Last modified on May 07, 2025 08:30

Issued by: CSIC

Date: 15/12/2023 (updated on 15 Apr 2025 )

Service contract: 2022/C3S2_381_Lot1_Contractor/SC1

1. Introduction 

This document describes the "Copernicus Interactive Climate Atlas gridded monthly dataset" (C3S Atlas Dataset in short) which constitutes an evolution of the IPCC AR6 Atlas gridded monthly dataset, also published in the CDS¹. This dataset has been entirely produced using climate information fully available in C3S CDS and a processing workflow available from the GitHub C3S Atlas repository² (including the underlying software and illustrative notebooks). The dataset underpins the different products visualised in the Copernicus Interactive Climate Atlas Application (C3S Atlas Application in short), available at https://atlas.climate.copernicus.eu

The dataset available in the CDS at https://cds.climate.copernicus.eu/datasets/multi-origin-c3s-atlas is always the latest version, which is version 2 (v2) at the moment. Information on changes with respect to previous versions is available at Sections 3 and 6 of this document.  

In addition to describing the dataset, we summarise the fit-for-purpose quality control and curation applied, based on insights from basic products (e.g., maps and climate stripes), and outline the archiving metadata and standards employed. 

2. CDS catalogues, variables and harmonization

2.1. CDS catalogues and datasets used

The Copernicus Interactive Climate Atlas gridded monthly dataset integrates information from a range of climatic observations, reanalysis products and projection datasets. Table 1 lists the datasets and the corresponding CDS catalogues used (datasets included in the latest version – v2 –  are highlighted in red). Version v2 includes three new observational datasets: NOAA Climate Prediction Center dataset (CPC) for precipitation, Berkeley Earth (BERKEARTH in short) for air temperatures, and the European Space Agency (ESA) SST Climate Change Initiative (SST-CCI) project for sea surface temperature. In addition, version v2 includes the regional reanalyses CERRA (adding extra details on ERA5 with the use of a high resolution limited area model for downscaling including data assimilation), an additional high-resolution source of information to assess recent changes over Europe. 

Note that the CORDEX catalogue comprises two distinct datasets with different horizontal spatial resolution³: (1) CORDEX-CORE, a homogeneous subset of CORE⁴ simulations at 0.22° and (2) CORDEX-EUROPE, covering Europe at a finer 0.11° resolution. To harmonise the spatial grids across datasets within the Atlas, a set of common nested regular sub-grids (2°, 1°, 0.5°, 0.25°, 0.125°, and  0.0625°) were used to (conservatively) interpolate some of the original datasets, particularly multi-model ensembles. However, certain products, such as ERA5-Land and ERA5⁵, are retained on their original regular grids provided by the CDS to preserve their native resolution. The reference grids used for the various datasets, along with additional auxiliary information, are available in the GitHub C3S Atlas repository⁶.

Some of the datasets available in the CDS are distributed across multiple catalogues, depending on temporal resolutions (hourly, daily, or monthly) or subsets of variables (e.g. single-level or multi-level). To produce the Atlas dataset, we selected the most appropriate catalogue for each variable or index (see Table 2), guided by the principle that primary variables and their derived indices (e.g. maximum temperature and number of days with maximum temperature over 35°C) should originate from the same source catalogue. For example, maximum temperature (TX) is needed at daily resolution to compute derived indices; therefore, for ERA5, the hourly catalogue (Reanalysis-era5-single-levels) was  used for this and other related variables. In contrast, variables such as sea surface temperature, which do not require high temporal resolution, were sourced from the monthly catalogue (Reanalysis-era5-land-monthly-means). Table 1 uses colour coding to indicate the temporal resolution of the original datasets in the CDS catalogues.

Table 1: Description of the projection, reanalyses and observational datasets and CDS catalogue entries used in the C3S Atlas Dataset. Colour codes indicate the different temporal resolutions (hourly, daily, and monthly, one or several for each catalogue). The column “horizontal resolution” indicates the different grids used (products regridded to the common regular sub-grids: 2°, 1°, 0.5°, 0.25°, 0.125°, 0.0625° are indicated with an asterisk). Note that the periods of the Atlas dataset correspond to full years. BERKEARTH stands for Berkeley Earth Foundation dataset, CPC for NOAA Climate Prediction Center datasets, and SST-CCI for European Space Agency (ESA) SST Climate Change Initiative (CCI) project. Red colour indicates new elements in version 2 (latest version) of the dataset (v2). 

