A method of wavelength selection and spectral discrimination of hyperspectral reflectance spectrometry

Abstract

Regularized regression was used in a discriminant analysis framework to identify the key spectral regions for the separation of hyperspectral reflectance spectra of grapevine leaves. Choice of regularization parameter values was guided by cross-validation: for the field-measured spectra, estimated validation errors<12% were used; whereas for the glasshouse-measured spectra, validation errors were estimated to be >60% so choice was based on training error of <20%. Out of the 1151 wavelength bands available in the data, the analysis selected 12 or so wavelengths that can be used to differentiate the groups of vines studied. Moreover these wavelengths were repeatedly observed to occur in spectral regions known to be linked to plant physiology and condition, specifically 500-550 nm; 660-690 nm; 700-760 nm; and 900-1450 nm.