Estimating the first flowering and full blossom dates of Yoshino cherry (Cerasus × yedoensis ‘Somei-yoshino’) in Japan using machine learning algorithms

Abstract

Climate change alters the phenology of various plants. For example, increasing temperatures shift the first flowering and full blossom days of Yoshino cherry trees and affect cultural events related to cherry blossoms. We developed models to estimate the first flowering and full blossom dates of Yoshino cherry in Japan based on temperature and phenological data observed at 82 stations in Japan for 68 years (1953–2020). Three machine learning algorithms, namely, the random forest (RF), artificial neural network (ANN), and gradient boosting decision tree (GBDT) algorithms, were utilized, and the hyperparameters were optimized using Optuna. The GBDT models produced the best estimation accuracy, with an overall root mean square error (RMSE) = 1.53 and 1.48 days for the first flowering date and full blossom date, respectively. Furthermore, our analysis using Shapley Additive Explanations (SHAP) revealed that in the RF and GBDT models, the low temperature in winter and high temperature in spring would advance the estimated first flowering and full blossom dates.