Input features and parameters optimization improved the prediction accuracy of support vector regression models based on colorimetric sensor data for detection of aflatoxin B1 in corn

Abstract

A new method for quantitative detection of aflatoxin B1 (AFB1) in corn based on colorimetric sensor technology was proposed in this study. First, we chose 12 kinds of color sensitive materials to make the colorimetric sensor array. The prepared colorimetric sensor array was used to collect the olfactory sensory information of corn samples with different degrees of mildew, and the information was characterized in the form of images, and using a new unsupervised feature selection with multi-subspace randomization and collaboration (SRCFS) algorithm combined with support vector machine regression (SVR) to optimize the image features of the colorimetric sensor to determine the best color components combination. The SVR model was established by using the best color components combination to achieve rapid quantitative analysis of the AFB1 in corn; in the process of SVR model calibration, optimization of SVR parameters C and g using particle swarm optimization (PSO). The research results showed that compared with the PSO-SVR model, the mean of correlation coefficient of prediction (RP) of the SRCFS-PSO-SVR model has increased from 0.97 to 0.98, and the mean of root mean square error of prediction (RMSEP) has decreased from 4.8 μg·kg−1 to 3.5 μg·kg−1; the RP of the best SRCFS-PSO-SVR model was 0.99 and the RMSEP was 2.8 μg·kg−1. The results reveal that the quantitative detection of the AFB1 in corn can be achieved by using colorimetric sensor technology combined with chemometric methods. Moreover, the SRCFS algorithm can efficiently find the feature combination of colorimetric sensor.