Seamless climate information for climate extremes through merging of forecasts across monthly to multi-year timescales: User application
MuhammadAdnanAbid1,2✉Emailadnan.abid@physics.ox.ac.uk
BeenaBalanSarojini1,2
AntjeWeisheimer1,2,3
MuhammadAdnan4
AbidAOPP4
1Oceanic and Planetary Physics (AOPP), Department of PhysicsUniversity of OxfordOxfordU.K
2
A
A
National Centre for Atmospheric Science (NCAS)U.K3European Centre for Medium-Range Weather Forecasts (ECMWF)ReadingU.K
4Department of PhysicsUniv. of OxfordOX1 3PUOxfordU.K
Muhammad Adnan Abid1,2, Beena Balan Sarojini1,2, Antje Weisheimer1,2,3
1) Atmospheric, Oceanic and Planetary Physics (AOPP), Department of Physics, University of Oxford, Oxford, U.K.
2) National Centre for Atmospheric Science (NCAS), U.K.
3) European Centre for Medium-Range Weather Forecasts (ECMWF), Reading, U.K.
Submitted to
Environmental Research Letters
September 2025
Authors Orcid:
1- Muhammad Adnan Abid (https://orcid.org/0000-0002-7389-7977)
2- Beena Balan Sarojini (https://orcid.org/0000-0002-3267-6369)
3- Antje Weisheimer (https://orcid.org/0000-0002-7231-6974)
Corresponding Author:
Muhammad Adnan Abid
AOPP, Department of Physics,
Univ. of Oxford, Oxford,
OX1 3PU, U.K.
Email: adnan.abid@physics.ox.ac.uk
ORCID: https://orcid.org/0000-0002-7389-7977
Abstract
There is a clear need of seamless climate information, which can provide forecast about the tailored climate extremes for agriculture sector on monthly multi-year lead timescales. Seamless forecast framework has been proposed to provide climate information to the climate user. In recent years, vineyard farmers in Raimat, Catalonia, have reported frost risk during spring months (March and April). Observed Frequency of Frost Days (FFD) over Catalonia shows high interannual variability for the period 1983–2022. The European Centre for Medium-Range Weather Forecasts (ECMWF) seasonal forecasting system (SEAS5), shows a varying level of FFD forecast skill during March-April season for monthly (month-1) to multi-year (month-22) lead timescales. A new temporal merging technique is developed to build a stable and seamless climate information of the tailored climate extremes on multi-year timescales. Merged forecast tends to enhance 5 to 7% of the forecast skill compared to individual start dates, added through large ensemble members which bring information from the earlier start dates, at no additional cost to the seamless system. A Ready-Steady-Action/Go framework is developed, for climate users about the potential frost risk on a monthly to multi-year lead timescales. Moreover, we found frost risk over Catalonia is predictable about 10 months prior to the target (March-April) season. FFD forecast skill over Catalonia is mainly contributed by long-term warming trends as noted for all start dates, while for the May start date (10-month lead) 60% of the forecast skill is attributed to climate variability.
Keywords:
Climate Information
Seamless forecast
Frost risk
ASPECT
Europe
Vineyard
1. Introduction
Forecasting tailored climate extreme anomalies across different timescales remains a challenge for the climate forecast models. Different societal sectors including agriculture, health, finance, and energy, are increasingly exposed to the adverse impact of the climate extremes. These sectors require skilful forecast at lead times ranging from monthly to multi-year to support effective risk management. Whilst widespread, reliable forecasts are not yet operationally available, climate information based on dynamical forecasting systems is useful to develop a dialogue with key climate user sectors, to understand their needs and develop climate forecasts (Doblas-Reyes et al., 2024; Solaraju-Murali et al., 2022). Building this climate information in engagement with climate users (e.g., agriculture, health, finance etc.), will help to cope the impacts of the tailored climate extremes for a climate resilient society (Ruti et al., 2020). Forecasts are available on different timescales, but a gap is identified in making these forecasts useful for emerging societal needs spanning across various timescales, particularly multi-year to seasonal timescale. Novel methodologies are required to combine forecasts and facilitate seamless climate information for the super-users. Seamless climate information across different timescales is mainly developed using a single model, which tends to minimise model uncertainties (Palmer et al., 2008). Therefore, in this study, European Centre for Medium-Range Weather Forecasts (ECMWF) multi-year and seasonal forecasts systems are combined to develop a new seamless system that will provide climate information to the climate forecast users, as of their requirements, on monthly to multi-year lead times.
Seasonal to decadal forecasts are publicly available to provide climate anomalies forecasts up to typically a few months to 10-year timescales respectively, while climate projections are available on centennial timescales. Decadal forecast fills the gap between seasonal forecast and the climate projections. In recent years, different temporal merging methodologies are developed to combine the climate information across decadal to climate projection timescales by constraining climate projections onto the decadal forecast to provide a skilful climate information on multi-decadal (~ 10–50 years) timescales (Borchert et al., 2019; Brunner et al., 2020; Mahmood et al., 2022). For example, sub-selection of ensemble members from climate projections, whose trajectory matches well with decadal forecast provides a skilful forecast of the climate anomalies on decadal to centennial timescales (Donat et al., 2024; Mahmood et al., 2021). Moreover, constraining the ensemble spread distribution of decadal forecasts tends to identify consistencies with the climate projections (Befort et al., 2020, 2022). However, there is a lack of work on combining monthly to multi-year lead timescales, which is addressd in this study.
