A Cost-Effective and Scalable Machine Learning Approach for Quality Assessment of Fresh Maize Kernel Using NIR Spectroscopy

In fresh maize breeding, developing robust and accurate near-infrared (NIR) calibration models traditionally requires significant time, cost, and labor. To address these challenges, a novel machine learning approach is proposed using a Prediction-Correction Neural Network (PCNN) that enables effective modeling from small sample sets augmented with synthetic data based on NIR spectroscopy. For key quality traits such as amylopectin, protein, crude fiber, and total sugar, the PCNN achieved residual predictive deviation (RPD) values between 2.821 and 4.862, and coefficients of determination ( R V 2 $R_V^2$ ) ranging from 0.869 to 0.951, using an average of only 32 calibration samples. For sugars including fructose, glucose, and sucrose, the model yielded RPD >2 and R V 2 ≥ 0.747 $R_V^2 \ge 0.747$ with just 62 samples. The PCNN method has also been successfully applied to NIR model development for small sample sets in intact kernel of fresh maize and other crops, including forage maize, rice, wheat, and barley. Compared to Partial Least Squares (PLS) and traditional Artificial Neural Networks (ANN), PCNN delivered RPD improvements of 38.99%−63.20% over PLS and 7.07%−25.82% over ANN. These results highlight the PCNN's high efficiency and accuracy, offering a scalable and cost-effective solution for rapid quality evaluation in fresh maize and other cereals.