Modeling environmental trends for environment identification in common bean breeding

Environmental trends (EnvT) are a measure used to quantify environmental quality for the expression of phenotypic performance and to understand how varieties respond to different environmental conditions. The identification of these trends enables the modeling and classification of environments into environmental types, providing a basis for a better understanding of genotype × environment interaction and more accurate cultivar recommendations for specific environmental conditions. The objective was to classify environments for the early identification of stressful environments in order to guide cultivar recommendations and optimize the planning of experimental networks in breeding programs. A historical dataset from the Embrapa Rice and Bean variety testing program, covering the period from 2017 to 2023, was used, comprising experiments conducted across the states of Paraná, Mato Grosso do Sul, Goiás, and Minas Gerais, over all growing seasons. A total of 60 environments were defined by the combination of year × location × growing season. Forty climatic variables from NASA POWER and 18 soil variables from SoilGrids were obtained. Climatic variables were obtained at a daily temporal scale and aggregated into 5, 15, and 30-day intervals, as well as over the entire crop cycle. Environments were classified as stressful, moderate, and favorable. Soil and climatic variables were used as predictors of EnvT through XGBoost, Random Forest, and Bayesian mixture models. The dataset was divided into traning and testing, with a ratio of 70/30 in a cross-validation scheme, using 100 repetitions. Random Forest showed the best performance across all evaluated temporal scales, achieving an accuracy of 53.3% and a balanced accuracy of 41.5% using variables aggregated into 15-day intervals. Among the 60 evaluated environments, 13 were classified as stressful, 22 as moderate, and 25 as favorable. The imbalance among environmental classes likely contributed to the reduction in accuracy, particularly balanced accuracy, indicating that class balancing strategies may improve predictive performance. Permutation importance analysis revealed that the most relevant variables for environment classification were maximum temperature, soil temperature, and evapotranspiration during the reproductive period; shortwave radiation, maximum temperature, relative humidity, accumulated precipitation, evapotranspiration, soil temperature, and sunshine duration during the vegetative period; and silt content in the 5-15 cm soil layer. These results highlight the predominance of climatic factors in determining environmental trends for common bean, although silt content also showed high importance. According to studies in soil science, the silt fraction contributes to soil water retention, which may mitigate potential heat and water stress effects, particularly during the vegetative and reproductive periods, when maximum temperature ranked among the most important variables. Environment classification represents a promising tool for the early identification of stressful environments, guiding cultivar recommendations and optimizing experimental network planning in breeding programs.