In recent years, Europe has experienced numerous climate extremes, such as fires, extreme flooding in Italy, droughts in Spain, flash flooding in Germany, and frost in France and Spain. Frost risk in southern France has affected the vineyard industry, incurring losses of billions of euros (Lamichhane, 2021). Europe contributes significant proportion to the global vineyard industry, because of the favourable climatic conditions, where 75% of European vineyard lies in Spain, Italy and France. The spring season is critical for the vineyard industry, where young grapevines sprout in March and April, so any temperature fluctuation within this period may affect their sprouting and thus their yield (Tomczyk et al., 2020). The vineyards in the southwest of Europe, particularly in Raimat, Catalonia, Spain, where local farming community reported losses due to the frost risk during spring months and require seamless climate information on monthly to multi-year to lead times. To address this problem, a new temporal merging technique is used to forecast frost risk during the spring (March-April) season over Spain, with a particular focus over Catalonia, seamlessly from monthly to multi-year lead timescales.
The new multi-year forecasts along with the seasonal forecasts are used to develop seamless climate forecast information for climate extremes in the present study. The extended seasonal (or multi-year) forecast are available as 13-month and 24-month long forecasts using different start dates from the ECMWF seasonal forecasting system, SEAS5. In the present study, we propose a novel framework that combines forecast from monthly to multi-year lead times, providing seamless climate information across multi-year timescales. This climate information across multi-year timescale will address the vineyard farmers requirements, by proposing a Ready-Steady-Action/Go framework (Goddard et al., 2014). The details of the temporal merging methodology including the datasets are discussed in section 2, while results are in section 3, and summary is presented in section 4.
2. Materials and Methods
2.1 Reanalysis and re-forecast datasets
The daily minimum temperature (tmin) available at 0.25 x 0.25-degree spatial resolution from the fifth generation European Reanalysis (ERA5) is used to estimate the Frequency of Frost Days (FFD) during the spring season for the period 1983–2022 (Hersbach et al., 2020). Spring frost frequency is estimated if tmin< 0°C, then that particular grid is identified as a frost day (Albert M.G. Klein Tank et al., 2009). FFD is defined as the accumulated frost days for each spring month (i.e., March to May) at each grid point. Observational uncertainty is estimated by using the high resolution Ensemble daily gridded observation (E-OBS) minimum temperature dataset at 0.1 x 0.1-degree spatial resolution for the period 1983–2022. E-OBS is derived from station networks from the European National Meteorological station data networks across Europe (Cornes et al., 2018).
The minimum temperature are obtained from ECMWF’s SEAS5 to estimate the forecast quality of the FFD for the period 1981–2022 (Johnson et al., 2019). The SEAS5 forecasting system is based on the Integrated Forecast System (IFS) atmospheric model coupled with the ocean model, NEMO3.4. The SEAS5 atmospheric model resolution is Tco319 (cubic octahedral grid with a spectral truncation of 319), with 91 vertical levels and a model top at 0.1 hPa. The hindcast is available for the period 1981–2016, while forecast is available from 2017–2022.
Here, we use new ECMWF high resolution multi-year (24-months) forecast simulations (ECMWF, 2025). These are similar to low resolution simulations (Weisheimer et al., 2022) but use same initialization and horizontal resolution (ocean and atmosphere) as SEAS5. FFD are defined for each forecast start date for each ensemble member. The anomaly of each forecast start date is estimated based on the hindcast period 1983–2016. The SEAS5 dataset is re-gridded to 0.25 x 0.25-degree spatial resolution to match ERA5 spatial resolution.
Forecasts is available for different start dates (shown in different colour bars) at different lead times with ensemble members ranging from 15 to 51, shown in Fig. 1(a). The forecast start dates are represented on x-axis, where M-represents month, and numeric is the lead month for the target season.
2.2 Temporal Merging Method
A new temporal merging method is developed for a seamless prediction system across all available start dates. Eq. (1) is used to pool ensemble members across all start dates.
In Eq. (1) “i” represents forecast from each start date, “K” is the total number of start dates available from all start dates, “
represents weight assigned to each forecast, which is considered equal (i.e., 1) for each forecast date,
is the forecast available spatially with ensemble members “m”. The choice of weights is not straight forward and this could lead to the overfitting issues. Moreover, this also depends on the use case, where seamless system can be modified accordingly. In the present study, we found using equal weights (i.e., W = 1) is appropriate for producing seamless forecasts of the FFD. Figure 1(b) shows the trajectories of FFD forecasted for the target (March-April) season in year 2024 from all available ensemble members across different start dates (shown in different colours) from monthly (M-0) to multi-year (M-22) lead timescales.
