Research On Sparse Representation And Space Constraint Based Classification Of Hyperspectral Image

With the rapid development of science and technology and the stronger and stronger ability of human to deal with data, hyperspectral image classification has become one of the most popular areas of image processing. Hyperspectral remote sensing image contains a large number of spatial information and spectral information, and has high spectral resolution and spatial resolution. Spectral information decides the spectrum characteristics of ground objects, spatial information decides the geometric features of objects, combine the spectral information and the spatial information of hyperspectal remote image is helpful to classify hyperspectal remote image. Therefore, making full use of spectral information and spatial information of hyperspectral remote sensing image to classify objects has become a hot spot of recent research.Recently, the sparse representation has been widely used in pattern recognition, signal processing and a lot of fields in image processing. This paper discusses and researches the related problems of hyperspectral remote sensing image classification around the sparse representation method, aimed at search theoretical background and specific application of sparse conception in hyperspectral remote sensing image classification problems, and improves the framework of sparse representation. The mainly works are as following:At first, the paper put forward a semi-supervised local protective dimensionality reduction and an improved joint collaborative representation classification method. First of all, we can obtain a projection matrix for every two types of training samples, the projection matrix makes the distance between samples in the same classes as little as possible, and as far as possible in different classes, and we can obtain a projection matrix for each two classes. When classifying, we can estimate which class is close to the test sample by every projection matrix. Statistical the information, we can get the category of the test sample. Some experiments are taken on Indian Pines and Salinas-A, the experimental results can prove that this algorithm can get a higher accuracy on hyperspectral remote sensing image classification.Secondly,we proposes a classification algorithm which combines semi-supervised local protection linear discriminate dimension reduction algorithm and improved nearest neighbor classifier. The basic idea of the algorithm is that through all labeled samples and some samples that donot have label imformation to find an optimal projection subspace, then trained an improved nearest neighbor classifier. At first,we obtain the dimension reduction data with the projection matrix, then use the training improve classifier to classify the test sample. Some tests on Indian Pines and Kennedy Space Center are taken to prove the superiority of our method. Compared with some algorithms, our algorithm can obtain a better accuracy.Thirdly, this paper proposes a two stage nonlocal similarity joint collaborative representation classification algorithm. At first, we obtain a certain number of training samples for each test sample by using the spectral information, and regard them as the dictionary of the test sample, and combine the nonlocal similarity of test samples and the dictionaries of test samples to represent the test samples, and then obtain the category of the test samples. We confirm the performance of the algorithm through the data set of Indian Pines and Kennedy Space Center.