Impact of G4U and 7-component target decomposition on PolSAR image semantic segmentation

Abstract

Classification of Polarimetric Synthetic Aperture Radar (PolSAR) images is an incredibly challenging task. In early days researchers have applied parametric classifiers like maximum likelihood and Wishart classifiers. Later it is found that each class may or may not follow the Gaussian distribution hence researchers have used nonparametric classifiers like ANN, Decision Tree and SVM etc. It is observed from their work that these classifiers give good classification accuracy, but a lot of time is consumed in extracting the features and training the model. To avoid this, in this paper, Semantic Segmentation with the U-net architecture is used. Compared to covariance or coherency matrix, target decomposition techniques provide more information about scattering mechanisms. Hence, the effect of Touzi and Gulab Singh’s 4- (G4U), 7- component (7SD) target decomposition techniques is studied on Semantic Segmentation. Comparative results show that G4U and 7SD decomposition gives almost similar classification accuracy. It is also observed that out of many of Touzi parameters, the first and second components of α, Φ, λ and τ can effectively classify various features and the classification accuracy is comparable with G4U and 7 SD decomposition. Moreover, the classification accuracy using nine elements of the coherency matrix is less than all the above-mentioned decomposition techniques.