Statistical downscaling of seasonal spring-frost predictions

Spring frost occurring after bud break of grapevines can hinder the plants’ growth, and cause large economic losses to the winemaking sector. On-field protection strategies exist at the seasonal horizon, but their cost-effectiveness depends on the availability of accurate risk predictions at an appropriate spatial resolution. In this work, we investigate the predictability of spring-frost events in the important winemaking region of Catalonia, at a high spatial resolution of 10 km, and at the seasonal horizon. This problem setup caters for the specific requirements of one of the case studies considered within the European project ASPECT (Adaptation-oriented Seamless Predictions of European ClimaTe), where researchers and end-users co-develop strategies of climate information production and delivery. To this end, we present a statistical downscaling method which leverages the predictable components of the large-scale atmospheric variability, and builds upon a robust statistical link identified in observations between spring-frost occurrence and atmospheric blocking events. The method is trained with reanalysis and observational data, and subsequently tested using 1-month lead-time retrospective predictions of blocking frequency from several operational seasonal prediction systems. The statistically-downscaled predictions are then evaluated probabilistically against standard verification benchmarks. It is found that blocking frequency variability is sufficiently well reproduced by the seasonal prediction systems over the areas important to Catalan frost, and that the skill of the downscaled predictions can improve (albeit not always with statistical significance) over the prediction references. • Spring-frost events in Catalonia are statistically associated with large-scale patterns of anomalous blocking frequency. • Seasonal predictions skilfully reproduce variability of atmospheric blocking events in March at one month lead time. • A statistical downscaling method based on large-scale patterns of blocking frequency is trained to predict frost risk at high spatial resolution. • Skill of statistically downscaled seasonal predictions of spring-frost occurrence can improve over existing prediction benchmarks.