One of the advantages of pooling ensemble members across all start dates, is the increased number of ensemble members. This tends to minimize internal variance and provides a more steady forecast skill for the target season.
2.3 Forecast Skill and Added value analysis
Forecast skill of the FFD is estimated by calculating the anomaly correlation between ensemble mean and observations/reanalysis anomaly index. A cosine latitude weighted area-averaged FFD index over the Catalonia region (0:3.5°E; 41:43°N) is estimated for ERA5 as well as for each ensemble member for all available forecast start dates. Then for each start date, the ensemble mean anomaly across all members is estimated and correlated with ERA5 to analyze the forecast skill of FFD over the Catalonia region.
The Continuous Ranked Probability Skill (CRPS) metric is used for the forecast verification (Hersbach, 2000). CRPS is a robust measure for the ensemble forecast verification, which estimates the probabilistic error based on the cumulative distribution function (CDF) of the forecast to the CDF of the observation, where smaller values represent less error and vice-versa. CRPSS at each merged forecast is compared with each individual start date as stated in Eq. (2).
Here, merged step is defined for the concatenated forecast for start date “i” and “i + 1”. For example, at the first merged step (MS1), the forecast is based on i start date concatenated with the forecast of i + 1-start date, and in the next merged step (MS2), forecast from i, i + 1, i + 2 are concatenated and so on. Here CRPSmerged step means CRPS estimated based on merged data using one or more start dates, while CRPSi represents the CRPS estimated for each individual start date (represented with “i” subscript).
A positive value indicates that the forecast based on merged data is better than the individual start date forecast, and vice-versa. The higher the CRPSS, the more value added compared to the individual forecast.
In order to assess the significance of temporal merging, we estimate the difference between the forecast skill based on merged data and that from a particular start date using the following equation.
Furthermore, anomaly forecast skill uncertainty is measured using the bootstrap replacement method, where 40 years time series is randomly sampled for 10,000 times (Befort et al., 2021; Wright et al., 2024). Shading represents 5–95% confidence interval based on the random samples.
3. Results and Discussion
3.1 Mean Frequency of the Frost days
Figure 2(a) shows the spatial mean number of FFD for March-April over Europe, based on ERA5. In north and northeast Europe, higher FFD values are observed compared to the southwestern Europe and Mediterranean region during the March-April season. In Spain, south of the Pyrenees mountain ranges, the FFD index over the Catalonia region (shown in black box) is defined for March and April, as shown in Fig. 2(b). Figure 2(b) shows high interannual variability in FFD over the Catalonia region. Most of the observed FFD occur during the month of March, and then in April, while only a few are observed in May (not shown). We find FFD over sub-domain covering Raimat within Catalonia are significantly correlated (0.93, statistically significant at 95% confidence level) with the larger Catalonia domain. Furthermore, FFD over the Catalonia region based on E-OBS, show a similar pattern to that of ERA5 but with a larger amplitude. The correlation coefficient (CC) between ERA5 and E-OBS is 0.95 (statistically significant at 95% confidence level), which shows the coherence and consistency of FFD over Catalonia among the different datasets. Furthermore, a constant bias of about 10 days is noted in mean FFD for all start dates in SEAS5 compared with ERA5 (not shown). The positive FFD bias over Catalonia could be related to the colder bias in minimum temperature in SEAS5 compared to ERA5. The mean bias is removed from each ensemble member prior to the calculation of forecast skill, which is discussed in next section.
3.2 Forecast Skill from monthly to multi-year start dates
Next, we analyse the bias corrected forecast skill of FFD during March-April for each start date, from the March start date (M-0) to the May start date 22 months prior to the target season (M-22). The anomaly of each ensemble member is calculated (shown in solid circle) and then the ensemble mean is estimated (solid line), and compared with the ERA5 anomalies (black solid line) in Fig. 3. Large spread is noted for FFD across all ensemble members for all start dates compared to ERA5. Also, SEAS5 ensemble mean peak FFD years for March start date (i.e., Lead M-0) resembles well with ERA5, while underestimated for other start dates.
Figure 4(a) shows the forecast skill of FFD for March-April at each start date, with the number of ensemble members for each start date on the secondary y-axis. Notable variation among FFD forecast skill is found from lead month 22 (M-22) to lead month 1 (M-1). Figure 4(a) identifies the start date that can potentially provide forecast to the vineyard users in the Catalonia region. In this case, we have identified that the forecast skill of the May (M-10) start date is higher compared to any other start date forecast, suggesting that it could potentially be used for the early warning (the “Ready” stage) of the frost risk to the vineyard users in the Catalonia region. Moreover, January start dates (M-2) show no forecast skill for FFD during the March-April target season, where interannual variability of FFD forecasts suggests that it does not matches well with ERA5 (Fig. 3h). For March start dates (M-0), a maximum forecast skill (0.73) is noted. This is mainly due to the initialization, and information at this scale is mainly related to forecasting weather systems. Furthermore, observational uncertainty is also estimated by replacing ERA5 with E-OBS. Overall, the forecast skill using E-OBS is comparable to that of ERA5 (Figure not shown), but with weaker skill magnitudes for most of start dates except for May start date (M-10), where it matches well with the ERA5 based forecast skill. Irrespective of observational data uncertainties, the May start date (M-10) shows a coherent prominent skill of FFD for the March-April season.
