Non-destructive prediction of soluble solids content in citrus using visible near-infrared spectroscopy

Abstract

The soluble solids content (SSC) of fruits is an important parameter that influences its internal quality. Visible near-infrared (Vis-NIR) spectroscopy is a effective means to detect the internal quality of fruits and vegetables. Measuring samples by instruments generates noise due to environmental factors and machine vibrations, which affects the accuracy of predictions. In this paper, we use standard normalized variables (SNV) and multiplicative scattering correction (MSC) to preprocess the spectral wavelengths, which can effectively reduce the effect of noise. In addition, spectral data contain many redundant variables and useless information, leading to poor prediction of the model. In order to solve this problem, this paper propose a wavelength selection method based on a hybrid strategy of Genetic Algorithm (GA) and Competitive Adaptive Reweighted Sampling (CARS) to screen the effective variables. And the final model is created by partial least squares (PLSR). The GA-CARS model with 84 selected variables has better predictive performance compared to the origin spectrum. In the experiments, samples are obtained from fresh citrus grown in farms around Guilin, and the spectra of citrus are detected in the range of 590 nm-940 nm with a Vis-NIR spectrometer. The experimental results showed that the performance of the prediction model is improved after wavelength screening (RMSEP=0.1581, R2=0.9245). Compared with the traditional algorithm, GA-CARS is an excellent method for screening variables, and the screened wavelengths combined with the model established by PLSR can be a rapid means to detect the SSC of citrus.