Affiliations: [a] College of Electrical and Information, Northeast Agricultural University, Harbin, Heilongjiang 150030, China | [b] School of Information Engineering, Lingnan Normal University, Zhanjiang, Guangdong 524048, China | [c] China Energy Engineering Group Heilongjiang Electric Power Design CO.LTO, Harbin, Heilongjiang 150078, China | [d] College of Agriculture, Northeast Agricultural University, Harbin, Heilongjiang 150030, China
Correspondence:
[*]
Corresponding author: Runtao Wang, College of Electrical and Information, Northeast Agricultural University, Harbin, Heilongjiang 150030, China. E-mail: [email protected].
Abstract: Purifying soybean seed is greatly beneficial for improving the quality of soybean planting and products. Since the seeds are usually evaluated in many fields for sowing and oilseed processing, they must be identified quickly and accurately for selecting a correct variety. In this study, an acceptable method combining spectral information operating in the spectral wavelengths from 400 nm to 1000 nm, with machine learning was proposed to classify 10 soybean varieties. After pre-processing original data, principal component analysis (PCA) was performed to reduce the spectra dimensionality. The texture feature parameters (energy, entropy, inertia moment and correlation) were extracted from three feature images selected by PCA. In this way, each sample was represented by 9 features including PC scores, mean value and standard deviations of inertia moment under three feature images. In the experimental study, soft independent modelling of class analogy (SIMCA), partial squares discriminant analysis (PLS-DA), genetic algorithm optimized BP neural network (GA-BP) and Takagi-Sugeno fuzzy neural network (T-S fuzzy neural network) were built to distinguish different soybean varieties. The performance of the classification models was analysed when they were applied to the feature matrix. The average accuracy of the training set was higher than 96% and the average accuracy of the testing set was higher than 84%. Although GA-BP network model got the highest predictive accuracy with 92%, T-S fuzzy neural network model was the best choice considering the accepted identification accuracy, better stability and less computation cost.