3.3 Forecast Skill based on temporal merging
Eq. (1) is adopted to develop a seamless forecast system utilising all available start dates from monthly to multi-year timescales with equal weights for each start date. We have chosen an optimal criterion, i.e., equal weighting, to combine the forecast across different start dates (Knutti et al., 2017). Figure 4(b) shows the forecast skill of FFD for the March-April season over the Catalonia region based on the temporally merged dataset, which shows stable forecast skill from lead month 16 to lead month 1 start dates. Pooling ensemble members across all start dates provides a large ensemble size (shown on secondary y-axis) to forecast the target season. This tends to reduce errors in the merged forecasts compared to individual start dates forecast, as shown in Fig. 5(a), providing positive feedback to stabilize the forecast skill. However, for the January start date, the forecast skill of the merged dataset slightly drops. We attribute this to the higher root-mean-square-errors (RMSE) noted for January and February, that have a negative feedback onto the skill (Fig. 4b).
Figure 4(c) shows the impact of the larger ensemble size onto the forecast skill of FFD, where 15 ensemble members are randomly selected from each merged step. A subsample of 15 ensemble members are selected randomly from the total members of the merged data. The forecast skill (shown in orange line) based on 15 randomly selected subsample of the ensemble members is estimated using 10,000 bootstrapped samples, where shading represents the 5–95% uncertainty range. The forecast skill of the 15 ensemble members is lower (shown in orange) than that of the total ensemble forecast skill available for each merged step (shown in blue). This demonstrates that a larger ensemble size has a positive impact onto the forecast skill consistent with earlier finding (Buizza & Palmer, 1998).
Figure 4(d) quantifies the value added by temporally merged forecasts compared with individual start dates (dashed black line), as well as the effect of ensemble size (dashed gray line). Mostly, a gain in the forecast skill of FFD is noted in the merged dataset compared to each start date. It is important to mention that the monthly to multi-year lead time forecast system provides average statistics for the March-April season, but not day-to-day weather, which is filtered out after averaging across the seasonal (in this case about 60 days) mean (Weisheimer & Palmer, 2014). For most of the start dates, the gain in the merged forecast skill range between 0 to 0.32, with maximum noted for January start date. Notably, the January start date exhibit negative forecast skill, indicating no predictive skill. In such cases, it also demonstrates that when individual start date lacks skill, then merged forecast with prior information can be beneficial. Also, an increase is noted in the forecast skill with larger ensemble size. Therefore, the seamless climate information application with larger ensembles provides stable and skilful forecast at extended lead times compared to the individual start dates.
3.4 Comparison of Forecast Skill between Temporally Merged data vs individual start dates
Next, we estimated CRPSS by calculating CRPS for each merged step and for each individual start date. This yields a CRPS score, discussed in Eq. (2). Figure 5(b) represents the added value of the merged forecast compared to individual start dates. The positive value indicates that the merged forecast is better than that of the individual start date forecast. Mainly, the forecast skill for the merged data adds value compared to the individual start date forecast, with positive values for all start dates, where higher contributions are noted for August and January (about 10%) start dates, while for the Nov (M-4) start date a minimum contribution is noted in the merged forecast.
3.5 Forecast Skill: Trends vs Detrended
Global warming trends contribute significantly to the seasonal forecast skill (Patterson et al., 2024). The trends in FFD are estimated for ERA5 and forecasts available from different start dates. In ERA5, an overall negative trend (-0.4 days/decade) in FFD is noted. Trends in the forecasts match well with the observations. Linear trends are removed from ERA5 as well as from each ensemble member. The forecast skill is re-estimated and compared with the raw (trended data) forecast skill.
Figure 6 shows that long-term warming trends contribute strongly to the forecast skill (shown in light red) of FFD during the March-April season for most of the start dates, except for the May (M-10) start dates. For the May start date (M-10), about 40% of FFD forecast skill during March-April season is attributed to the long-term global warming trend, with the remaining 60% being due to internal natural climate variability. Moreover, for March start date (M-0), a negligible difference is noted for the detrended (0.70) versus trended (0.73) forecast skill. This shows most of the forecast skill at lead-0 is determined by the interannual variability signal while long lead forecast is mainly determined by the long-term trends.
4. Summary
A dialogue has been established with the vineyard industry in the Catalonia region, where they require frost risk information on monthly to multi-year lead times for planning and mitigating the climate risk. A marginal forecast skill is noted for the frost risk across which is a challenge for any forecasting system to skilfully forecast climate extreme anomalies at localised regional scale, and this can be addressed with higher spatial resolution models (Stevens et al., 2024).
