Seasonal climate forecast can inform the European agricultural sector well in advance of harvesting

Agricultural production faces increasing exposure to climate extremes (heat stress, drought, excessive rainfall) that can cause major yield and quality losses. Seasonal climate predictions, offering lead times up to a year, could inform critical farm decisions (sowing, variety selection, fertilization, field operations, irrigation) and support market and policy responses. However, challenges remain due to limited skill in key regions like Europe, scale dependencies, and the interpretation of probabilistic uncertainty. This study assesses the skill and reliability of the ECMWF SEAS5 seasonal forecasting system in predicting agro-climate indicators relevant to European winter wheat, using the co-designed Clisagri service. It provides a spatial assessment of prediction performance across the growing season, evaluates predictability of different event types (e.g., drought, excessive wetness, temperature stress), and highlights opportunities linked to increased lead time of skillful predictions.

Prior work has shown seasonal forecasts can add value for agricultural decision-making across various regions, yet Europe often exhibits lower skill and strong dependencies on event scale and timing. Interpreting probabilistic forecasts remains a barrier to uptake. Land-surface initialization can improve seasonal skill for crop forecasts, and large-scale teleconnections (NAO, atmospheric blocking, ENSO-related indices) are known to influence European climate and yields. Persistence properties of hydrological balance anomalies underpin drought predictability, while threshold-based wetness and some temperature extremes are less predictable at seasonal scales. These insights inform the selection and evaluation of agro-climate indicators used here.

Data: ECMWF SEAS5 seasonal forecasts (via Copernicus Climate Change Service) with reforecasts for 1993–2019, 25-member ensembles, six initializations (November, February, March, April, May, June), each with 7-month lead time. Variables: daily minimum and maximum temperature and precipitation. Forecasts (native ~36 km atmospheric resolution) were upscaled to 1° by C3S. Bias adjustment was performed via quantile mapping using the MarsMet reference dataset (25 km, ~4000 stations across Europe), upscaled to SEAS5 resolution for adjustment. Agro-climate indicators: Indicators from the Clisagri service tailored to winter wheat, computed for sensitive phenological stages (germination to maturity). A dynamic phenology model, calibrated with European field and observational data for current varieties, simulates stage timing per grid cell and year. Drought is quantified using the Standardized Precipitation Evapotranspiration Index (SPEI) computed non-parametrically over stage-specific windows (e.g., sowing–leaf development, tillering, stem elongation, booting–heading, heading–maturity, and entire season). Excessive wetness indicators count days with precipitation above defined thresholds during specific stages; temperature stress indicators count days below/above thresholds (cold/heat) during key stages. Thresholds and indicator definitions were co-designed with farmers. Merging observed and forecast data: For each initialization, observed MarsMet weather is used up to the forecast start date, then merged with bias-adjusted SEAS5 for the remainder of each indicator window. November runs cover events up to June (cannot extend to maturity in most regions due to 7-month lead). Skill and reliability assessment: Indicators are categorized into terciles for evaluation: flowering timing (early/normal/late); hydrological balance (SPEI < −0.84, −0.84 to 0.84, > 0.84); and event counts (three categories based on impact thresholds). Skill was quantified using the Fair Ranked Probability Skill Score (FRPSS) against climatological reference (ensemble-size corrected). Reliability was assessed via reliability diagrams using bootstrap resampling (1000 iterations) and categorized as perfect, very useful, marginally useful, not useful, or dangerously useless. Spatial summaries include the share of arable land with FRPSS > 0 within predefined European regions (Iberia, Italy, France, South-Eastern, Central, Northern, UK & Ireland, Eastern).

