Near infrared hyperspectral imaging technique has been used for adulteration detection in food samples for several years. However, the sensor cannot screen beyond a certain material thickness. This work studies a method to determine the penetration depth of near infrared radiations in the context of detection. The case of wheat flour in a polylactic acid sample holder is investigated. A sample holder is specially designed to have the wheat flour thickness vary from 0.5 to 3.5 mm. Hyperspectral images are acquired and the partial least squares regression method is used to determine the amount of polylactic acid in each spectrum. Partial least squares prediction results are interpreted using sensor considerations and the Kubelka–Munk theory. Thereafter, the Kubelka–Munk model is fitted and the penetration depth is determined for each wavelength using the reflectance profiles. Similarities between these results and partial least squares regression coefficients lead to the conclusion that partial least squares, combined with the near infrared spectra, is able to characterize the detection depth. The value is calculated by fitting two linear models on the partial least squares prediction results. As a result, 1.80 mm is found to be the detection depth, defined as the maximum thickness of wheat flour to ensure the detection of polylactic acid through the background. Reflectance profiles also show that the penetration depth is highly dependent on the wavelength. This study aimed at providing a method that could be used to evaluate the penetration depth and the detection depth in other contexts than a polylactic acid target in wheat flour.

The possibility of using near infrared hyperspectral imaging spectroscopy to quantify the water content in commercial biscuits was investigated. Principal component analysis was successfully applied to hyperspectral images of commercial biscuits to monitor their water contents. Variables were selected and water contents quantified using analysis of variance, followed by multiple linear regression, and the results were compared with those obtained with variable importance in projection partial least squares. Initially equal to 212, the number of variables after application of analysis of variance was equal to 10. Analysis of variance–multiple linear regression gave better results: the coefficient of determination (R²) was higher than 0.92 and the root mean square error of validation was less than 0.015. The “prediction images” obtained were very relevant and can be used to study biscuit defects. The methodology developed could be implemented at the industrial level for biscuit quality control and for online monitoring of the uniform distribution of water in the superficial layer of biscuits.