Research Of Hyperspectral Image SVM Classification Combining Spatial Information

The spectral resolution of hyperspectral images is high so that it can precisely characterize the distinct materials of objects on the surface of the earth, thus making hyperspectral imagery more suitable for classification than multispectral imagery. Due to its features, hyperspectral data aroused much attention among the remote sensing community.In classifying hyperspectral images, the higher dimentionality of data increases the accuracy for the classifier to detect and distinguish various classes on the ground, and brings about new issues such as the high cost of true sample labeling, insufficient training samples, demanding operational speed for computers, and the "Hughes" phenomenon caused by the numerous spectral bands. The support vector machine(SVM for short) can solve most of the problems mentioned above. In the thesis, we tried to introduce spatial information into SVM classification of hyperspectral images, combined it with spectral using two distinct strategies and developed several SVM-based classifiers. The main contents are as follows:(1) The algorithm that combine texture extracted by fixed-window-based methods with SVM. Since one of the representations of spatial information in images is the texture, we designed SVM-based classifiers that contain texture, which is extracted using the fixed-window-based methods. First, we compared histogram, autocorrelation, edgefrequency and some other texture features given by gray level co-occurrence matrix and selected the most informative ones. Then, we combined the selected texture features with spectral features and used the compsite kernels SVM to classify the images. Experiments show that the algorithm improves the classification results slightly.(2) Algorithms that combine image segmentation with SVM. Since the shapes of the ground objects are also one of the representations of spatial information in images, we adopted the idea of image segmentation to build three classifiers, i. e. watershed segmentation, K-means clustering and minimum spanning forest(MSF for short) segmentation. Experiments show that the algorithms have stronger capabilities for classifying wide ground objects, while the accuracies of classifying narrow ground objects are decreased.(3) Iterative algorithms that combine spatial information of pixels’nerborhood with SVM. After analyzing the shortages of the fixed-window-based methods and the image-segmentation-based methods, we further proposed the spatial-contextual composite kernel support vector machine(SCCSVM for short) algorithm. In iterations and extractions of neighbor information, SCCSVM automatically finds a balance between the spatial and spectral information, and gets the best results. Further more, in order to reach a compromise between computational complexity and classification performance, we proposed two more algorithms which run faster. Experiments show that all of the three algorithms give better results than the rest.