Three Dimensional Texture Computation in Feature Space for Hyperspectral Image Cubes

• Main Authors: Jhe-syuan Lai, 賴哲儇
• Other Authors: Fuan Tsai
• Format: Others
• Language: zh-TW
• Published: 2009
• Online Access: http://ndltd.ncl.edu.tw/handle/7dc2ej

碩士 === 國立中央大學 === 土木工程研究所 === 97 === The characteristics of remote sensing imagery exhibit a majority of irregular and complex patterns. Because texture analysis can achieve good results in extracting spatial features from complex images by considering the relationship among adjacent pixels, it is an important method in remote sensing image analysis. Texture analysis of remote sensing imagery mainly uses statistics-based gray level co-occurrence matrix (GLCM) to extract features and improve the classification results. The traditional GLCM is in two-dimensional (2D) form. Because hyperspectral imagery in the feature space has the characteristic of volumetric data, it has a great potential for three-dimensional (3D) texture analysis. Previous studies have extended the computation of traditional GLCM to a 3D form, and performed better in classification. However, the core of 3D computation of GLCM was still in a 2D matrix. To truly explore volumetric texture characteristics, this study further extended texture matrix to a tensor field (Gray Level Co-occurrence Tensor Field, GLCTF) that uses three voxels to extract subtle features from image cubes, and utilizes third order statistical computation. For classification applications, the kernel size for texture computation has a significant impact to the results. For 3D texture computation, kernel size can be determined effectively with semi-variance analysis in the spatial domain. However, in a hyperspectral image cube, one of the dimensions is spectral. Therefore, semi-variance analysis might yield improper kernel size in this dimension. To address this issue, this study developed an algorithm based on separability measures to identity appropriate kernel size in the spectral dimension for 3D texture computation. The developed algorithms were applied to 3D texture computation of Magnetic Resonance Images (MRI), whose dimensions are all spatial, to test its validity. Experimental results demonstrate that GLCTF performs better as expected in real volumetric datasets. Consequently, the method was further extended to extract subtle features from hyperspectral image cubes. Evaluations of the classification results indicate that semi-variance analysis and separability measures can determine more appropriate kernel sizes for 3D texture computation and GLCTF in most indexes has better classification results in the test cases.