Deep Supervised and Contractive Neural Network for SAR Image Classification

Abstract

The classification of a synthetic aperture radar
(SAR) image is a significant yet challenging task, due to the presence of speckle noises and the absence of effective feature representation. Inspired by deep learning technology, a novel deep supervised and contractive neural network (DSCNN) for SAR image classification is proposed to overcome these problems. In order to extract spatial features, a multiscale patch-based feature extraction model that consists of gray level-gradient cooccurrence matrix, Gabor, and histogram of oriented gradient descriptors is developed to obtain primitive features from the SAR image. Then, to get discriminative representation of initial features, the DSCNN network that comprises four layers of supervised and contractive autoencoders is proposed to optimize features for classification. The supervised penalty of the DSCNN can capture the relevant information between features and labels, and the contractive restriction aims to enhance the locally invariant and robustness of the encoding representation. Consequently, the DSCNN is able to produce effective representation of sample features and provide superb predictions of the class labels. Moreover, to restrain the influence of speckle noises, a graphcut-based spatial regularization is adopted after classification to suppress misclassified pixels and smooth the results. Experiments on three SAR data sets demonstrate that the proposed method is able to yield superior classification performance compared with some related approaches.

Method

This paper is distinguished by three main novel contributions.
1) A DSCNN is proposed to develop an efficient feature representation automatically. On each layer of DSCNN, the SCAE requires the feature representation to produce good reconstructions of the input features and, at the same time, to give good predictions of the class labels. The supervised penalty is incorporated to capture the relevant information between features and labels, which tends to discover intrinsical characters of the samples. The contractive penalty with Jacobian term is adopted to enhance the locally invariant and robustness of the encoding representation, which prefers to degrade the influence of speckle noises. Therefore, the whole DSCNN significantly improves the discrimination and robustness of feature representation.
2) A multiscale patch-based feature extraction (MPFE) model is developed to obtain the initial features from the SAR image. The patch-based approach is used to combine homogenous pixels together, which not only reduces the computation complexity but also extracts similar characteristics and removes speckle noises. Moreover, multiscale features are extracted based on gray level-gradient co-occurrence matrix (GLGCM), Gabor transform, and histogram of oriented gradient (HOG) to contain different spatial features on different sizes. This model can extract low-level and midlevel features that contain the details (e.g., edges, lines, and contours) of adjacent pixels.
3) A spatial regularization based on graph-cut is adopted for postprocessing of classification. Through graph-cutbased spatial smoothing, the connectivity is improved, and the misclassified points are reduced. This process provides more competitive accuracies of classification and overcomes the serious contamination of speckle noises further.

Results

Images acquired by three different spaceborne SAR instruments operating at different bands, TerraSAR-X (X-band), Radarsat-2 (C-band), and ALOS-2 (L-band), are applied in the experiments. It is demonstrated that the proposed DSCNN-SR method produces optimal results compared with some related approaches. The proposed model is proved to yield superior classification accuracies on spaceborne SAR images with different bands and different polarizations. Specifically, our DSCNN network can obtain a discriminative representation of sample features, which is able to yield competitive classification results only with simple softmax classifier. Based on sensitivity analysis on the percent of training samples, our network is verified to obtain better performance with fewer training samples. Moreover, in front of the DSCNN, the MPFE model is proposed to extract features from the original SAR image, which enriches the primitive feature representation and reduces speckle noises. After classification by DSCNN, graph-cut-based spatial regularization is applied to reduce the misclassified points, which further improves classification accuracies. In addition, our model performs preferable computational efficiency, which may be significant for application in large-scale and highresolution remote sensing images. According to our experiments and analyses, the proposed
method yields superb classification results and performs preferable computational efficiency. Since the structure of deep network is always complex, we intend to explore a more automatic and effective architecture in future work. Another future research direction is to extend our DNN for dual- or multipolarized data.

Note

NOTE:This paper is originated from《Deep Supervised and Contractive Neural Network for SAR Image Classification》