Data harmonisation across the various datasets and catalogues is critical aspect of this data preparation work. For example, ERA5 and ERA5-Land do not provide daily data directly, requiring daily values to be derived from hourly data. However, both datasets use different conventions for encoding hourly data: ERA5 provides hourly accumulated values, while ERA5-Land provides accumulations from the start of the forecast (00 UTC) over a 24-hour period. As a result, different post-processing steps are needed to compute  consistent daily values⁷. These differences must be carefully addressed during harmonisation to ensure that all variables share a consistent definition, temporal aggregation method, and unit standardisation (see Section 2.3).

2.2. Variables and indices

Table 2 shows the set of 35 variables included in the dataset, comprising both primary variables (highlighted in bold) and derived indices (see Table 4 for details). This expands the 21 climate variables and indices featured in the original IPCC AR6 Atlas gridded monthly dataset. These variables are organised in different Climatic Impact Driver (CID) categories: heat and cold, wet and dry, drought, wind and radiation, snow and ice, ocean, and circulation. 

Table 2: Description of the climate variables and indices included in the second version of the C3S Atlas (v2), with new additions and changes relative to version 1 (v1) highlighted in red. The first column indicates the Climatic Impact Drivers (CID) categories: heat and cold, wet and dry, drought, wind and radiation, snow and ice, ocean, and circulation. The “bias adjustment (BA)” column indicates the bias adjustment methods applied to each index, either simple linear scaling (“S”), the ISIMIP method (“I”), or both. A bias correction selection widget has been added to the CDS form, and in the files this is indicated with 'bals' and 'baisimip' being appended to the variable code. Primary variables are shown in bold, while derived indices appear in regular font (see Table 4 for details). The column "CDS Names" indicates the names used  in the CDS catalogue of the C3S Atlas Dataset and further details on thresholds and temporal aggregation are provided in the "Description" column. 

Table 3 lists the variables from Table 2 that are computed across the different datasets referenced in Table 1. Note that bias-adjusted indices are calculated only for climate projection datasets (currently CMIP5/6, CORDEX-CORE and CORDEX-EUR-11) and not for reanalysis or observational datasets. 

Table 3. Description of variables/indices (in rows) calculated for the different datasets (in columns). Data not available is indicated by ‘-’ and data not considered due to insufficient data (only available for a reduced subset) is indicated by (*). The table indicates the availability for the two versions of the SPI and SPEI indices (see Section 2.3 for details). 

2.3. Index calculation

Most variables and indices included in the C3S Atlas Dataset, defined in Table 2, are derived from straightforward temporal aggregations or threshold-based computations. However, several more complex cases are described below.

Cooling degree-days (CD) and Heating degree-days (HD) have been implemented following the definition used in the IPCC AR6 Atlas (IPCC, 2021: Annex VI: Climatic Impact-driver and Extreme Indices).

The Standardised Precipitation Index (SPI-6) is a monthly index that compares 6-month accumulated precipitation against the long-term distribution for the same location and accumulation period, as the number of standard deviations from the median. Similarly, the Standardised Precipitation Evapotranspiration Index (SPEI-6) compares the 6-month accumulation of precipitation minus potential evapotranspiration (PET) to its long-term distribution. In version 2 of the dataset, the definition of PET (and consequently the SPEI index) has been updated to use the Hargreaves formulation (Hargreaves, 1994), replacing the Thornthwaite method used in v1. These indices were computed using using the Python-based xclim software. The updated PET component is provided alongside the SPEI index in the new version of the dataset.

Both SPI and SPEI involve standardisation using a specific reference period to compute long-term distributions. The standard reference period used is 1971-2010 for observations, reanalyses, and projections. In addition, the full 1971-2100 period is used for projections as an alternative reference period providing both versions in the dataset. In the C3S Atlas viewer the combination of the two normalisations are used (see more details in the C3S Atlas User Guide). Both approaches, using a historical reference periods and pooling historical and future scenario data to define a reference are well established practices in the literature to analyse future projections (e.g. Thota et al. 2025 and Spinoni et al. 2020, respectively). Therefore, the Atlas dataset provides both so they can be used for different purposes.

