A vision-based deep learning framework enables high-accuracy prediction of Geng rice eating quality and facilitates QTL mapping.

Geng rice (Oryza sativa subsp. japonica) from northeast China dominates the country's high-value rice market because of its superior eating quality. However, current methods for quality evaluation rely on either labor-intensive, subjective sensory protocols or low-accuracy, calibration-heavy near-infrared spectroscopy (NIRS), thus hindering breeding efforts for high eating quality and market development. Here, we report a vision-based deep learning framework that combines multi-population fine-tuning with industrial vision-language model (VLM) pre-training to predict the eating quality of Geng rice. Trained on datasets from natural and recombinant inbred (RI) populations, our optimal model (Model 4) showed high cross-population stability. It achieved R2 values of 0.98 for a natural-population validation set (35 cultivars), 0.57 for the independent DA-RI population (201 lines), and 0.61 for a randomly collected set of 30 cultivars from northeast China and 28 cultivars from southern China, consistently outperforming the widely used Satake STA1B analyzer. This approach also enabled us to map a novel, robust quantitative trait locus, qIVOE7, for Geng rice eating quality on chromosome 7. Further analysis suggested that Model 4 relied primarily on the hue dimension of the HSV (hue, saturation, and value) color space for its predictions. This framework provides a highly accurate predictive model and a simple tool for the evaluation of rice eating quality, helping to accelerate the breeding and market development of high-quality rice.