A new framework is proposed using a temporal merging technique to seamlessly forecast tailored climate extremes on monthly to multi-year lead timescales using ECMWF’s seasonal forecasting system SEAS5. We noted forecast skill of FFD for most start dates is driven primarily by long-term global warming trends. For the May start date (10-months prior to the target season, March-April), the FFD forecast skill arises from a combination of long-term trends and internal climate variability, while other start dates, the forecast skill is mainly attributed to the long-term trends. Ensemble members are pooled together across all available start dates. For example, we use a long lead time (i.e., Nov (M-16) start date) to forecast frost risk; and this information is updated, when a new start date is available (i.e., May (M-10)) an so on, as we approach the target seaso. The updated forecast is merged with the prior forecast start date (i.e., Nov (M-16)) to produce a new set of forecasts for the target season. In this way, members are pooled from all start dates until reaching the final start date before the target season. All start dates are pooled with equal weights to seamlessly forecast FFD risk during March-April (target) season from monthly to multi-year lead timescales. The weights can be made modified as of requirement, where they can be assigned to the individual forecast based on the sources of the forecast skill information, which could help to optimize the signal-to-noise ratio. Moreover, pooling ensemble members across all start dates provides a large ensemble members, that tends to reduce errors and positively feedbacks onto the forecast skill. We excluded the March (M-0) start date in temporal merging, where a maximum forecast skill (0.73) is noted due to initialization and interannual variability signal. However, this will be used for final actions along with multi-year merged forecast.
Temporal merging provides stable seamless forecasts from monthly to multi-year timescales for frost risk during the spring (March-April) season (Fig. 4b). It tends to reduce errors and provides a stable forecast across all start dates. Furthermore, selective sampling of ensemble members from merged forecast can enhance the forecast skill. The ensemble members from different start dates are correlated with observations. For most of the start dates, 1 to 3 ensemble members are identified that fulfils 95% statistically significant criteria, except for Jan (M-2) and Feb (M-1) start dates, where no members are found, while for merged data 5% (11/232) of the total ensemble members resembles with the observations, and shows 0.37 FFD forecast skill. This forecast skill is 85% higher than of the total merged ensemble forecast skill. A “Ready-Steady-Action/Go” strategy is proposed to the climate users, as shown in Fig. 7. Long lead time forecasts warn farmers to be ready for possible risk, which is followed by continuous risk monitoring in the steady stage as new forecasts become available. Finally, preventive actions can be deployed in the Action/Go stage, based on the merged forecasts together with shorter lead-time weather forecasts. Furthermore, this method shows continuous monitoring and iterative updating, help to recover the forecast skill in large ensembles, thereby reinforcing confidence in the potential hazard forecast at the extended lead times.
The seamless climate information provides a Ready-Steady-and-Action/Go strategy to the vineyard farmers on monthly to multi-year lead times. It is important to note that there could be other factors, which may affect vineyards, but these are beyond the scope of the current study (Mindlin et al., 2024). In the future, this temporal merging methodology could be applied in a multi-model framework. Also, it could be tuned (e.g., by adjusting weights and analysing the skill of other extremes) to build seamless climate information for other sectors such as health, governance and finance.
A
Acknowledgments
MAA, BBS and AW are supported by the UK Research and Innovation (UKRI) under the UK government’s Horizon Europe project (ASPECT) under grant No. [101081460] and National Centre for Atmospheric Science (NCAS). ASPECT is coordinated by the Barcelona Supercomputing Centre (BSC) and we thank Albert Soret, Marta Terrado, Veronica Torralba and Carlos D. Torre for leading the vineyard superuser case interactions. We also thank Magdalena Alonso Balmaseda (ECMWF) for producing the 24-months SEAS5 integrations. We further thank David Sexton (UK Met Office) and Dan Befort (ECMWF) for useful discussions.
5.
Data Access statement
European Reanalysis fifth generation (ERA5) dataset was obtained from Copernicus (https://climate.copernicus.eu/climate-reanalysis). E-OBS dataset was obtained from following link (https://www.ecad.eu/download/ensembles/download.php). 24-month multi-year forecast, 13-month forecast and seasonal forecast dataset from ECMWF was obtained from MARS (https://apps.ecmwf.int/mars-catalogue).
A
Competing InterestThe authors declare no competing interests.
A
Funding Information
This work is conducted under the Horizon Europe project (ASPECT) under grant No. 101081460.
A
Authors Contribution
AW developed the initial concept and secured the funding. MAA lead this work including data collection, formal analysis, original manuscript and revised draft. BBS and AW contributed to the discussion of original and revised manuscript.