Full details of the processing workflow (including index calculation) are available in the GitHub C3S Atlas repository which includes the software and notebooks documenting the entire workflow, ensuring reproducibility and facilitating reuse.

2.4. Harmonisation of variables across catalogues

Table 4 details the specific CDS variables from the catalogues displayed in Table 1 that were used to construct the variables and indices presented in Table 2. If focuses only on the primary climate variables (excluding derived indices) which are highlighted in bold in Table 2. These raw variables originate from different catalogues, each with varying temporal resolutions, units and definitions⁸. The table also outlines the harmonisation and conversion processes applied to ensure consistent definitions and units across the dataset. The first column provides the variable code along with its standard name and units. The second column lists the CDS variable names and describes the harmonisation procedures applied across datasets to obtain standardised variables and units. Colour coding reflects the temporal aggregation (hourly, daily and monthly) of the catalogue from which each variable was sourced. N/C indicates variables that were "Not Considered" for some particular reason. 

In addition to the name and unit standardisation described in Table 2, the harmonisation pre-processing also includes calendar conversion to a standard calendar⁹ and the use of a consistent coordinate system (longitude range from -180° to 180°). Furthermore, all variables and indices derived from the source datasets are interpolated onto common grids to produce the final harmonised dataset.

Table 4: Harmonisation work across the different catalogues carried out for the variables of the dataset. Colour codes indicate the temporal resolutions used for the different datasets available in the CDS catalogues, which correspond to the colours indicated in Table 1 (green for hourly, blue for daily and purple for monthly). Unit harmonisation (when needed) is indicated with "==>".

2.5. Available ensembles for climate projection products

The global and regional climate projection datasets listed in Table 1 are based on multi-model ensembles that include simulations from both historical and future scenarios (RCPs or SSPs). To compute climate change signals, simulations from both time periods are required. As a result, the Atlas dataset includes only those models that provide data for the historical period and at least one future scenario. Figures 1, 2 and 3 show the final ensembles used for each dataset (CMIP6, CMIP5, and CORDEX-EUR-11) organised by variable (shown in columns). Asterisks indicate variables used to compute derived indices (see Table 4), which share the same model availability. Note that the CORDEX CDS catalogue was previously used in the IPCC Atlas gridded monthly dataset to create regional datasets over the continental domains covered by CORDEX, at the common 0.44° resolution (not included in the current version). This ensemble is described in the documentation accompanying the corresponding IPCC AR6 Atlas entry in the CDS catalogue. The C3S Atlas Dataset includes the datasets with highest resolution available from CORDEX, in particular the 0.11º version of the European regional dataset (CORDEX-EUR-11) described here and a worldwide 0.22º mosaic produced with CORDEX-CORE simulations described in Section 2.5.

Figure 1: Final CMIP6 ensemble used in the C3S Atlas dataset, showing the available models (rows) across the different variables (columns). Colours represent the historical and future SSP scenarios available in the CDS.

Figure 3: Final CMIP5 ensemble used in the C3S Atlas dataset, showing the available models (rows) across the different variables (columns). Colours represent the historical and future RCP scenarios available in the CDS.

Figure 2: Final CORDEX-EUR-11 ensemble used in the C3S Atlas dataset, showing the available models (rows) across the different variables (columns). Colours represent the historical and future RCP scenarios available in the CDS. 

Note that the CDS catalogue for CMIP6 includes only a single realisation per model. In contrast, for CMIP5, multiple realisations are included in some cases. Figure 3 displays the CMIP5 realizations used in the C3S Atlas Application, typically one per model to ensure consistency across all projection datasets. However, for models that provided multiple realisations used in CORDEX simulations, all such realisations have been included. The full C3S Atlas Dataset, as available in the CDS, includes all realisations for each model.

