Advancing grapevine phenology predictions: a comparison of regression-based statistical learning approach and conventional models in mediterranean environment

Abstract

Climate conditions can significantly affect plant dormancy and growing season, impacting crop development and its productivity. Temperature is considered the primary driver of grapevine phenology, with endodormancy and ecodormancy largely regulated by chilling and forcing accumulations, respectively. Nevertheless, additional climatic factors, such as precipitation, radiation and atmospheric circulation patterns, may also contribute and regulate variability of different phenological phases. As hotspot of climate change, Mediterranean regions are witnessing negative compounding effects by enhanced warming and changes in precipitation patterns on crop systems and linked agro-industry. Among these, viticulture is a particularly relevant economic crop in Mediterranean regions. This research aims to assess some conventional phenological modeling approaches to project several phenological stages of grapevine in Sardinia (Italy), comparing the results to regression-based stepwise Statistical Learning that can potentially account for effect of several climate variables. Combined models (i.e. Chilling Hours-Growing Degree Hours) improve prediction of budburst phenological timing (RMSEs: 7 days) in relation to more complex models (i.e., PhenoFlex). The Growing Degree Days model was more accurate than complex phenological models in predicting blooming and veraison (RMSEs of 6 and 4 days, respectively). The regression-based statistical learning approach achieved phenological prediction performances comparable to the best of the conventional models (RMSEs: 4–7 days). Regression-based statistical learning approach was able to infer contribution of several climate variables on grapevine phenology, at a larger seasonal scale. Implementing both modeling approaches showed similar spatial patterns, except for budburst phase which indicated earlier timings over mountainous regions through regression-based statistical learning contrary to conventional model, highlighting the influence of maximum and minimum temperatures from previous year.