A Simple Method of Predicting Autumn Leaf Coloring Date Using Machine Learning with Spring Leaf Unfolding Date

• Published in: Asia-Pacific journal of atmospheric sciences Vol. 58; no. 2; pp. 219 - 226
• Main Authors: Lee, Sehyun, Jeong, Sujong, Park, Chang-Eui, Kim, Jongho
• Format: Journal Article
• Language: English
• Published: Seoul Korean Meteorological Society 01.05.2022; Springer Nature B.V; 한국기상학회
• Subjects: Atmospheric Sciences; Autumn; Climatology; Deciduous trees; Earth and Environmental Science; Earth Sciences; Ecological function; Geophysics/Geodesy; Leaves; Machine learning; Mathematical models; Mean temperatures; Modelling; Numerical models; Original Article; Phenology; Physiology; Plant phenology; Prediction models; Spring; 대기과학; Spring leaf unfolding; Plant phenology; Autumn leaf coloring date; Machine learning; Prediction model
• ISSN: 1976-7633; 1976-7951
• DOI: 10.1007/s13143-021-00251-4

Predicting plant phenology is considered the foundational for the forecast of ecosystem function and dynamics from species level to global level. However, the exact prediction of plant phenology remains limited because of the challenges associated with adding exact environmental and physiological cues to numerical models. In this study, we developed a simple data-based prediction model for leaf coloring dates of temperate deciduous trees by applying machine learning to datasets obtained from the newly established South Korean national-scale phenology network (NPN). Ground observations of spring leaf unfolding dates for 2009–2018 obtained from NPN together with data on the environmental drivers of leaf coloring (summer mean temperature, altitude) were utilized for the model. The model can be evaluated to have simulated the characteristics of observed leaf coloring dates relatively accurate, with only a two-day difference between the average observed and predicted leaf coloring dates. In addition, the model yielded an RMSE value of approximately 7 days, which is within the acceptable error criteria when compared to the sampling frequency, despite the use of only three input variables. Data-based machine learning using existing spring leaf unfolding data as an input help us predict autumn phenology better, even without precise species-specific physiological knowledge on leaf coloring mechanisms. Consequently, a phenology network across the globe based on steady observations will be favorable datasets for a phenology prediction model that can be applied widely.