Median-mean line based collaborative representation for PolSAR terrain classification

Abstract

A collaborative representation (CR) based method is proposed for polarimetric synthetic aperture radar (PolSAR) data classification in this work. Although CR can well smooth the PolSAR data and remove the speckle noise but it may degrade the class boundaries in heterogeneous regions. To deal with this difficulty, a weighted CR with considering the edge information is proposed. In addition, to further utilize the contextual information, the residual terms of CR are smoothed while the misfitting terms are minimized. Moreover, the median-mean line metric is used to degrade the outlier effects with involving interpolation or extrapolation of mean and median values. The proposed method called median-mean line based CR (MMLCR) leads to superior PolSAR classification results particularly when a limited number of training samples is available. For example, 94.79% overall classification accuracy is achieved for classification of the Flevoland dataset containing 15 classes with just using 10 training samples per class.

Keywords

• Median-mean line
• Collaborative representation
• polarimetric SAR
• Classification