A Vision-Based Deep Learning Framework enables High-Accuracy Prediction of Geng Rice Eating Quality and Facilitates QTL Mapping.

Northeast China's Geng rice (Oryza sativa subsp. japonica) dominates the high-value rice markets in China due to its superior eating quality. However, current evaluation methods rely on either labor-intensive, subjective sensory protocols or low-accuracy, calibration-heavy near-infrared spectroscopy (NIRS), constraining breeding for high eating quality and market development. Here, we report a vision-based deep learning framework combining multi-population fine-tuning with industrial vision-language model (VLM) pre-training for Geng rice eating quality prediction. Trained on natural and recombinant inbred (RI) population datasets, our optimal model (Model 4) showed high cross-population stability. It achieved R2 values of 0.98, 0.57, and 0.61 in a natural population validation set (35 cultivars), an independent DA-RI population (201 lines), and a randomly collected set (30 Northeast and 28 Southern cultivars), respectively, consistently outperforming the widely used Satake STA1B analyzer. Furthermore, our approach enabled the mapping of a novel, robust quantitative trait locus, qIVOE7, for Geng rice eating quality on Chromosome 7. Further analysis suggested that Model 4 appears to rely on the Hue dimension of the HSV color space for its predictions. This framework provides a high-accuracy prediction model and an easy-to-use tool for rice eating quality evaluation, accelerating high-quality rice breeding as well as the development of the high-quality rice market.