Abstract: Based on the analysis to current algorithms,some new approaches are proposed and experimented.Similarity measure based on spectral polygon considers both albedo and wavelength at the same time so its precision is higher than other methods.Spectral similarity measure based on multi-scale wavelet transformation can combine similarity measure with dimensionality reduction,also have high precision.By experiments it proves that these new approaches can be used to spectral feature-based hyperspectral RS image retrieval effectively.