Somei-Yoshino Flowering Date Prediction and Interpretation Considering Uncertainty through SHAP Value Distribution

In this study, we predict the flowering date of Somei-Yoshino across Japan using a one-dimensional convolutional neural network (1D-CNN) model with daily mean temperature as input. By combining this predictive model with MC dropout, we obtain prediction distributions. Furthermore, applying SHAP (SHapley Additive exPlanations) to the model with MC dropout allows us to obtain SHAP values for each feature of individual data samples, enabling an analysis of their distribution. By consistently identifying characteristics of SHAP value distributions, we can assess the reliability of features in models with inherent uncertainties. We expect this approach to provide new insights that traditional point estimation methods for predicting flowering dates cannot offer. In this study, we first analyze the consistency between the base model’s SHAP values and domain knowledge about cherry blossoms to verify the model’s validity. Next, we select specific locations and years to eliminate variability by location and year to compare temperatures and SHAP values, thereby deepening our understanding of how regional climate conditions affect flowering dates. Subsequently, we apply MC dropout to this base model to obtain SHAP value distributions and conduct a meticulous analysis of their shapes. Finally, based on these analyses, we investigate the relationship between the variance of prediction distributions and SHAP value distributions and quantify the reliability of feature influences with uncertainties on flowering dates. This comprehensive analysis will improve model interpretability and prediction accuracy.