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

Abstract

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