Polarimetric SAR Image Terrain Classification Based On Sparse Coding Dictionary And Deep Learning

Nowadays, polarimetric synthetic aperture radar(SAR) is not only one of the most important applications in the field of remote sensing, but also a hot spot of many research institutions all over the world. In this context, in order to improve the interpretation of polarimetric SAR image and promote the application of polarimetric SAR system in our country, working on polarimetric SAR image classification research has great practical significance. The propose of polarimetric SAR image classification is classifying the land cover with the same or similar scattering characteristics into a category in order to lay the foundation for the subsequent interpretation. Classification of polarimetric SAR image is a problem with high-dimensional data, we will face the challenge of a large amount of data and computation. This thesis mainly study polarimetric SAR image classification based on sparse coding and deep learning which includes the following three sections:1.A method for terrain classification of ploarimetric SAR image based on improved K-SVD and sparse coding is proposed. First, we learn a dictionary via improved K-SVD. Then, we code testing samples via learned dictionary. Finally, we use a sparse coding classifier based on residuals to classify testing samples. It must be emphasized that this approach can get a discriminative dictionary, so we can use a sparse coding classifier based on residuals to make classification directly. This approach is easy to compute and has good performance.2.A polarimetric SAR image terrain classification approach based on collaborative representation and SAE is proposed. First we represent test samples collaboratively with dictionary learned by K-SVD. Then, we put coefficients into an SAE. Finally, we put learned coefficients into a SVM classifier to make classification. The computation complexity reduces greatly because of the use of collaborative representation. The SAE removes redundant information, accelerating the speed of the algorithm.3.A novel method for polarimetric SAR image terrain classification based on ICA and importance sampling is proposed. First, we use ICA to sparse coding. And then, we choose representative training samples by importance sampling. Next, we put coefficients into a SAE. Finally, we use a SVM classifier to make classification. It's show that we can get a great dictionary because of the use of ICA. The algorithm chooses good training samples via importance sampling, improving the quality of polarimetric SAR image classification.