Model updating strategy for classification of maize seeds using hyperspectral imaging

Abstract. Hyperspectral imaging is a promising technique for identifying variety of seeds. But the hyperspectral features of seeds are vulnerable to the influence of years. Present modeling methods mostly depend on the training samples in the same year and prediction accuracy of these models will be decreased when these models are used to discriminate the same kind of seeds from the other different years. To improve the stability and robustness of the model, hyperspectral image coupled with model updating strategy based on support vector data description (SVDD) algorithm and Least squares support vector machine (LSSVM) was proposed. SVDD was used to find new samples which were not contained in the original model to update present SVDD model, respectively. Recognition experiments on four kinds of maize seeds under different years showed that the model updating strategy based on SVDD and LSSVM worked well, and the prediction accuracy of model was improved from 84.13% to 93.91% by adding 11.5% to 13.15% of samples under testing from other years to the original training set.