References
Albert MG, Klein Tank FW, Zwiers, Xuebin Zhang (2009) &. Guidelines on Analysis of extremes in a changing climate in support of informed decisions for adaptation
Befort DJ, Brunner L, Borchert LF, O’Reilly CH, Mignot J, Ballinger AP, Hegerl GC, Murphy JM, Weisheimer A (2022) Combination of Decadal Predictions and Climate Projections in Time: Challenges and Potential Solutions. Geophys Res Lett 49(15):e2022GL098568. https://doi.org/https://doi.org/10.1029/2022GL098568
Befort DJ, O’Reilly CH, Weisheimer A (2020) Constraining Projections Using Decadal Predictions. Geophys Res Lett 47(18). https://doi.org/https://doi.org/10.1029/2020GL087900. e2020GL087900
Befort DJ, O’Reilly CH, Weisheimer A (2021) Representing Model Uncertainty in Multiannual Predictions. Geophys Res Lett 48(5). https://doi.org/https://doi.org/10.1029/2020GL090059. e2020GL090059
Borchert LF, Pohlmann H, Baehr J, Neddermann N-C, Suarez-Gutierrez L, Müller WA (2019) Decadal Predictions of the Probability of Occurrence for Warm Summer Temperature Extremes. Geophys Res Lett 46(23):14042–14051. https://doi.org/https://doi.org/10.1029/2019GL085385
Brunner L, McSweeney C, Ballinger AP, Befort DJ, Benassi M, Booth B, Coppola E, de Vries H, Harris G, Hegerl GC, Knutti R, Lenderink G, Lowe J, Nogherotto R, O’Reilly C, Qasmi S, Ribes A, Stocchi P, Undorf S (2020) Comparing Methods to Constrain Future European Climate Projections Using a Consistent Framework. J Clim 33(20):8671–8692. https://doi.org/https://doi.org/10.1175/JCLI-D-19-0953.1
Buizza R, Palmer TN (1998) Impact of Ensemble Size on Ensemble Prediction. Mon Weather Rev 126(9):2503–2518. https://doi.org/10.1175/1520-0493(1998)126%3C2503:IOESOE%3E2.0.CO;2
Cornes RC, van der Schrier G, van den Besselaar EJM, Jones PD (2018) An Ensemble Version of the E-OBS Temperature and Precipitation Data Sets. J Geophys Research: Atmos 123(17):9391–9409. https://doi.org/https://doi.org/10.1029/2017JD028200
Doblas-Reyes FJ, St Clair L, Baldissera Pacchetti A, Checchia M, Cortekar P, Klostermann J, Krauß JEM, Muñoz W, Mysiak ÁG, Paz J, Terrado J, Villwock M, Volarev A, M., Zorita S (2024) Standardisation of equitable climate services by supporting a community of practice. Clim Serv 36:100520. https://doi.org/https://doi.org/10.1016/j.cliser.2024.100520
Donat MG, Mahmood R, Cos P, Ortega P, Doblas-Reyes F (2024) Improving the forecast quality of near-term climate projections by constraining internal variability based on decadal predictions and observations. Environ Research: Clim 3(3):035013. https://doi.org/10.1088/2752-5295/ad5463
ECMWF (2025) 24-months SEAS5 reforecast dataset. https://doi.org/https://doi.org/10.21957/5e60-rf82
Goddard L, Baethgen WE, Bhojwani H, Robertson AW (2014) The International Research Institute for Climate & Society: why, what and how. Earth Perspect 1(1):10. https://doi.org/10.1186/2194-6434-1-10
Hersbach H (2000) Decomposition of the Continuous Ranked Probability Score for Ensemble Prediction Systems. Weather Forecast 15(5):559–570. https://doi.org/10.1175/1520-0434(2000)015%3C0559:DOTCRP%3E2.0.CO;2
Hersbach H, Bell B, Berrisford P, Hirahara S, Horányi A, Muñoz-Sabater J, Nicolas J, Peubey C, Radu R, Schepers D, Simmons A, Soci C, Abdalla S, Abellan X, Balsamo G, Bechtold P, Biavati G, Bidlot J, Bonavita M, Thépaut JN (2020) The ERA5 global reanalysis. Q J R Meteorol Soc 146(730):1999–2049. https://doi.org/10.1002/QJ.3803
Johnson SJ, Stockdale TN, Ferranti L, Balmaseda MA, Molteni F, Magnusson L, Tietsche S, Decremer D, Weisheimer A, Balsamo G, Keeley SPE, Mogensen K, Zuo H, Monge-Sanz BM (2019) SEAS5: The new ECMWF seasonal forecast system. Geosci Model Dev 12(3):1087–1117. https://doi.org/10.5194/gmd-12-1087-2019
Knutti R, Sedláček J, Sanderson BM, Lorenz R, Fischer EM, Eyring V (2017) A climate model projection weighting scheme accounting for performance and interdependence. Geophys Res Lett 44(4):1909–1918. https://doi.org/https://doi.org/10.1002/2016GL072012
Lamichhane JR (2021) Rising risks of late-spring frosts in a changing climate. Nat Clim Change 11(7):554–555. https://doi.org/10.1038/s41558-021-01090-x
