Abstract: In this paper, we propose a novel approach based on compressed sensing (CS) and multi-scale pyramid in synthetic aperture radar (SAR) image classification. Firstly, a multi-scale PolInSAR feature space is constructed by wavelet transform and feature extraction on the original image; then, CS provides a transform for the measurement domain and recovers the sparse features in the data domain on the image patches in each scale; finally, the combination of multi-scale sparse features generates the final multi-scale pyramid representation of the image for classification. Motivated by the limitations of sparse coding and general pyramid methods, we not only take the advantages of observation matrix in dimension reduction, but also perform analysis on texture features in different scales. Experimental results on the first batch of PolInSAR data show the presented approach's efficiency.