Speaker
Description
Detecting changes between polarimetric radar observations is important for monitoring the temporal evolution of the biogeophysical properties of natural scenes. The long wavelength of BIOMASS’s fully polarimetric P-band radar provides sensitivity to forest properties and their change over time.
Under the standard complex Gaussian model for single-look complex SAR data, polarimetric covariance matrices estimated by averaging independent samples redondant (complex Gaussian...) follow a scaled complex Wishart distribution. This model provides the basis for the generalised likelihood ratio test (GLRT) for change detection developed by Conradsen and al. Despite its optimal characteristics, this approach can be computationally expensive and does not explicitly distinguish changes in backscatter from changes in the polarimetric scattering mechanisms.
This paper proposes two new descriptors of polarimetric variations.
The first is based on a Log-Euclidean (LE) representation, which accounts for the geometry of positive-definite covariance matrices. while involving simple calculations of means and distances. This geometrically consistent representation may improve detection performance, while closed-form expressions may reduce computational cost, compared to Conradsen and al. Using an intrinsic Gaussian model on this space, a GLRT is developed to discriminate between the hypothesis of a common mean across both dates and that of distinct means for each date.
The second descriptor explicitly uses the eigenvalues decomposition of the covariance matrices. In contrast to the Log-Euclidean approach, this modelling approach aims to capture the geometry of both eigenvectors and eigenvalues through Lie group theory, thereby enabling a more accurate identification of scattering mechanisms and a more reliable estimation of their associated powers. Then, a GLRT detector is derived using a Gaussian model defined on the Lie-group of both eigenvectors and eigenvalues. It enables a deeper characterization of polarimetric changes by discriminating variations in backscattered power and changes in the underlying scattering mechanisms.
The proposed descriptors are evaluated using multitemporal BIOMASS data acquired over forested areas in the Brazilian Amazon. They are compared with the Wishart-based GLRT to assess their detection performance and their ability to characterise polarimetric variations associated with forest changes.
This work is funded by the Centre National d’Études Spatiales (CNES) and CESBIO.