Mahmood R, Donat MG, Ortega P, Doblas-Reyes FJ, Delgado-Torres C, Samsó M, Bretonnière P-A (2022) Constraining low-frequency variability in climate projections to predict climate on decadal to multi-decadal timescales -- a poor man’s initialized prediction system. Earth Sys Dyn 13(4):1437–1450. https://doi.org/10.5194/esd-13-1437-2022
Mahmood R, Donat MG, Ortega P, Doblas-Reyes FJ, Ruprich-Robert Y (2021) Constraining Decadal Variability Yields Skillful Projections of Near-Term Climate Change. Geophys Res Lett 48(24). https://doi.org/https://doi.org/10.1029/2021GL094915. e2021GL094915
Mindlin J, Vera CS, Shepherd TG, Doblas-Reyes FJ, Gonzalez-Reviriego N, Osman M, Terrado M (2024) Assessment of plausible changes in Climatic Impact-Drivers relevant for the viticulture sector: A storyline approach with a climate service perspective. Clim Serv 34:100480. https://doi.org/https://doi.org/10.1016/j.cliser.2024.100480
Palmer TN, Doblas-Reyes FJ, Weisheimer A, Rodwell MJ (2008) Toward seamless prediction: Calibration of climate change projections using seasonal forecasts. Bull Am Meteorol Soc 89(4):459–470. https://doi.org/10.1175/BAMS-89-4-459
Patterson M, Befort DJ, O’Reilly CH, Weisheimer A (2024) Drivers of the ECMWF SEAS5 seasonal forecast for the hot and dry European summer of 2022. Q J R Meteorol Soc 150(765):4969–4986. https://doi.org/https://doi.org/10.1002/qj.4851
Ruti PM, Tarasova O, Keller JH, Carmichael G, Hov Ø, Jones SC, Terblanche D, Anderson-Lefale C, Barros AP, Bauer P, Bouchet V, Brasseur G, Brunet G, DeCola P, Dike V, Kane MD, Gan C, Gurney KR, Hamburg S, Yamaji M (2020) Advancing Research for Seamless Earth System Prediction. Bull Am Meteorol Soc 101(1):E23–E35. https://doi.org/https://doi.org/10.1175/BAMS-D-17-0302.1
Solaraju-Murali B, Bojovic D, Gonzalez-Reviriego N, Nicodemou A, Terrado M, Caron L-P, Doblas-Reyes FJ (2022) How decadal predictions entered the climate services arena: an example from the agriculture sector. Clim Serv 27:100303. https://doi.org/https://doi.org/10.1016/j.cliser.2022.100303
Stevens B, Adami S, Ali T, Anzt H, Aslan Z, Attinger S, Bäck J, Baehr J, Bauer P, Bernier N, Bishop B, Bockelmann H, Bony S, Brasseur G, Bresch DN, Breyer S, Brunet G, Buttigieg PL, Cao J, Ziemen F (2024) Earth Virtualization Engines (EVE). In Earth System Science Data. Copernicus Publications 16(4):2113–2122. https://doi.org/10.5194/essd-16-2113-2024
Tomczyk AM, Szyga-Pluta K, Bednorz E (2020) Occurrence and synoptic background of strong and very strong frost in spring and autumn in Central Europe. Int J Biometeorol 64(1):59–70. https://doi.org/10.1007/s00484-019-01793-z
Weisheimer A, Balmaseda MA, Stockdale TN, Mayer M, Sharmila S, Hendon H, Alves O (2022) Variability of ENSO Forecast Skill in 2-Year Global Reforecasts Over the 20th Century. Geophys Res Lett 49(10). https://doi.org/https://doi.org/10.1029/2022GL097885. e2022GL097885
Weisheimer A, Palmer TN (2014) On the reliability of seasonal climate forecasts. J Royal Soc Interface 11(96):20131162. https://doi.org/10.1098/rsif.2013.1162
Wright MJ, Weisheimer A, Woollings T (2024) Multi-Decadal Skill Variability in Predicting the Spatial Patterns of ENSO Events. Geophys Res Lett 51(12). https://doi.org/https://doi.org/10.1029/2023GL107971. e2023GL107971
Figure List
Figure 1
a) Monthly to Multi-year forecast simulation ensemble size (shown as bars, with colours denoting forecasts of differing length) from ECMWF Seasonal forecast system 5 (SEAS5), for the period 1981–2022. b) an example of Frequency of Frost Days (FFD) forecast trajectories for spring (March-April) 2024. Each line represents ensemble members from different start dates (colours represent different start dates). The non-linear x-axis represents different start dates.
Figure 2: a) Spatial map of the Frequency of Frost Days (FFD) using ERA5 for the period 1983–2022 (Units: days). White region represents masked mountainous areas, The red box is the focus of the study, i.e., Catalonia region [0:3.5 °E, 41:43 °N]; focus of the study; b) Area-averaged FFD over shown in black box over northeast Spain for March (red), April (blue), and Spring (March-April mean; black line) season using ERA5 for the period 1983–2022. The solid grey line represents FFD for March-April season based on the E-OBS dataset over the same region. Correlation between E-OBS and ERA5 is 0.95.
