Evaluating the performance of models predicting the flowering times of twenty-six apple cultivars in England

The timing of the transition between endodormancy and ecodormancy remains uncertain. However, with advancements in phenology modelling, we can now fit models which allow for variable transitions between chilling and forcing models. Previous studies have primarily focused on single-cultivar parameterisation, and few have explored multi-cultivar comparative modelling. In this paper, we address this gap by evaluating three parameterisation approaches based on the recently developed PhenoFlex framework using a large flowering time dataset of twenty-six apple cultivars collected at the same location in England. The three parameterisation approaches were: cultivar-specific, group-specific with the groups derived using the K-means algorithm on mean bloom and variation of bloom dates, and a common model (for all twenty-six cultivars). The three PhenoFlex models fitted to each of three groups of cultivars based on their flowering time and the common model fitted to all cultivars achieved similar predictive performance, better than predictions using the average bloom date of each cultivar. The best approach to apply would depend on the amount of data present. The common model works best with large number of cultivars with small datasets (∼10 years), the mean flowering date grouped works best with medium numbers of datasets (∼20 years) and the cultivar-specific model should only be used when each cultivar has at least 30 years of data, however, it is more biased, so it is likely to predict bloom dates later than the observed bloom dates. Finally, the PhenoFlex model was shown to perform better than the StepChill model, where no overlapping is allowed between chilling and heat models. The result of this study indicates that the PhenoFlex model can be used to determine apple flowering time at the species level.