Factors affecting the non-destructive detection of water contents in fresh corn cobs by near-infrared spectroscopy

This study aims to obtain the high-quality near-infrared (NIR) spectra of fresh corn cobs. A systematic investigation was made to explore the effects of experimental parameters on the spectral features of fresh corn cobs and modeling validation using NIR diffuse reflectance spectroscopy. According to the cob stick-shaped characteristics, the multi-dimensional comprehensive test was carried out to collect 900~1700 nm spectral data under four parameters, namely, light source type, light intensity, detection distance, and light source angle. The halogen lamp cups were used as the fiber optic light sources. Among them, the halogen lamp cups were selected with the power of 20 and 40 W, detection distances of 10 and 50 mm, and halogen lamp cups with angles of 30°, 45°, and 60°, in order to analyze the spectral differences and distribution patterns of curves. The spectral differences were determined for the distribution patterns. The standard deviation and spectral area extreme deviation indexes were used to evaluate the spectral quality. Further validation tests were carried out on the model. The spectra were evaluated at 30° and 45° halogen lamp cup angles by multiplicative scatter correction (MSC), standard normal variate (SNV), first derivative (FD), and trend correction. Furthermore, the new model was established to predict the water content using partial least squares (PLS) and support vector machines (SVM). The performance of the model was compared after the derivative (FD) and detrending (DT) pre-processing. The experimental results showed that sufficient spectral response and less interference were achieved in the halogen lamp cup with a power of 20 W and a detection distance of 40 mm. The standard deviation and spectral area polarity of the curve were 0.83 and 187.2, respectively. The better quality of the spectral curve and higher performance of the model were also obtained in the halogen lamp cup with the clamp angle of 45°, compared with 30°. The better performance was found in the SVM prediction model after SNV preprocessing. The coefficients of determination were 0.943 and 0.880, respectively, in the correction and prediction datasets, while the root mean square errors were 0.708 and 0.932, respectively, and the residual prediction deviation was 2.956. The finding can provide technical support to the nondestructive test on the intrinsic quality of fresh corn cobs using near-infrared diffuse reflectance spectroscopy.