Abstract: In order to obtain the best hyperspectral characteristic wavelength for wolfberry grading, a feature wavelength selection method for hyperspectral image analysis based on information entropy was presented. Under different wavelengths, firstly, the self-information of each sample image was calculated, and the mean self-information of hyperspectral images of each class of wolfberry was calculated; secondly, the mutual information between two arbitrary sample classes was calculated to obtain the mean mutual information between the corresponding hyperspectral images. Furthermore, the ratio of the mean mutual information to the sum of the mean self-information which corresponded to each sample class under a certain wavelength was calculated and defined as A. Finally,?it was found that A value could be taken as a quantitative index to select the optimal hyperspectral image wavelength for wolfberry grading. The analytical results showed that the optimal wavelength was 950 nm. Then the texture features of all wolfberry samples under the selected wavelength were extracted, and Fisher discriminant analysis (FDA) was employed to classify six classes of wolfberries for the purpose of verification. The results of this study showed that wavelength selection of hyperspectral image analysis based on information entropy is highly feasible for wolfberry grading.

Key words: hyperspectral image, information entropy, optimal wavelength, wolfberry, classification