Abstract:

To investigate the feasibility of using hyper-spectral imaging technique to detect potato ring rot, hyperspectral
imaging operated in reflectance mode in the wavelength range of 980–1 650 nm was applied to 120 potato samples
(60 qualified and 60 ring rot). The effects of multiple scattering correction, standard normal transformation and savitzky-golay +
first derivative on model performance were compared. Multiple scattering correction was chosen as the best spectral preprocessing
method. Then, 9 characteristic wavelengths (993, 1 005, 1 009, 1 031, 1 112, 1 162, 1 165, 1 225, and 1 636 nm)
were extracted based on the partial least squares method. Two linear discriminant analysis (LDA) models, Fisher-LDA and
Mahalanobis distance-LDA, and four support vector machine (SVM) models, linear kernel SVMs, SVM with radial basis
kernel, polynomial kernel SVM and SVM with Sigmoid kernel, were built for ring rot potato at characteristic wavelengths.
The results showed that Mahalanobis distance-LDA was better than Fisher-LDA while Sigmoid kernel performed best
among all the SVM models. The LDA models were overall better than the SVM models. Thus, the LDA model of potato
ring rot based on Mahalanobis distance was the best model. Its recognition rate was 100% for calibration set and 93.33% for
validation set. This study indicated that hyper-spectral imaging technology can be used to identify potato ring rot.

Key words: hyper-spectral imaging, potato, ring rot, non-destructive detection