Figure 3: Anomalous Frequency of Frost Days (FFD, units: days) based on each start date from Month-22 to Month-0 lead time for the target Spring (March-April) season compared with ERA5 FFD anomalies for the period 1983–2022. FFD for each ensemble members are on primary y-axis, while ERA5 FFD anomaly (shown in black colour) and the ensemble mean anomaly are shown on secondary y-axis. Prediction skill is defined as the anomaly correlation between ensemble mean for each start date and ERA5 of March-April FFD is shown at top for each start date figure.
Figure 4: a) Forecast Skill of FFD (black line on primary y-axis) during Spring (March-April) season based on anomaly correlation for each start date for the period 1983–2022; while ensemble size for each start dates is on secondary y-axis; b) Forecast Skill of temporally Merged data (blue line on primary y-axis), while ensemble size for merged data is shown on secondary y-axis; c) Comparison of forecast skill based on fixed randomly selected 15-ensemble members from temporally merged data (shown in orange line) compared with the merged data (shown in blue line); d) Change in forecast skill of FFDs between merged and each start date (black dashed line) while effect of large ensemble member size is shown in (grey dashed line). Shading (a-c) shows uncertainty range within 5–95% confidence level.
Figure 5: a) Root Mean Square Error (RMSE) of forecast for each start date (red) compared to the merged data (blue); b) Continuous Ranked Probability Skill Score (CRPSS) of merged data compared with each start date (grey bars). Horizontal dashed grey line is mean of the CRPSS for all merged steps. Secondary x-axis represents merged steps (MS), where MS1 means merged forecast based on Nov (M-16) and May (M-10), and so on, while MS8 means merged forecast based on all available start dates.
A
Fig. 6
Comparison of FFD Forecast skill during March-April using trended and detrended forecasts from each start date. Shading represents the uncertainty, ranges between 5–95% confidence levels. Light red shading represents uncertainty range of the forecast including trends, while grey shading represents the forecast based on detrended dataset.
A
Fig. 7
Framework for the seamless climate information based on climate user’s input as of their requirement, Ready-Steady-Action/Go. (following; Goddard et al. 2004).
A
Fig. 1
a) Monthly to Multi-year forecast simulation ensemble size (shown as bars, with colours denoting forecasts of differing length) from ECMWF Seasonal forecast system 5 (SEAS5), for the period 1981–2022. b) an example of Frequency of Frost Days (FFD) forecast trajectories for spring (March-April) 2024. Each line represents ensemble members from different start dates (colours represent different start dates). The non-linear x-axis represents different start dates.
A
Fig. 2
a) Spatial map of the Frequency of Frost Days (FFD) using ERA5 for the period 1983–2022 (Units: days). White region represents masked mountainous areas, The red box is the focus of the study, i.e., Catalonia region [0:3.5 °E, 41:43 °N]; focus of the study; b) Area-averaged FFD over shown in black box over northeast Spain for March (red), April (blue), and Spring (March-April mean; black line) season using ERA5 for the period 1983–2022. The solid grey line represents FFD for March-April season based on the E-OBS dataset over the same region. Correlation between E-OBS and ERA5 is 0.95.
A
Fig. 3
Anomalous Frequency of Frost Days (FFD, units: days) based on each start date from Month-22 to Month-0 lead time for the target Spring (March-April) season compared with ERA5 FFD anomalies for the period 1983–2022. FFD for each ensemble members are on primary y-axis, while ERA5 FFD anomaly (shown in black colour) and the ensemble mean anomaly are shown on secondary y-axis. Prediction skill is defined as the anomaly correlation between ensemble mean for each start date and ERA5 of March-April FFD is shown at top for each start date figure.
A
Fig. 4
a) Forecast Skill of FFD (black line on primary y-axis) during Spring (March-April) season based on anomaly correlation for each start date for the period 1983–2022; while ensemble size for each start dates is on secondary y-axis; b) Forecast Skill of temporally Merged data (blue line on primary y-axis), while ensemble size for merged data is shown on secondary y-axis; c) Comparison of forecast skill based on fixed randomly selected 15-ensemble members from temporally merged data (shown in orange line) compared with the merged data (shown in blue line); d) Change in forecast skill of FFDs between merged and each start date (black dashed line) while effect of large ensemble member size is shown in (grey dashed line). Shading (a-c) shows uncertainty range within 5–95% confidence level.
A
Fig. 5
a) Root Mean Square Error (RMSE) of forecast for each start date (red) compared to the merged data (blue); b) Continuous Ranked Probability Skill Score (CRPSS) of merged data compared with each start date (grey bars). Horizontal dashed grey line is mean of the CRPSS for all merged steps. Secondary x-axis represents merged steps (MS), where MS1 means merged forecast based on Nov (M-16) and May (M-10), and so on, while MS8 means merged forecast based on all available start dates.
Figure 6
Comparison of FFD Forecast skill during March-April using trended and detrended forecasts from each start date. Shading represents the uncertainty, ranges between 5–95% confidence levels. Light red shading represents uncertainty range of the forecast including trends, while grey shading represents the forecast based on detrended dataset.
Figure 7
Framework for the seamless climate information based on climate user’s input as of their requirement, Ready-Steady-Action/Go. (following; Goddard et al. 2004).
