SAR Image Classification via Deep Recurrent Encoding Neural Networks

Abstract

With the advancement of imaging techniques, it permits to produce higher resolution SAR data and extend data amount. Therefore, intelligent algorithms for high-resolution SAR image classification are demanded. Inspired by deep learning technology, an end-to-end classification model from the original SAR image to final classification map is developed to automatically extract features and conduct classification, which is named deep recurrent encoding neural networks (DRENNs). In our proposed framework, a spatial feature learning network based on long–short-term memory (LSTM) is developed to extract contextual dependencies of SAR images, where 2-D image patches are transformed into 1-D sequences and imported into LSTM to learn the latent spatial correlations. After LSTM, nonnegative and Fisher constrained autoencoders (NFCAEs) are proposed to improve the discrimination of features and conduct final classification, where nonnegative constraint and Fisher constraint are developed in each autoencoder to restrict the training of the network. The whole DRENN not only combines the spatial feature learning power of LSTM but also utilizes the discriminative representation ability of our NFCAE to improve the classification performance. The experimental results tested on three SAR images demonstrate that the proposed DRENN is able to learn effective feature representations from SAR images and produce competitive classification accuracies to other related approaches.
SAR image classification is a fundamental step for SAR image understanding and interpretation, which aims to label the pixels with different categories.

Method

The rest of this paper is organized as follows. The background of basic deep learning is reviewed in Section II. Section III presents the whole framework for SAR image classification, where recurrent layers and encoding layers are described in detail. Section IV presents the experimental results on three SAR data to verify the effectiveness of the proposed network. Finally, conclusions are summarized in Section V.

Result

In this paper, DRENNs are proposed for SAR image classification, which aim to develop an end-to-end automatic classification model from the original SAR image to the final classification map. Experiments on three SAR data demonstrate that the proposed network can produce superior classification results compared with some related approaches. It is testified that the developed recurrent network based on LSTM is able to learn contextual features automatically, which prefers to discover intrinsical correlations among pixels and outperform other handcrafted spatial filters. Therefore, our recurrent network is more suitable for a large amount of data especially for high-resolution images, due to its automatic feature extraction ability. Moreover, NFCAE is proposed to enhance the discrimination of feature representation, which is verified to improve classification accuracies. Specifically, nonnegative constraint is developed to improve the sparsity and reconstruction quality of feature representation; Fisher constraint is proposed to fine-tune weights to enforce intraclass compactness of feature representation. Based on our results, the proposed network yields higher accuracies in the categories of the road and the building, which indicates NFCAE has the capability to represent effective local information for improving small regional classification. In summary, the whole DRENN not only combines the contextual dependence learning power of LSTM but also utilizes the discriminative representation ability of our NFCAE, which are both conducive to improve the final classification accuracies. Since the computational complexity of our deep network is
large, more efficient architecture will be explored in our future work. Since high wind can affect the backscatter and our data are acquired without high wind weather, the proposed method may be not suitable for data acquired in high wind weather. In addition, our deep NN will be extended for multipolarized data classification.

Note

NOTE:This passage is originated from 《SAR Image Classification via Deep Recurrent Encoding Neural Networks》