Nondestructive Classification of Potatoes Based on HSI and Clustering

Abstract

A rapid classification method of potatoes based on the combination of hyperspectral imaging(HSI) technology and integrated learning algorithm was proposed in this paper. Here, potatoes were divided into six types: intact ones, green skin, germination, dry rot, wormhole and damage. Firstly, visible-near infrared (VNIR) hyperspectral imaging system with the band range of 400-1000nm was used to collect the potato hyperspectral image information in the experiment. Further, after the image masked, K-means clustering method was used to segmentate images. Extract the average spectrum of the defect areas and the intact areas as the classification data set. Based on the traditional machine learning algorithm (support vector machine, decision tree) and the integrated learning algorithm (random forest, gradient promotion decision tree), the classification model of potato defects was established and compared. The results show that among all classification algorithms, the classification accuracy of potato defects can be significantly improved by using the decision tree of gradient lifting. By comparing the feature importance of each band, the model accuracy was maintained above 80%. Furthermore, in order to improve the discrimination ability of data and reduce the dimension of data, linear discrimination analysis method was used to process spectral data, and the accuracy of the established model was finally improved to 84.62%.