- Flowering time prediction: Significant skill from November in central and eastern Europe and Turkey; lack of skill across much of western Europe. Reliability is useful for early and late flowering categories in central/eastern Europe; the normal category is generally least reliable. Skill and reliability increase with later initializations (Feb–May), achieving widespread skill by April–May. - Drought (SPEI) indicators: With November initialization, SPEI for sowing–leaf development is largely skillful across Europe, especially where the window includes observed data. SPEI for tillering and stem elongation shows lower skill early; skill increases with Feb+ initializations and when part of the window is observed. SPEI between heading and maturity (indicator 5) is only predictable from February runs onward; UK & Ireland show dangerously useless reliability in Mar–May despite Feb skill. The full-season SPEI (indicator 6) shows significant skill and reliability already from February when roughly half the season remains to be forecast. - Excessive wetness: Indicators based on days above precipitation thresholds (notably indicators 12 and 13) show no or very limited skill and poor reliability across regions; indicator 8 (rain >10 mm during tillering) improves with later initializations but remains limited. - Temperature stress: Cold stress (Tmin < 2 °C between booting and flowering) is reliably predicted in UK & Ireland from February. Heat stress indicators (Tmax > 31 °C during booting–flowering; Tmax > 35 °C during flowering–maturity) show significant skill/reliability mainly in Iberia, Italy, France, South-Eastern, and Eastern Europe; little to no skill in Northern Europe and UK & Ireland. - 2018 case study (heading–maturity SPEI): The severe spring–summer drought in central/northern Europe and concurrent wetness in southern Europe was detected only by the June initialization; earlier runs failed to capture it. June runs increased skill in southern Europe but remained low/no skill in parts of central, eastern, northern Europe, and UK & Ireland where SPEI relied heavily on forecast data. - Continental extent of impacts: Roughly a quarter of European arable land is affected annually by drought or wet conditions during heading–maturity; drought share exceeded 20% in 2006, 2007, 2015, 2018, and 2019. Drought-affected areas are generally predicted more accurately than wetness-affected areas, with June initializations achieving the highest European-level reliability.

Seasonal forecasts can inform European wheat decision-making, notably enabling early-season choices (variety selection, management planning) via skillful flowering predictions in central/eastern Europe from November. Predictability depends on both initialization time and event type. SPEI-based hydrological balance predictability generally improves as the season progresses, benefiting from integrated observed conditions and the persistence of anomalies; the full-season hydrological balance (indicator 6) is already skillful by late winter/early spring. Conversely, threshold-based wetness indicators lack seasonal predictability, suggesting a need to exploit large-scale drivers (teleconnections, circulation regimes) and higher resolution to improve skill. Drought during heading–maturity is critical but shows limited early-year predictability in western and parts of central Europe, likely due to limited impact of initial conditions and spring regime transitions, including biases in jet position and strength affecting blocking and Atlantic Low regimes. In central/eastern Europe, stronger dependence on initial conditions and positive correlations between late-spring and summer SPEI suggest soil moisture memory enhances predictability. Practical implications span farm-level operations to policy (market stabilization, planning, risk responses).

This study demonstrates that ECMWF SEAS5 seasonal forecasts, translated into co-designed Clisagri agro-climate indicators, can provide actionable information for European winter wheat producers. Flowering time is predictably and reliably forecast early in the season over central/eastern Europe, and drought-related hydrological balance indicators attain useful skill from late winter, while excessive wetness indicators remain largely unpredictable at seasonal scales. The work highlights the value of integrating observed conditions and exploiting persistence in hydrological anomalies, and it identifies regional and event-type limitations. Future research should evaluate higher-resolution and multi-system large-ensemble forecasts, develop alternative methods leveraging large-scale atmospheric/oceanic patterns and teleconnections to improve extremes prediction (e.g., tropical North Atlantic SST influences), and continue co-design with end users. Advancements in high-resolution Earth system simulations may further enhance seasonal predictability of agriculturally relevant extremes.

- Limited or no seasonal predictability for excessive wetness indicators across Europe, constraining decision usefulness for wet-related risks. - Regional skill gaps persist, notably in western and parts of central Europe for drought during heading–maturity, due in part to limited influence of initial conditions and spring circulation regime predictability. - November initializations cannot cover maturity-stage indicators due to 7-month lead limitations in the available C3S product. - Indicator impacts can vary with local environmental factors (e.g., soil properties), which are not explicitly resolved in the predictability assessment. - Skill partly reflects incorporation of observed conditions prior to initialization, and predictability may be sensitive to bias-adjustment and merging procedures. - Wetness-affected area prediction generally underperforms drought-affected area prediction at the European scale.