Chemically Interpretable Wavelength Selection for Near-Infrared Spectroscopic Prediction of Starch Content in Corn Kernel Powder Using Two-Dimensional Correlation Spectroscopy

Abstract

Starch is a major chemical component of corn kernel powder and an important indicator of its processing quality, nutritional value, and commercial utilization. Near-infrared spectroscopy (NIRS) provides a rapid and non-destructive analytical approach for starch determination. However, wavelength selection in existing studies remains largely one-dimensional, which may limit both feature interpretability and model robustness across different varieties. Thus, this study proposes a wavelength selection method based on two-dimensional correlation spectroscopy (2DCOS) for predicting starch content in corn kernel powder from multiple varieties using NIRS. A total of 145 spectral samples (940–1660 nm) are collected from corn kernel powder derived from different varieties and production regions. Seven preprocessing strategies, including single and combined methods, are evaluated. The proposed 2DCOS method is further compared with the widely used uninformative variable elimination (UVE). Partial least squares regression, convolutional neural networks, and support vector machines are then used to develop prediction models. The results show that the 2DCOS reduced the original 128 wavelengths to 25 effective variables and outperformed UVE in both stability and chemical interpretability. The SG+D1-2DCOS-linear-kernel-SVM model achieves an Rp of 0.986, an RMSEp of 0.435%, and an RPD of 5.891. Its prediction error is lower than that of the best UVE-based model and approaches that of the best full-spectrum model while using substantially fewer wavelengths. These findings demonstrate the potential of 2DCOS-guided wavelength selection for developing compact NIRS models and interpreting spectral variation associated with starch content in corn kernel powder.