell, UK

Forestry V - Forestry

Speaker

Dr Pierre-Louis FRISON (Universite Paris-Est Marne-la-Vallée - MATIS - IGN / CESBIO)

Description

SAR data are well suited for the retrieval of biomass from vegetated areas. In particular, the ability of low frequencies to deeply penetrate dense vegetation, allows P band radar data to retrieve high biomass levels that can be encountered over forests, where other frequencies show their limitations. Up to know many biomass retrieval methods are based on multi-linear regressions using the different polarized channel intensities acquired in HH, HV, and VV polarization. Such linear behaviour is not necessarily observed by the measured parameters with respect to biomass. The present study aims to assess the potential on Support Vector Regression (SVR) for biomass retrieval from polarimetric SAR data. Support Vector algorithms have been widely used especially for classification of remote sensing data. It is especially well suited to take into account multiple parameters, not necessarily homogeneous in a physical point of view. The SVR algorithm allows to better fit learning data, displaying linear behaviour or not. The polarimetric parameters consists in different parameters obtained from polarimetric decomposition algorithms (the entropy, anisotropy, Freeman parameters, degree of coherence between HH and VV polarizations,…) The method is assessed on SAR and in situ data collected in 2007 and 2008 over Remningstorp and Krycklan during the BioSAR-1 and BioSAR-2 campaigns. These two data set, collected over two different test sites, with different topographic environment, at different dates (March April and May 2007, and October 2008), allow to evaluate the robustness of the methods with respect to spatial and temporal variations. These results will be analysed in the frame of the 7th ESA Earth Explorer mission BIOMASS, for which the number of training plots is a critical point .