Study Of Spectral-Spatial Classification Methods For Hyperspectral Imagery

Compared to other types of remote sensing data, hyperspectral data contain much graphic information and spectral information. As a consequence, hyperspectral imaging techniques are widely used for many applications, including environmental monitoring, mineralogy, astronomy, surveillance and defense. Nevertheless, the high dimensionality of the pixels, the undesirable noise, the high spectral redundancy and the spectral and spatial variability, in conjunction to limited ground truth data, present challenges to the analysis of hyperspectral imagery. It was recorded that most of the previous classification techniques process each pixel independently without considering spatial information, such as spatial structures, adaptive spatial neighborhoods, etc. Recently, some research on hyperspectral image classification focus on the incorporation of spatial context features with spectral signatures in the classifier, which inspired us on this work.In this thesis, we consider spectral-spatial classification as the main line of our study. By using both of spatial information and spectral information, we proposed two novel spectral-spatial classification methods for hyperspectral data as follows:Based on the study of hyperspectral anisotropic diffusion partial differential equation (PDE) model, we use the algebraic multigrid method to construct a multigrid structure for multiscale representation of hyperspectral data. By taking advantage of the vertices in each coarsen-grid for the marker-based hierarchical segmentation, we can segment the hyperspectral data into different segmentation regions. The next step consists in employing a maximum vote decision rule to combine the pixel-wise classification map and the segmentation maps. Finally, a final classification map is achieved by choosing the optimal grid level to extract representative spectra. The experiments based on three different types of real hyperspectral datasets with different resolutions and contexts, demonstrate that our method can the highest accuracy when compared to several marker-based spectral-spatial classification method.Based on the edge-preserving image filtering method, we extend the bilateral filter to optimize the support vector machine (SVM)probability results, by using the filtered SVM probability data and considering both of spatial information and spectral information for each pixel. Then, we build new Gibbs energy model for each class of interested, and each energy model is solved by the standard minimum cut algorithm. As a consequence, the range of each class is extracted and form the final spectral-spatial classification map. Experiments show that our spectral-spatial classification method can achieve the best classification map with highest accuracy and more homogeneous areas, compared to the traditional methods.As for the above two spectral-spatial classification methods, we have developed a software for hyperspectral data analysis and processing using C++ language. Our software not only provides many functions for remote sensing data, such as image reading, displaying, saving, filtering, segmenting, classifying and accuracy assessment, but also provides several state-of-the-art spectral-spatial classification methods of hyperspectral data, which can be used for the subsequent research of spectral-spatial classification methods.