Integrating AMMI, GGE and machine learning to understand yield stability and trait contributions in rice under organic and inorganic production systems

The development of rice varieties with stable performance across contrasting production systems is essential for sustainable crop improvement and requires a thorough understanding of genotype × environment interaction (GEI). In this study, 66 rice genotypes (62 landraces and 4 checks) were evaluated in an alpha lattice design with three replications under organic and inorganic production systems over two consecutive years (2023 and 2024). The objectives were to identify genotypes with favourable combinations of yield performance and stability and to determine the contribution of yield-related traits to single plant yield (SPY). Combined ANOVA revealed significant (p < 0.01) effects of genotype (G), environment (E) and GEI for all the studied traits. For single plant yield, GEI accounted for 46.71% of the total variation, surpassing the genotypic effect alone and highlighting differential genotype responses across production systems. AMMI analysis showed that IPCA1 explained 59.5% of the GEI, indicating a highly structured interaction pattern. Based on the AMMI Stability Value (ASV) and mean single plant yield, genotypes G39 (Polinel), G7 (Buthakaima), G47 (Sempalai), G4 (Arcot kitchadi), G44 (Rathasali II) and G57 (Vaigunda red) exhibited favourable combinations of yield performance and stability across production systems, whereas G40 (Ponmani samba) and G29 (Mani samba) showed superior mean SPY with specific adaptation to organic and inorganic production systems, respectively. GGE biplot analysis further confirmed the presence of distinct mega-environments and crossover GEI. Among the evaluated machine learning (ML) models, the Random Forest (RF) model showed the best predictive performance and was therefore selected for Shapley Additive Explanations (SHAP)-based interpretation of the relative contribution of yield-related traits to single plant yield. SHAP analysis identified the number of productive tillers per plant as the major contributor to SPY (33.59%), followed by days to 50% flowering (17.20%) and hundred seed weight (12.78%) under pooled conditions. The integrated use of AMMI, GGE and SHAP-assisted machine learning provided improved insights into genotype performance and yield-related trait architecture under contrasting production systems. Further validation of the identified genotypes across additional locations, seasons and plot-level trials is necessary prior to their broader utilization in rice improvement programmes. Significant GEI and crossover interactions were observed under contrasting organic and inorganic systems. AMMI and GGE analyses identified stable, broadly adapted, and specifically adapted rice landraces. SHAP-based machine learning identified productive tillers per plant as the major contributor to single plant yield. This study integrates AMMI, GGE, and SHAP-based machine learning for rice stability analysis under contrasting production systems.