California's agricultural sector is a global leader in fruit and nut production, with many crops (e.g., walnuts, pistachios, cherries, pears, plums) requiring significant winter chill for proper dormancy breaking. Insufficient chill leads to delayed flowering, uneven bud break, poor fruit quality, reduced yields, and impaired vegetative growth. Studies project significant future declines in chill accumulation, with impacts already observed in California's Central Valley. While previous research explored relationships between chill accumulation and climate variability indices (ONI, PNA, PDO), these indices explained only a small portion of interannual variability. This study explores an alternative approach: using interannual temperature predictions from dynamic climate models to forecast chill accumulation. The central research question is whether these models possess sufficient skill to predict winter chill accumulation (November-December-January-February, or NDJF) in California at various lead times. This is crucial for informing growers' decisions and mitigating risks associated with insufficient chill.

Prior research indicates a projected decline in chill accumulation in California's fruit-growing regions, with negative impacts already evident. Studies have explored the link between chill accumulation and large-scale climate patterns like the El Niño-Southern Oscillation (ENSO), Pacific-North American (PNA) teleconnection, and Pacific Decadal Oscillation (PDO), but these connections explained less than 16% of interannual variability. Other studies have examined the skill of dynamic climate models in predicting temperature in California, with varying results, demonstrating limited skill in predicting temperature at sub-seasonal to seasonal timescales. However, while these models may not accurately predict daily temperature variations, the study hypothesizes that they might be more effective in predicting overall seasonal chill accumulation.

The study used daily maximum (Tmax) and minimum (Tmin) temperature data from the Parameter-elevation Regressions on Independent Slopes Model (PRISM) dataset (1982-2018) as a reference. Monthly hindcasts of temperature from five global climate models (NCEP CFSv2, CCCma CanCM4, ECMWF SEAS5, CMCC SPSv3, UKMO GloSea6-GC3.2) participating in the North American Multi-Model Ensemble (NMME) and available through the Copernicus Climate Data Store were used. Hindcasts for the NDJF season were evaluated at 1-month, 1-4 month, 2-5 month, and 3-6 month lead times. The models' monthly Tmax and Tmin were downscaled to daily values using PRISM's daily climatology and monthly temperature anomalies. Chill accumulation was quantified using two metrics: Chill Hours (CH), accumulating hours between 0-7.2°C, and Chill Portions (CP), using the Dynamic Model in the ChillR package which accounts for the mitigating effects of warm temperatures on chill accumulation. The skill of the forecasts was evaluated using anomaly correlation coefficients (ACC) between model-predicted and reference chill values, and by assessing the accuracy of categorical chill forecasts (above normal, normal, below normal). The study also analyzed crop-specific chill sufficiency using CH thresholds for walnuts, pistachios, cherries, plums, and pears, determining if the model correctly predicted whether total chill would be above or below the threshold. A survey of 341 California farmers was conducted to gauge their concern regarding declining chill accumulation.

A significant majority (70%) of surveyed California farmers expressed concern about declining chill accumulation. For one-month lead forecasts, the multimodel average showed ACCs exceeding 0.5 for CP and CH in 82% and 81% of California areas, respectively. The accuracy of categorical chill forecasts exceeded 50% in 40% (CP) and 43% (CH) of California areas. The multimodel forecasts correctly predicted crop-specific chill sufficiency (whether total chill would be above or below the threshold) over 50% of the time for most of the regions, especially walnuts. The models accurately captured interannual variability in chill, particularly during years with substantial chill decreases (e.g., 1995, 2005, 2014-2015). However, forecast skill decreased significantly with longer lead times (beyond one month), showing diminished ACCs, lower accuracy in categorical forecasts, and reduced ability to capture interannual variability. While CP provided more accurate quantification than CH, the study notes that CH remains the preferred metric among many growers.

The results demonstrate the potential of using multimodel seasonal forecasts to predict winter chill one month in advance, informing growers’ decisions regarding chill management strategies. The multimodel approach proved more effective than relying on individual models, likely due to compensating for individual model biases. The stronger skill in predicting crop-specific chill sufficiency compared to continuous chill amount highlights the value of threshold-based forecasts for practical applications. The limitations of longer lead-time forecasts highlight the need for further model development to improve sub-seasonal to seasonal prediction capabilities. The lower skill in predicting CH compared to CP suggests the need for outreach and education to promote the adoption of CP, a more accurate chill measure. The study's findings are relevant to other regions and crops facing insufficient winter chill, particularly in the tropics where temperate fruits are cultivated at higher altitudes.

This study showed that multimodel seasonal forecasts offer valuable skill in predicting one-month lead winter chill forecasts in California, improving growers’ ability to manage risks associated with insufficient chill. The ability to predict crop-specific chill sufficiency thresholds was particularly strong. While longer-lead forecasts showed limited skill, the current one-month forecast can still be useful for growers. Future research should focus on improving model skill at longer lead times and on refining methods for communicating these forecasts to growers. Extending this approach to other regions and crops is also warranted.

The study's reliance on monthly hindcasts, instead of daily data, may limit the accuracy of the chill calculations, particularly for CH. The downscaling method used to obtain daily temperature values could introduce some uncertainties. The study's focus on a limited number of crops and regions may limit the generalizability of the findings. Further research is needed to improve the skill of longer lead-time forecasts, and to address the discrepancies between CH and CP, as well as between the needs of growers and the superior metrics provided by CP.