Fisher Vectors for PolSAR Image Classification

JAVIER REDOLFI; JORGE SÁNCHEZ; ANA GEORGINA FLESIA

IEEE GEOSCIENCE AND REMOTE SENSING LETTERS

IEEE-INST ELECTRICAL ELECTRONICS ENGINEERS INC

Lugar: New York; Año: 2017 vol. 14 p. 2057 - 2061

1545-598X

In this letter, we study the application of the FisherVector (FV) to the problem of pixel-wise supervised classificationof PolSAR images. This is a challenging problem since informationin those images is encoded as complex-valued covariancematrices. We observe that the real part of these matricespreserves the positive semidefiniteness property of their complexcounterpart. Based on this observation, we derive an FV from amixture of real Wishart densities and integrate it with a Potts-likeenergy model in order to capture spatial dependencies betweenneighboring regions. Experimental results on two challengingdatasets show the effectiveness of the approach.