Calibration for an Ensemble of Grapevine Phenology Models under Different Optimization Algorithms

Vine phenology modelling is increasingly important for winegrowers and viticulturists. Model calibration often required before practical applications. However, when multiple models optimization methods are applied different varieties, it rarely known which factor tends to mostly affect the results. We mainly aim investigate main source of variability in errors flowering timings two varieties vine Douro Demarcated Region (DDR) Portugal; this based on five model simulations that use optimal parameters estimated by three algorithms (MLE, SA SCE-UA). Our results indicate can be affected initially assumed parameter boundary. Restricting initial distribution a narrow range impedes algorithm from exploring full space searching parameters. This lead largest variation models. At an identified appropriate boundary, difference between represents uncertainty, while choice contributes least overall uncertainty. The smaller among or (tools analysis) compared could reliability calibration. All show similar terms obtained goodness-of-fit: RMSE (MAE) 5–6 (4–5) days with negligible mean bias moderately good R2 (0.5–0.6) ensemble median predictor. Nevertheless, predictive performance result differently values, due equifinality multi-modal issue combinations give occurs non-linear structure those near-linear one. Yet, former found outperform latter ones predicting timing DDR. Overall, our findings highlight importance carefully defining boundary decomposing total variance prediction errors. study expected bring new insights will help better inform users about these factors involved Nonetheless, each change depending specific situation. Details how continuous improvement important.