|Abstract:Hierarchical spectrum recognition is pointed out in the aspect of hyper-spectral image classification using Hyperion images of Lanzhou and Yuyang in China, by the comparison study of SVM supervised classification.There are 4 problems for hyper-spectral image classification: spectral information fidelity image fusion, extraction of sensitive band, removal of "salt and pepper effect" which avoids misclassification for "same body with different spectrum" phenomenon. Hierarchical spectrum recognition puts forward 4 methods for solving these problems,which are WP-GS image fusion, derivative transformation, object oriented segmentation with 4 scales and multispectral SAM. Hierarchical spectrum recognition can identify 9 kinds land type exactly, based on sensitive bands extracted from derivative transformation image. Visual examination and quantifiable evaluation have been executed to verify authenticity. By contrast, SVM supervised classification is tested, with Gram-Schmidt Spectral sharpening/ Savitzky-Golay convolution filtering/PCA transformation.Using the method of hierarchical spectrum recognition proposed in this paper, the above 4 problems for hyper-spectral image classification are solved, and good classification results is achieved, with overall classification accuracy and kappa coefficient as 89.52%, 0.852 in Lanzhou, 91.12%、0.873 in Yuyang. Growth of 18.68% in Lanzhou and 17.80% in Yuyang for overall classification accuracy, as well as17.52% in Lanzhou and 16.89% in Yuyang for kappa coefficient are achieved compared with SVM supervised classification.Hierarchical spectrum recognition can provide more accurate recognition results are provided for Hyperion images classification in comparison with SVM supervised classification .