Polarimetric SAR Image Terrain Classification Based On Deep Ridgelet Neural Networks

Polarimetric synthetic aperture radar(SAR Polarimetric) image contains abundant target information, which has become a hotspot in the field of remote sensing identification in recent years. Among them, the terrain classification method plays an important role in the interpretation of polarimetric SAR image. It has extremely important application value and great significance in the military, civil and other fields. With the coming era of big data, the volume of data is increasing, and the information also becomes more and more complex, the typical polarimetric SAR image terrain classification methods, especially the shallow networks have had much difficulty in image feature extraction and classification, and the learning ability of characteristics has also met a great challenge. The deep learning methods give new ideas in dealing with the polarimetric SAR image terrain classification. This thesis did research on polarimetric SAR image terrain classification based on the deep neural networks with the focus on feature extraction. A novel algorithm is proposed, and the algorithm is put to use in the experiments on real data as follows:(1) A novel method on polarimetric SAR image terrain classification based on deep wavelet neural networks is proposed. Aiming at solving the limitations shallow network's in feature representation and feature learning, the method used wavelet function as the activation function of the hidden neurons in a neural network. It must be emphasized that wavelet function has a stronger learning and generalization ability than typical sigmoid function. The model consists of two layers of wavelet neural network. The results of experiment showed that the proposed method has a better accuracy than the existing typical methods on polarimetric SAR image terrain classification.(2) An improved method on polarimetric SAR image terrain classification based on deep ridgelet neural networks is proposed. This method focused on the disadvantages of wavelet in dealing with high dimensional non-point singularity, and used ridgelet in geometric multiscale analysis tool as the activation function of the hidden neurons in a neural network. It consists of two layers of ridgelet neural network expanding the advantages of wavelet into higher space. On the basis of scaling and telescopic parameters of wavelet function, a description of direction is given to represent more effectively for feature extraction of polarimetric SAR image. Experiments can verify the validity of the method. Compared with the deep wavelet neural networks, this method has more prominent capability in feature representation.(3) An improved method on polarimetric SAR image terrain classification based on deep quantum ridgelet neural networks is proposed. The typical neural network had the inherent defects on slow convergence rate, thus, the idea of quantum state is effective in combination with ridgelet neural network. It consists of two layers of quantum ridgelet neural network to ensure the accuracy of the polarimetric SAR image terrain classification and reduced the complexity of training network. Meanwhile, aiming at dealing with the unity of features in typical polarimetric SAR image terrain classification, a NSCT for extracting texture features was introduced into this method. The combination of polarimetric features and texture features led to an inprovement of image representation. Experiments showed that the algorithm had a high accuracy and a fast convergence rate on classification.