Machine learning-enhanced near-infrared spectroscopy for high-throughput phenotyping of sweetpotato sugars across raw and cooked states

Sweetpotato is a major root crop with high yield and nutritional benefits. However, existing methods for evaluating sugars level are inefficient, limiting the breeding and processing of high-quality varieties. This study utilized near-infrared spectroscopy (NIRS) coupled with machine learning algorithms to develop a high-throughput assay for fructose, glucose, sucrose, and maltose in sweetpotatoes across their raw, steamed, and baked states. Leveraging representative samples, characteristic spectral variables, and advanced learning algorithms, twelve optimal models were established for the four sugar indicators under three processing states. These models exhibited outstanding performance in calibration ( R 2 C : 0.941–0.984), cross-validation ( R 2 CV : 0.926–0.976), external validation ( R 2 V : 0.898–0.971), and the ratio of prediction to deviation (RPD: 5.83–10.3), confirming their robust predictive capacity. The findings suggest that these machine learning-enhanced NIRS models enable rapid, high-throughput analysis of sweetpotato sugars, significantly benefiting both breeding programs and food processing applications. • Machine learning enhances NIRS for high-throughput sweetpotato sugar phenotyping. • Robust models quantify sugars across raw, steamed, and baked sweetpotato states. • Optimized models ensure high accuracy and reliability for sugar content prediction. • Efficient NIRS reduces costs and time for postharvest quality assessment. • Insights aid sweetpotato breeding and consumer-oriented product development.