Can a multi-model ensemble improve phenology predictions for climate change studies?

Predicting phenology, the timing of developmental events, is critical for understanding how plants respond to the changing climate. Many prediction models have been developed during the last decades, but their use has been limited because of incomplete understanding of internal processes and lack of observation datasets needed for calibration and validation. Dependency on species and locations further complicates the model selection procedure which is an essential part of phenology predictions. To overcome the limitations raised by using a single model, we propose a multi-model ensemble that simplifies model selection and provides competitive performance. We hypothesize that 1) no single individual model consistently outperforms the others and 2) an ensemble model performs equally as or better than any individual models. Nine individual models based on the concept of thermal-time accumulation and their ensembles were cross-validated with 137 datasets of four species collected from multiple locations and years in the United States and South Korea. Non-parametric tests concluded that the performance of a simple mean ensemble model was as good as the best individual model and outperformed the others. Differences between individual models were not statistically significant. The use of ensemble, however, does not preclude any bias in the interpretation caused by characteristics of the underlying models. When the ensemble was classified into groups: 1) with and 2) without chilling components, to assess spring phenology of flowering cherry species in the long-term projections, the predictions of two ensemble groups diverged considerably under RCP8.5 scenario. Our results suggest that a simple ensemble model can be a good phenology prediction tool for avoiding the pitfalls of model selection and reducing inherent uncertainties in climate change studies, but also highlight the importance of implementing the underlying mechanisms of key physiological processes into individual models used in an ensemble.