Conventional techniques
used to measure oil content in the food
are laborious, rely on chemical agents, and have a negative environmental
impact. In this study, near-infrared hyperspectral imaging was used
as a rapid and nondestructive tool to determine the oil content and
its distribution in commercial flat-cooked and batch-cooked potato
chips. By evaluating various algorithmic models, such as partial least-squares
regression (PLSR), ridge regression, random forest, gradient boosting,
and support vector regression, in combination with preprocessing methods
like multiplicative scattering correction, standard normal variable
(SNV) transform, Savitzky–Golay filtering, normalization, and
baseline correction, the most effective preprocessing method and model
combination was determined to be SNV-PLSR. Moreover, by employing
the optimized PLSR model, a highly accurate oil content prediction
model was developed, achieving a coefficient of determination (R2) of 0.95. To identify the wavelengths that contributed most
significantly to the model’s predictive power, variable importance
in projection (VIP) analysis was utilized. A dimensionally reduced
PLSR model using only 68 selected wavelengths was developed based
on the VIP analysis. This simplified model maintained similar performance
to that of the full-spectrum model while using a smaller data set.
The model was also used to apply the hyperspectral images of potato
chips at the pixel level to visualize the oil distribution in potato
chips with the intent to provide a real-time approach to quality control
for the potato chip industry.