Seasonal climate forecasts show skill in predicting winter chill for specialty crops in California

California produces about three-quarters of U.S. fruits and nuts, with many species (e.g., walnut, pistachio, cherry, pear, plum) requiring substantial winter chill for dormancy release and normal phenology. Insufficient chill leads to delayed and uneven flowering, poor fruit quality, extended harvests, reduced vigor, and lower yields. Prior studies project significant future declines in chill, and insufficient chill has already been observed in California’s Central Valley. Interannual forecasts of chill anomalies could aid risk management, but teleconnection-based predictors (ONI, PNA, PDO) explain less than ~16% of interannual chill variability in California, limiting their utility. The study explores an alternative: using seasonal temperature forecasts from dynamic climate models to predict NDJF (November–February) chill accumulation at varying lead times. The hypothesis is that while models have limited skill in daily temperature variability, they may better predict season-aggregate chill. Objectives are to (1) assess the potential of global seasonal forecasts to predict NDJF chill anomalies and crop-specific chill sufficiency at different lead times in California, and (2) gauge California farmers’ concern regarding declining chill accumulation.

Previous work indicates declining chill under climate change and observed insufficiencies in California’s Central Valley. Teleconnection indices (ONI, PNA, PDO) account for a small fraction (<16%) of chill variability (1979–2019) in major growing regions, suggesting limited predictive power for interannual chill. Studies of seasonal forecast temperature skill over California (e.g., NMME) report modest ACCs (~0.1–0.6) at zero to one-month leads, with ensemble means outperforming individual models and rapid skill decay with lead time. Dynamic models show limited temperature predictability at subseasonal-to-seasonal scales in California, but season-aggregate metrics like chill might still be predictable despite poor day-to-day temperature prediction. Additional literature highlights model biases related to coastal cloudiness, topography, and circulation anomalies, and notes growers’ reliance on Chilling Hours while the Dynamic Model’s Chill Portions often better reflect chill under warming conditions.

Study area and crops: The analysis covers California with a focus on ecoregions and top counties where walnut, pistachio, cherry, plum, and pear are cultivated. Crop masks were derived from the 30 m USDA-NASS Cropland Data Layer (2022) aggregated to the 4 km PRISM grid to restrict evaluation to crop-growing areas. Data: Reference daily Tmax/Tmin from PRISM (4 km) for 1982–2018. Monthly hindcasts of temperature for 1993–2016 from five seasonal systems: NCEP CFSv2, CCCma CanCM4, ECMWF SEAS5, CMCC SPSv3, and UKMO GloSea6-GC3.2, obtained from the Copernicus Climate Data Store or IRI (for CanCM4). For each model, ensemble members were averaged to form a model mean; models were equally weighted to form a multimodel mean. Lead-time configurations: NDJF season forecasts assessed at: 1-month lead (initialized monthly and aggregated across NDJF), 1–4 months (initialized Nov 1), 2–5 months (Oct 1), and 3–6 months (Sep 1). Because NDJF spans two calendar years, NDJF of year Y combines ND (Y) with JF (Y+1). Downscaling and daily/hourly construction: Model monthly Tmax/Tmin anomalies were downscaled to daily fields at 4 km by combining model monthly anomalies and standard deviations with PRISM daily climatology and observed monthly standard deviations, preserving PRISM’s fine-scale variability. Hourly temperatures were synthesized from daily Tmax/Tmin, latitude, and solar geometry using ChillR, with a sine curve for daytime warming and a logarithmic nighttime cooling scheme. Chill computation: Two metrics were used: Chilling Hours (CH; hours in 0–7.2°C) and Chill Portions (CP; Dynamic Model via ChillR). CH and CP were computed over NDJF for each lead time from 1993–2015 (2016 excluded due to missing JF 2017). Both metrics were calculated for reference (PRISM-driven) and for each model hindcast, then aggregated spatially as needed (e.g., county averages restricted to crop masks). Evaluation metrics: Deterministic skill was quantified using anomaly correlation coefficients (ACCs) between model and reference NDJF seasonal anomalies of CP and CH; statistical significance assessed via Pearson correlation (p<0.05). Categorical skill was assessed by classifying each year’s standardized anomaly (z-score) into above-normal (>0.5), normal (−0.5 to 0.5), and below-normal (<−0.5) categories separately for models and reference, then computing percentage of correct matches across years. Crop-specific chill sufficiency: Using crop-specific CH thresholds (approximate NDJF CH requirements): walnut ~700, pistachio ~1000, plum ~900, pear ~1350, cherry ~1200 (CP thresholds variably available), forecasts were deemed correct if model and reference were on the same side (above/below) of the threshold in a given year. Percent correct across years was calculated for each crop and lead time within crop-growing regions. County-level comparison: For top-producing counties per crop (from CDL 2022), county-averaged NDJF CP and CH were compared between reference and multimodel forecasts at varying leads. ACCs were computed for county-averaged time series. Farmer survey: Concern about declining chill was extracted from a statewide Qualtrics survey (USDA NIFA project). 12,933 invitations, five reminders; responses screened for authenticity; 341 valid farmer responses analyzed. The Likert-type item asked the extent of concern about climate-related impacts, including reduced chill. Demographics and regional distributions were summarized to contextualize concern.

