Comparison of wavelength selected methods for improving of prediction performance of PLS model to determine aflatoxin B1 (AFB1) in wheat samples during storage

Wheat is a widely grown grain crop around the world and is highly susceptible to environmental factors during storage and transportation, resulting in the production of fungal toxins that are harmful to humans. Of these, aflatoxin B1 (AFB1) is the most prevalent and most toxic. In view of this, this study used a self-built portable near-infrared spectroscopy system to predict the AFB1 content of wheat during storage and investigated and compared the prediction effects of different wavelength selection algorithms on the constructed PLS model. Firstly, the NIR spectra of wheat samples at disparate storage stages were acquired using the NIR spectroscopy system. Secondly, the raw NIR spectra were pretreated by Savizkg-Golag (SG) smoothing, standard normal variate (SNV) and normalization in turn. Finally, three variable optimization methods, which were variable combination population analysis (VCPA), variable iterative space shrinkage approach (VISSA) and competitive adaptive reweighted sampling (CARS), were applied to select the characteristic wavelength variables of the pre-processed spectra. Partial least squares (PLS) models based on the optimized features of the three methods were established, respectively. The results obtained showed that the CARS-PLS model had the best overall effect. The root mean square error of prediction (RMSEP) for the best CARS-PLS was 2.0965 μg∙kg-1, the prediction coefficient of determination (Rp2) was 0.9935, and the ratio of prediction to deviation (RPD) was 7.3279. The CARS variable screening method was used to effectively select the characteristic wavebands associated with AFB1 in wheat, compressing the number of wavelength variables, simplifying the model structure and improving model performance. The results reveal that the self-built portable NIR spectroscopy system enables to determine the AFB1 in wheat during storage. Furthermore, through the feature optimization of spectral wavelength variables can effectively exclude undesired wavelength variable information.