Non-linear Modeling For Quality Inspection Of Greenage

Non-destructive testing of greenage ingredient indicators is important for the development of its deep processing products.Spectral analysis technology has been widely used in the nondestructive testing of agricultural products.However,when there are too many samples and the number of spectral bands is large,the relationship between the spectral response and the concerned targe value is nonlinear,thus linear models cannot obtain enough prediction accuracy dut to the limits defined by their structure.This paper takes greengage as the experiment subjects and takes extreme learning machine(ELM)and support vector regression(SVR)nonlinear quantitative analysis models to investigate the quantitative prediction ability of nonlinear models based on near infrared spectral images.In this paper,spectral data collection and physicochemical measurement of acidity(pH)values were carried out on greengage fruits.Based on the full-ranges spectra,ELM,SVR nonlinear and principal component regression(PCR)and partial least squares regression(PLS)linear prediction models were built to predict greengage pH values.The results show that for the original full-range spectrum,the nonlinear model prediction accuracy is significantly better than the linear ones.This paper also studied the effect of reprocessing and feature wavelength selection methods on the prediction accuracy of quantitative prediction models.Eight preprocessing methods including multiple scatter correction(MSC),standard normal variate transformation(SNV),Savitzky-Golay(SG)5 point smoothing,first-order differential,second-order differential,SG 7 point smoothing,first-order differential,second-order differential were used to preprocess the spectral data for the aformentioned linear models and nonlinear models.The elimination of uninformative variables(UVE)and successive projections algorithm(SPA)methods were used to select the characteristic wavelengths,based on which,PLS,ELM and SVR models were built.Comparative studies show that spectral preprocessing and characteristic wavelength selection can improve the prediction accuracy of the linear models.However,the accuracy of all nonlinear models surpassed their linear counterparts although their accuracy decreased slightly after preprocessing or characteristic wavelength selection;In particular,the feature wavelength selection can greatly simplify the model,improving the spectral acquisition speed and efficiency of spectral-imaging-based quality inspection.This study has guiding significance for the development of agricultural product quality inspection systems based on spectral imaging and nonlinear models.