A robust feature wavelength extraction strategy for hyperspectral information:Three cases of potato quality evaluation

Abstract

Because the existing feature wavelength extraction methods for hyperspectral information are only for some specific property of sample, which has some limitations, a feature wavelength extraction strategy with generalization and robustness is proposed. Namely, a part of the same or similar feature wavelengths are extracted by Wilks Λ statistic coupled with principal component analysis from hyperspectral information of different tested samples, and another part are extracted by partial least squares regression coefficient coupled with texture information. And then the two part of feature wavelengths can be integrated together to form the final feature wavelengths that represent the comprehensive properties of these tested samples. Take potato as the research object, and according to this feature wavelength extraction strategy, 8 feature wavelengths applicable to different potato samples were extracted, and three quality analysis cases were successfully implemented: the variety identification model, quality grade detection model and VC content prediction model. The results show that, based on the 8 feature wavelengths extracted by this strategy, the correction rate of the variety identification is 100 %; and the correction rate of the quality grading is above 92 %; and the R2P of the VC content prediction is 0.9715, which shows that the extracted feature wavelengths can effectively represent the different qualities of potato, and has robustness and generalization. This also provides a reference for the study of feature wavelength extraction methods of other agricultural products.