Integrating genotype-by-environment interaction and machine learning to evaluate yield-related traits in Korean rice cultivars

Rice yield is a complex trait influenced by genotype-by-environment interactions (GEI), with climate change further intensifying these effects. Therefore, evaluating crop productivity and adaptability in diverse environments is essential in plant breeding. Here, we evaluated the yield-related traits and yield performance of 25 Korean rice cultivars under early- and late-season transplanting in two consecutive years. By utilizing GEI analysis and machine learning (ML) models, we aimed to identify superior genotypes for each trait based on transplanting seasons and assess trait contribution to rice yield. On average, in the late season compared to the early-season, days to heading were significantly promoted by 25 days, grain filling rate increased by 5%, and brown rice recovery rate improved by 2%, while culm length and panicle number significantly decreased by 4 cm and 1 panicle, respectively. The genotype main effect plus GEI biplot and ML-based rice yield prediction using gradient boosting identified G7 (Unkwang), G12 (Sobi), and G20 (Hyeonpum) as recommend cultivars with high stability and yield suitable for both early- and late- transplanting. ML-based SHapley Additive exPlanation (SHAP) feature importance revealed that spikelet number per panicle was the highest contributor to rice yield during early-season, whereas panicle number was the top contributor during late-season transplanting. These results provide valuable guidance to breeders in selecting ideal parents that are well-suited for specific breeding objectives, such as improving adaptability to different transplanting seasons, and optimizing rice yield under varying environmental conditions.