The traditional gray level co-occurrence matrix (GLCM) is in two-dimensional form. Because hyperspectral imagery in the feature space has the characteristic of volumetric data, it has a great potential for three-dimensional texture analysis. Previous studies have successfully extended traditional 2D GLCM to a 3D form (Gray Level Co-occurrence Matrix for Volumetric Data, GLCMVD) for extracting features in hyperspectral image cubes by considering pixel-pairs in 3D spatial relationship. However, the core of texture computation was still in a 2D texture matrix form. To truly explore volumetric texture characteristics, this study further extended traditional GLCM to a tensor form (Gray Level Co-occurrence Tensor, GLCT) that uses three voxels to extract subtle features from image cubes. For classification applications, the kernel size for texture computation has a significant impact to the results. This study developed an algorithm based on semi-variance and separability analysis to identity appropriate kernel sizes for three-dimensional computation. Experimental results demonstrate that GLCT performs better in classification than GLCMVD for the texture analysis of hyperspectral image cubes. In addition, the developed algorithm can obtain more reasonable kernel sizes for three-dimensional computation of hyperspectral remote sensing datasets.