COS 154-5 - Enhancing the accuracy of flowering time using high-resolution daily temperature data

Abstract: Climate changes have caused changes in the timing of spring flowering in temperate forests, and the direction and magnitude of these changes are divergent among different species and regions. If a plant's flowering time is not synchronized with the activity of its mutualistic pollinators, this can have negative impacts on the plant's ability to reproduce and survive. Moreover, many local festivals centered around the phenology of individual species, and the festival's success hinges on timing it correctly. Thus, there is a growing need for accurate predictions of flowering time. Based on degree-days calculation, we developed process-based models that simulate the timing of flowering of 7 woody species (Cornus officinalis, Rhododendron mucronulatum, Forsythia koreana, Prunus yedoensis, Rhododendron yedoense f. poukhanense, Rhododendron schlippenbachii, and Robinia pseudoacacia) at national scales. We utilized flowering phenology data collected over nine years from 8 to 24 sites and high-resolution (1km) daily temperature restoration data. In addition, to improve the accuracy, we introduced four types of correction factors (L1, L2, R1, R2) using either linear regression (L-type) or random forest (R-type) algorithms. These correction factors were estimated using a variety of predictors, such as geographic coordinates, altitude, temperature, and photoperiod. We found that the R1-type factor, which was estimated using the random forest algorithm based on the mean temperature over the analysis period rather than the temperature of the phenological year, significantly improved the model's predictive performance for all species, except for R. pseudoacacia (p < 0.05). The Root Mean Squared Error decreased by 3-37% for each species, and in particular, the coefficient of determination increased by about 0.8 in R. schlippenbachii. These findings imply that the geographical and climatic characteristics of a region should be considered to improve the reliability of predicting flowering time with a process-based model. Furthermore, we applied the R1-type correction factor across the country for each species and produced a nationwide map of correction factors at 1-kilometer resolution. Our advanced prediction algorithm for spring flowering time can be used to provide high-quality flowering forecasts.