Integration of spectral and image features of hyperspectral imaging for quantitative determination of protein and starch contents in maize kernels

Rapid and accurate detection of protein and starch contents is important to ensure maize quality. However, existing methods for rapidly detecting protein and starch contents often suffer from drying samples to improve accuracy but significantly increase the overall detection time. To bridge this research gap, this study aims to integrate spectral and image features of visible near-infrared hyperspectral imaging for rapid determination of protein and starch contents with high accuracy under the effect of water on maize kernels. First, the spectral information at 399.75–1005.80 nm, color (the first, second, and third-order moments of H, S, and V color channels), and texture (contrast, correlation, energy, homogeneity, and entropy) features data were extracted. Then, 40 prediction models, including various preprocessing methods and modeling algorithms, partial least squares regression (PLSR), support vector regression (SVR), and extreme learning machine (ELM), were established and compared. Furthermore, the characteristic wavelengths are selected by successive projection algorithm (SPA) and uninformative variables elimination (UVE), and further fused with color and texture features data to improve model accuracy. The results showed that the best prediction models of both protein and starch contents were built based on fusion data. The standard normal transformation (SNV)-SPA-Color-ELM-Protein model and multiplicative scatter correction (MSC)-UVE-Texture-ELM-Starch model were optimal, which can achieve rapid and accurate detection in maize kernels under the effect of water.