The lack of precise and efficient methods for evaluating the quality of purple-fleshed sweet potato has become a major bottleneck in accurate germplasm identification and genetic improvement. To address this challenge, this study employed near-infrared spectroscopy (NIRS) to construct and optimize high-throughput predictive models for key quality traits of purple-fleshed sweet potato, including total starch, crude protein, reducing sugar, total flavonoid content, total phenolic content and total anthocyanin content. A total of 150 representative samples were selected, and six high-performance predictive models were successfully developed and optimized using a dual optimization strategy combined with machine learning algorithms. The models achieved high accuracy, with coefficients of determination (R2C) for calibration ranging from 0.936 to 0.992, cross-validation (R2CV) from 0.918 to 0.987, and external validation (R2V) from 0.903 to 0.987. The ratio of prediction deviation (RPD) ranged from 6.55 to 19.80, and the range error ratio (RER) ranged from 21.9 to 63.4, indicating strong model stability and prediction capability. The predictive models developed in this study offer an efficient and practical approach for the high-throughput analysis of purple-fleshed sweet potato quality. They enable quantitative evaluation of nutritional components and facilitate the screening of superior germplasm, thereby providing important technical support for quality improvement and the sustainable development of the purple-fleshed sweet potato industry.
Jiang et al. (Wed,) studied this question.