A novel information extraction method of spectral images based on 3D spectral angle statistics is proposed. By computing the spectral angle between adjacent pixels of the image in the horizontal, vertical and diagonal direction respectively, a 3D information statistical model was then constructed. This model reflects the similarity between adjacent pixels which represent some kind of materials. Uniform areas and edge information of the same material, which will be used for the training sample collection in supervised classification, can be extracted from the image if different threshold values are set and slices are extracted from different axes in the statistical model. Compared with the traditional statistical tools, such as the histogram and scatter diagram, this statistical method has higher robustness and reliability. And it can obtain more information extracted from the spectral images.