Feature Line Based Feature Reduction of Polarimetric-Contextual Feature Cube for Polarimetric SAR Classification

Extraction of discriminative features is an efficient step in any classification problem such as synthetic aperture radar (SAR) images classification. Polarimetric SAR (PolSAR) images with rich spatial features in two first dimensions and polarimetric characteristics in the third dimension are rich source of information for providing classification maps from the ground surface. By applying the spatial operators such as morphological filters by reconstruction, data dimensionality of the PolSAR is increased and needs feature reduction. In this work, median-mean and feature line embedding (MMFLE) is proposed for dimensionality reduction of the polarimetric-contextual cube in PolSAR images. MMFLE is stable with respect to outliers by utilizing the median-mean line metric. By an appropriate definition of scatter matrices, MMFLE maximizes the class separability. In addition, MMFLE is specially a superior feature reduction method when a small training set is available because it uses the feature line metric to model the data variations and generate virtual samples. With 10 training samples per class, MMFLE achieves 94.15% and 83.01% overall classification accuracy, respectively in Flevoland and SanFranciso PolSAR datasets acquired by AIRSAR.