Application driven key wavelengths mining method for aflatoxin detection using hyperspectral data

The first step of developing Aflatoxin intelligent sorter is to determine the key wavelengths for aflatoxin detection. In order to find more accurate wavelengths, in this paper, three kinds of sensor system are built separately: the first sensor is hyper-spectrometer by ASD spectrometer, the second is the multispectral camera system based on Liquid crystal tunable filter (LCTF) and the third one is the hyperspectral camera based on Grating spectrometer module (GSM). Under 365 nm UV LED illumination, using these three systems, three hypersepctral datasets of 45, 41 and 73 peanut samples before and after aflatoxin contaminated have be collected separately. In order to select the best key wavelengths, four feature selection methods (Fisher, SPA, BestFist and Ranker) and four classifier models (KNN, SVM, BP-ANN, RandomForest) were analyzed and compared. Using all selected wavelengths based on different datasets, a weighted voting method was proposed and 10 key wavelengths (440 380 410 460 420 370 450 490 700 600 nm) were selected. Based on the best model (Random Forest), the best integrated average recognition rate is 94.5%. And then, using these key wavelengths and the best classification model, a new design system for aflatoxin sorter base on a Ploygon mirror was proposed. The structure of this system is simple, detection accuracy is high, which can be applied to online sorting of aflatoxin detection.