Detection of kiwifruit dry matter content based on hyperspectral technology using uninformed variable elimination coupled with successive projection algorithm

Abstract

The internal parameters of kiwifruit are mostly detected using

traditional destructive physical–chemical methods, which are not

only labor and time consuming but also inconvenient in operation.

The hyperspectral imaging technique is now considered a new

non-destructive method for detecting the quality parameters

of kiwifruits. However, most studies focused on detecting the

soluble solid content, hardness, and ripeness of this fruit. Thus,

the detection precision of this imaging technique needs to be

improved. Moreover, few of these techniques are involved in the

detection of the dry matter content. A non-destructive detection

method based on the hyperspectral imaging technique is proposed

in this study to detect the dry matter content of kiwifruit online

rapidly and precisely. First, the hyperspectral images of kiwifruit

were analyzed, the interested regions therein were extracted,

and denoising was preprocessed using the multiplicative scatter

correction. Second, the redundancy of the 217 pieces of full

band spectral information was researched, and 66 characteristic

spectral bands were initially screened out through uninformed

variable elimination (UVE). The collinearity among these bands

was eliminated using successive projection algorithm (SPA), and

five characteristic spectral bands were extracted. Finally, the dry

matter content of the kiwifruit was detected by taking least squares

support vector (LSSVM) as the detector, by employing particle

swarm optimization (PSO) to optimize LSSVM’s parameters, and

by entering the five bands into the LSSVM later. Test results show

that: (1) the redundancy and the collinearity of the full spectral

bands can be eliminated effectively by combining SPA with UVE

so that the extracted low-dimensional characteristic spectral

bands can reflect the dry matter content of kiwifruit better. (2)

The detection indicators of UVE+SPA+LSSVM to the training set

is that the coefficient of correlation (R) = 0.91, root-mean-square

error (RMSE) = 0.28, and the detection indicators to the prediction

set is that R = 0.89, RMSE = 0.31, indicating that the detection

precision is higher than the other methods. This study shows that

the non-destructive detection method proposed in this paper can

detect the dry matter content of kiwifruit rapidly and efficiently.

This method serves as a theoretical basis for the industrialized

classification of kiwifruit that is based on the internal parameters.