Model updating for the classification of different varieties of maize seeds from different years by hyperspectral imaging coupled with a pre-labeling method

The use of hyperspectral imaging technology combined with chemometrics is an effective nondestructive method for sorting seed varieties. However, the performance of the method is susceptible to the influence of time and depends on the training set used in the modeling process. The accuracy of classification models maybe deteriorate when they are used to differentiate the same variety of seeds harvested in different years, due to new variances in the test set are introduced by changes in the cultivation conditions, soil environmental conditions and climatic changes from one year to another. To maintain the accuracy and robustness of model, a model-updating algorithm for differentiating maize seed varieties from different years based on hyperspectral imaging coupled with a pre-labeling method was proposed in this work. The pre-label of each unlabeled sample was obtained using the original classification models developed by the least squares support vector machine classifier. The representative unlabeled samples, which had reliable pre-labels, were selected for updating classification models based on Pearson correlation coefficients. After model updating, the average classification accuracies were improved by 8.9%, 35.8% and 9.6%, compared with those of non-updated models for three test sets, respectively. This shows the effectiveness of the proposed method for classifying maize seeds of different years. (c) 2017 Elsevier B.V. All rights reserved.