Stability and transferability of broadly trained phenology models in a changing climate

A variety of phenology process-based models have been developed to simulate environmental influences on the timing of spring and autumn phenophases. Similar performances between different types of mechanistic models have raised questions about reliability of their predictions. To assess the biological relevance of phenology models, we used a seven-decade dataset of five species across 170 sites and 1700 m elevation in Switzerland. We evaluated nine leaf emergence and ten senescence models over time and space. We explored how optimal parameter values and influences vary, reflecting transitions in model aptitude and phenology responses to drivers. Leaf emergence models showed improved predictions at external sites over time, while emergence dates converged across Switzerland. In contrast, leaf senescence models often failed to outperform the null model predicting the mean date of training data and showed divergent performance trends. Trends in optimal parameters indicated species-specific responses to emergence drivers, with cold-climate suited species favouring earlier thresholds for warmth accumulation in spring, while the trends were opposite for warm-climate suited species, except for beech showing stable parameters likely due to strong photoperiod constraints. Warming increased the importance of chilling-related parameters for leaf emergence, while senescence parameter sensitivities remained stable. Spatial analyses revealed that complex models were less robust to training and validation at different elevations than simple models, and that phenological responses may vary non-linearly with elevation, likely due to local adaptations. Senescence models performed better with validation at high elevations, where climatic variables such as cooling temperatures play a large role, while predictions were more challenging at other elevations. These findings highlight the need for further refinement of process-based models to account for all driving influences on plant phenology, particularly for leaf senescence models. Our work demonstrates the potential for process-based modelling techniques to better understand phenology responses to climate change.