Nondestructive detection of the pH value of silage maize feeds based on hyperspectral images

Abstract: The fermentation of whole maize silage relies mainly on the reaction of lactic acid bacteria. Among them, the soluble carbohydrates in the raw material can be converted rapidly into organic acids (mainly lactic acid), resulting in a rapid decrease in the pH of the maize silage. At the same time, the degradation of nutrients can be inhibited in the silage free by other aerobic microorganisms, thus preserving the nutrients of the feed. As such, the pH value can be expected to serve as the main influencing factor on the quality of the maize silage. Particularly, the rate of PH reduction has a great impact influence on the preservation of protein. Specifically, the pH values range from 3.4 to 3.8, 3.9 to 4.1, 4.2 to 4.7, and greater than 4.8 in the good, fair, medium, and spoiled quality maize silage. Therefore, it is very necessary to accurately and rapidly detect the pH value in the maize silage. This study aims to achieve the rapid and nondestructive detection of the pH value in the maize silages using hyperspectral techniques. A pH content prediction model was developed for the different qualities of maize silages. The mean spectra of silages were sampled from the range of 936-2539nm by a hyperspectral imaging system. The spectral information contained the useful physicochemical information of silage maize feed and the interference information (such as the dark current and light scattering), because the hyperspectral data was susceptible to the instrument noise and surrounding environment. There were also more burrs in the original spectral curve of silage maize feed after detection. Therefore, spectral preprocessing was used to reduce the interference signals for the subsequent modeling, in order to improve the accuracy and stability of the detection model of silage maize feed quality. Firstly, six pre-processing methods were used to treat the feed spectral data, including the multiplicative scatter correction (MSC), standard normalized variables (SNV), Savitzky-Golay smoothing (S-G), orthogonal signal correction (OSC), the first-order derivative, and the second-order derivative. A partial least squares (PLS) regression model was then constructed to derive two well-effective preprocessing, namely the MSC and S-G. Secondly, the feature wavelength was selected using competitive adaptive reweighted sampling (CARS), variable combination population analysis (VCPA), and iterative retention information variable (IRIV). The number of feature variables extracted by the MSC preprocessed spectra were 23, 9, and 43, accounting for 9.5%, 3.7%, and 17.8% of the full spectral band, respectively. The numbers of feature variables extracted by the S-G were 34, 7, and 16, accounting for 14.0%, 2.9%, and 6.6% of the full spectral band, respectively. Finally, the partial least squares regression (PLSR) and extreme learning machines (ELM) were used to build the pH prediction models of feed full band and feature band. The predictive model of the characteristic band was improved to compare with the full-band prediction model, indicating the greatly reduced calculation amount. The MSC-CARS-PLSR was the optimal algorithm in the combination with the correlation coefficients of 0.962 4 and 0.957 6 for the calibration and prediction sets, with the root mean square error (RMSE) of 0.421 3 and 0.426 6, respectively. The best prediction of silage pH can be achieved to combine the PLSR model. The finding can also provide a reliable and effective way for the nondestructive testing of maize silage pH.