Optimized Random Forest Framework for Integrating Cultivar, Environmental, and Phenological Interactions in Crop Yield Prediction

Accurate rice yield prediction remains a major challenge due to the complex and nonlinear interactions among cultivar, environment, and phenology. Existing approaches often focus on analyzing individual components while ignoring their interdependencies, which results in limited predictive accuracy and generalizability. To overcome these problems, this study proposes a novel interpretable random forest model that integrates cultivar, environmental, and phenological dimensions. Different from conventional approaches, the proposed method incorporates a factor-combination optimization strategy to identify the most effective information for yield estimation. For analysis, 24 key determinants were screened, including the geographical location, meteorological conditions, phenological events, and cultivar traits. The RF models were also evaluated when built with seven factor combinations. The results reveal the following: (1) Meteorological conditions play a dominant role during the vegetative growth period, including net solar radiation (r = 0.42), daylength (r = 0.38), and thermal summation (r = 0.29). On the other hand, thermal summation (r = 0.28), mean minimum temperature (r = &minus:0.23), and mean temperature (r = &minus:0.20) are most relevant during the reproductive growth period. (2) The full-factor model achieves optimal performance (RMSE = 601.45 kg/ha and MAE = 454.98 kg/ha, R2 = 0.77). (3) Importance analysis reveals that meteorological factors provide the greatest contribution (53.59%), followed by phenological factors (20.39%), geographical factors (17.20%), and cultivar (8.82%), respectively. The results also reveal that threshold effects of key determinants on yield, and identify mid-April to early May as the optimal sowing window. These findings demonstrate that integrating cultivar, environment, and phenology factors creates a powerful predictive model for rice yields.