A novel and improved near-infrared spectroscopic assay for precise determining sugarcane stalk quality

The global sugarcane industry demands rapid and accurate assessment of cane quality to ensure efficient supply chain management. While our previous study established near-infrared spectroscopy (NIRS) models for mature sugarcane quality traits, those models exhibited reduced predictive accuracy when applied to samples from earlier developmental stages or with low sugar content. This study aimed to enhance the applicability of NIRS-based sugarcane quality quantification through model optimization. We analyzed 200 stem samples from 40 sugarcane germplasms across developmental stages ranging from elongation to maturity. Utilizing partial least squares (PLS), artificial neural networks (ANN), multilayer perceptrons (MLP), and convolutional neural networks (CNN), we developed enhanced near-infrared spectroscopy (NIRS) models to accurately quantify sucrose, reducing sugars, and total sugar content by integrating these diverse samples into a single dataset. Results demonstrated significant enhancements in cross-validated determination coefficients ( R²cv ) and ratio of performance to deviation ( RPD ) for PLS models. Specifically, R²cv values for sucrose and total sugar reached 0.99, with a maximum RPD of 10.18. ANN outperformed PLS in predicting both dry and fresh weight sugar ratios, consistently achieving R²cv > 0.90 and RPD > 3.00. Whereas, the performance of MLP and CNN models were relatively poor. This optimized model substantially reduced outlier proportions, decreasing dry weight glucose outliers from 23.56 % to 2.30 %, confirming its superior accuracy. Analysis of a six-year dataset revealed a stability coefficient as high as 0.85 and facilitated the identification of extreme germplasm with high (> 17 %) and low (< 9 %) sucrose content. This study provides a robust, high-throughput method for sugarcane stalk quality detection, enabling streamlined supply chain management. • Improved NIRS models boosting accuracy application across developmental stages. • NIRS models achieve exceptional R² (up to 0.99) for key quality traits. • ANN outperforms PLS, consistently delivering high prediction accuracy. • Optimized models significantly reduce outliers, enhancing reliability & precision. • 6-year stability validates robust long-term application for supply chain QC.