The influence of calibration data diversity on the performance of temperature-based spring phenology models for forest tree species in Central Europe

Global temperatures are increasing due to human-driven climate change, with notable implications for the flowering phenology of many forest tree species. Modelling the thermal requirements of these species is critical for projecting the impacts of climate change on forests and for developing appropriate adaptation strategies. Fitting models to phenological observations requires long time series of data, but such data are scarce. Researchers would benefit from combining databases from different locations to fit a single model. The aims of this study are to model the thermal requirements for flowering of the most relevant angiosperm tree species in central Europe and to determine if the accuracy of the models can be improved by limiting the geographic spread of the calibration data. To this end, we fitted the PhenoFlex phenology modelling framework using various subsets of records from the Pan-European Phenology database, which were paired with local temperature data. We used all available data for five species (Acer platanoides, Alnus glutinosa, Betula pendula, Corylus avellana and Fraxinus excelsior) to fit general thermal requirement models. We also fitted models using subsets of the dataset, limiting the calibration sets to data from climatically homogeneous regions and different geographical extents. The general models had average mean absolute errors of 8.51–15.15 days, indicating that they are effective in forecasting flowering onset for central Europe. Predictions did not improve when fitting models with data from temperature-homogeneous areas or from within small geographical extents. These findings suggest that fitting several models to cover parts of an extensive region does not necessarily perform better than fitting a single model for the whole region. This implies that including data from different locations within central Europe when calibrating models would increase the size of calibration datasets without causing a significant increase in model errors. This may help alleviate problems of data scarcity.