Classification Of Polarimetric SAR Images Using Multilayer Autoencoders And Superpixels

Polarimetric Synthetic Aperture Radar(Pol SAR) is one of the most important applications in geoscience and remote sensing. Land cover classification using polarimetric synthetic aperture radar images has increasingly become an important application. It mainly focuses on the target scattering characteristics at present. Looking for a more effective and a higher accuracy in the field of Pol SAR classification has been the mainly research direction. This paper proposes a novel Pol SAR SAR images classification using multilayer autoencoders and superpixels. In the method, both the information in each pixel and the spatial relations between pixels are considered,improving the accuracy of segmentation. We use the intensities information of RGB image to compute the superpixels. And the intensity information is used to optimize the final result. The main contents include the following aspects:Frist, this paper proposes a new Pol SAR images classification method based on multilayer autoencoders algorithm. This method successfully introduces the multilayer autoencoders into the field of Pol SAR. It makes full use of the multilayer autoencoders structure to learn the feature of the input data automatically. It can explore non-linear relationship between the coherent matrix elements. The algorithm is based on pixel and the feature extraction process is unsupervised algorithm. The classification process uses a softmax classifier. The algorithm uses a simple structure, with a better accuracy. Due to the input feature vectors of multilayer autoencoders in the natural image are extracted from raw pixels based on the image patches, the selection of feature vectors for classification can be crucial in Pol SAR data. In the Pol SAR data, each pixel is often represented as a scattering matrix. A series of effective feature vectors based on pixels are extracted from the multilook coherency matrix of Pol SAR data.Second, a new polarimetric synthetic aperture radar(Pol SAR) images classification method based on multilayer autoencoders algorithm with over-segmenting the images by superpixels is proposed in this paper. Firstly, a set of feature vectors that extracted from the multilook coherency matrix of the polarimetric SAR data are used to train the multilayer autoencoders network. Secondly, a set of feature vectors that extracted from the multilook coherency matrix of the polarimetric SAR data are used to train the multilayer autoencoders network. Finally, k-nearest neighbor(KNN) and statistical information are used to produce the final result. In the method, both the information in each pixel and the spatial relations between pixels are considered, improving the accuracy of segmentation.This work was supported in part by the National Natural Science Foundation of China under Grant 61072106 and 61271302; the National Basic Research Program(973Program) of China under Grant 2013CB329402.