• Farmer concern: Among 341 surveyed farmers, 70% reported concern about decreasing chill accumulation over recent decades. Nearly half were from San Joaquin Valley; 25% from the Superior region; 95.3% produced fruits and nuts.
• ACC skill at 1-month lead (multimodel): For CP (CH), ACC>0.5 covered 82% (81%) of California, 84% (88%) of the Central Valley, and 98% (77%) of Southern California (San Diego, Los Angeles, Inland South). Across crop-growing regions, ACC>0.5 covered 84–89% (90–94%).
• Lead-time degradation: Areas with ACC>0.5 declined to 32% (35%) at 1–4 months, 19% (20%) at 2–5 months, and 12% (18%) at 3–6 months for CP (CH). Within crop regions, ACC>0.5 covered only 30–47% (21–35%), 5–15% (21–33%), and 1–12% (19–22%) for the same lead-time tiers.
• Individual model performance: At 1-month lead, ACC>0.5 areas ranged 34–81% across models; GloSea6-GC3.2, SPSv3, and SEAS5 outperformed CFSv2 and CanCM4. Beyond 1 month, areas with ACC>0.5 fell below 43% across models; CFSv2 relatively better among longer leads. Multimodel generally exceeded individual models.
• Categorical forecasts (NDJF CP/CH): At 1-month lead, correct category >50% of the time in 40% (CP) and 43% (CH) of California; 41% (CP) and 52% (CH) in the Central Valley; and 83% (CP) and 59% (CH) in Southern California. In crop regions, 42–50% (CP) and 54–57% (CH) areas exceeded 50% correctness for walnut, pistachio, cherry, and plum; pear much lower (~0.23% CP; 0.32% CH). When the category was correct, the bias in CP (CH) was <20% in 71% (78%) of cases statewide and 61% (88%) in the Central Valley.
• Crop-specific chill sufficiency (CH thresholds): Multimodel forecasts correctly predicted above/below-threshold NDJF CH more than 50% of the time across nearly all walnut-growing areas at all lead times. For pistachio, cherry, plum, and pear, the percentage of areas exceeding 50% correctness typically ranged 86–99%, 74–99%, 97–99%, and 96–99% respectively, with lead-time influence generally <16%.
• Observed vs predicted at county scale (1-month lead): Multimodel ranges frequently enveloped reference CP (and often CH) series; however, some county–crop combinations (e.g., pistachio and plum in Fresno; walnut and cherry in San Joaquin and Stanislaus) showed reference CH outside the predicted range more than half the time. County-level ACCs for CP (CH) ranged 0.4–0.8 (0.5–0.7), significant at p<0.05 for almost all crop–county pairs except pear in Lake County. Predictive spread widened and ACCs declined at longer leads.

Season-ahead information on winter chill can help growers mitigate yield and quality losses associated with insufficient dormancy fulfillment. Despite modest temperature forecast skill in California, aggregating to seasonal chill enables meaningful predictability at short leads. One-month-lead multimodel forecasts achieve substantial skill in both continuous (ACC) and categorical (above/normal/below) predictions, especially in the Central Valley and Southern California, and capture sharp low-chill years. CH forecasts are somewhat less accurate than CP, though many growers still rely on CH due to simplicity; improved CH-focused outreach may enhance adoption. Skill diminishes beyond one month, likely tied to weak precipitation predictability influencing temperature and to model limitations in representing circulation and regional mesoscale processes. Multimodel ensembles are preferred over single systems given heterogeneous performance. The forecasts add value beyond climatology by flagging anomalously low-chill seasons, supporting proactive management (e.g., rest-breaking treatments, kaolin sprays, altered pruning, irrigation for evaporative cooling, cultivar/rootstock choices). Extension of this approach to other regions requires verifying local chill requirements and evaluating local model temperature skill; daily model data could further improve performance if computationally feasible.

Seasonal forecasts from global dynamical models can skillfully predict NDJF chill in California at one-month lead. Multimodel ACCs exceed 0.5 across most of California and the Central Valley, and categorical predictions are correct more than half the time over substantial areas. Forecasts capture interannual variability and years with sharp chill declines and are particularly effective for predicting binary crop-specific chill sufficiency relative to thresholds. However, skill declines at longer leads (≥1–4 months). These findings support using short-lead multimodel forecasts to inform chill management decisions. Future work should: incorporate daily model outputs where feasible; improve regional bias correction and representation of mesoscale processes; refine crop-specific chill thresholds (including CP-based thresholds) by location; and enhance communication tools tailored to growers’ reliance on CH and operational windows.

• Forecast skill degrades rapidly beyond a one-month lead, with substantial reductions in ACC and categorical accuracy.
• Limited temperature predictability in California, partly linked to poor precipitation and circulation predictability, constrains chill forecast skill.
• Model structural and resolution limitations (topography, coastal cloudiness, vegetation, parameterizations) introduce spatially varying errors.
• Crop-specific sufficiency forecasts (threshold-based) indicate direction (above/below) but not the magnitude of deficit/surplus.
• Chill requirements are location- and cultivar-specific; thresholds may not transfer across regions without local validation.
• Analysis used monthly hindcasts transformed to daily/hourly; direct daily model outputs might improve performance but were not used due to computational constraints.
• Pear categorical performance (especially for CP) was notably low in some areas; some county–crop CH series fell outside multimodel ranges frequently.