<![CDATA[Moisture content is a critical quality attribute of corn starch that affects shelf life, functional performance, and commercial value. This study developed and externally validated a rapid and non-destructive method to quantify corn starch moisture using near-infrared (NIR) spectroscopy and chemometric/machine-learning regression. Commercial corn starch was conditioned at approximately 76% relative humidity (saturated NaCl) for 20 days to generate moisture variability, and spectra were acquired using a SpectraStar XT-R instrument (900-2200 nm). Three spectral pre-processing strategies (MSC, SNV, and Savitzky-Golay first derivative) were evaluated prior to model development. A total of 951 samples were split by stratified sampling into calibration (70%, n = 666) and independent prediction (30%, n = 285) sets. Three models were compared: partial least squares regression (PLSR), support vector regression optimized by particle swarm optimization (SVR-PSO), and a one-dimensional convolutional neural network (1D-CNN). The best performance was achieved by PLSR with SNV (R2p = 0.929, RMSEp = 0.274%, RPD = 3.755), while SVR-PSO with MSC showed comparable accuracy (R2p = 0.929, RMSEp = 0.273%, RPD = 3.762). The 1D-CNN yielded lower predictive performance (best R2p = 0.841). Overall, NIR spectroscopy combined with optimized pre-processing and conventional regression models provides an accurate alternative to gravimetric drying for quality control of corn starch.]]>