2.6. Mosaic approach for CORDEX-CORE

One of the main novelties of the dataset is related to the CORDEX-CORE project¹⁰. CORDEX-CORE is a highly ambitious initiative aimed at providing regional climate projections with global coverage (see Figure 4) and high resolution (0.25°x0.25°) by combining of 2 RCMs nested to 5 GCMs selected to span the widest range of uncertainty (see Table 6). The dataset includes, for the first time, the CORDEX-CORE simulations spatially blended using the mosaic approach described in Diez-Sierra et al. (2022). This dataset addresses two main issues related to the CORDEX-CORE simulations: (1) avoiding domain selection in overlapping areas by choosing the domain that best fits each one of the IPCC AR6 reference regions and (2) avoiding the multiple native projections of the different domains by providing the data in a regular global mesh. Due to its global coverage and higher resolution, this dataset constitutes the main source of information to analyze regional climate change globally, so it is a strategic dataset for the C3S Atlas.

The two RCMs used in CORDEX-CORE are REMO2015 and RegCM4; the three GCMs used are MOHC_HadGEM2-ES, MPI-M_MPI-ESM-MR (with the LR version used in some domains), and NCC_NorESM1-M (sorted in high to low climate sensitivity); note that MIROC-MIROC5 and NOAA-GFDL-GFDL-ESM2M are used as backups for high and low sensitivity, respectively. The list of simulations is given in Table 5, which describes the matrix of precise model versions used in the different domains. Figure 5 shows the final ensembles computed for the different variables (in columns) for the CORDEX-CORE dataset.

Figure 4: Spatial coverage of the CORDEX-CORE simulations (displaying mean temperature change for a 3º global warming relative to 1981-2010). Image from the C3S Atlas Application (https://atlas.climate.copernicus.eu)

Table 5: CORDEX simulations for the different domains used to build the CORDEX-CORE dataset (note that EUR-11 simulations are upscaled to the global 0.22 grid). Empty cells indicate that the default versions for the two RCMs (RegCM4_7 and REMO2015) are used. Different versions and/or backup GCMs are indicated in the table.

Figure 5: Final CORDEX-CORE ensemble used in the C3S Atlas Dataset, showing the available models (rows) across the different variables (columns). Colours represent the historical and future RCP scenarios available in the CDS (see Table 5 for the specific configuration of the global mosaic).

2.7. Bias adjustment

Bias adjustment is a recommended processing step for threshold-related indices, as highlighted in the IPCC AR6 (Cross-Chapter Box 10.2: Relevance and Limitations of Bias Adjustment). Numerous methods have been proposed and compared in the literature, each with distinct advantages and limitations (e.g. Casanueva, 2022). For threshold-based temperature indices, simple linear scaling is often an effective choice; by adjusting the mean, it generally yields satisfactory results when compared to more complex approaches such as ISIMIP3. In the C3S Atlas Dataset, both raw and bias-adjusted versions of threshold-dependent indices are provided, following IPCC AR6 best practices. The ISIMIP3 method is applied for both temperature and precipitation indices, while the simple linear scaling method is also applied to temperature indices to highlight differences between basic and more advanced methods, and allowing users to explore this source of uncertainty.

All calculations are done using the ibicus package (see Table 4 for details) which implements a variety of bias adjustment methods (including linear scaling and ISIMIP3 among others). 

Table 4. Functions and parameters of the Python Ibicus package (https://ibicus.readthedocs.io) used in the C3S Atlas for bias adjusted variables. Note that the ISIMIP method requires different parameters for bias adjusting temperature and precipitation. 

Full details of the processing workflow (including bias adjustment) are available in the GitHub C3S Atlas repository which includes the software and notebooks documenting the entire workflow, ensuring reproducibility and facilitating reuse.

3. Technical information and archiving

3.1. File format and archiving

Files have been generated using the NetCDF-C library version 4.4.1.1 and HDF5 library version 1.10.1, using the NETCDF4 data model. The resulting files in NetCDF format follow the CF 1.9 metadata convention, which supports string-type NetCDF variables (used to define some attributes in the files, such as members). The attribute convention for data discovery is compliant with ACDD-1.3, including the Coordinate Reference System (CRS) and other descriptive elements.
Data is stored in separate files for each experiment, scenario, and variable/index, with all members included in the same file, identified by the member attribute, and covering the full time periods listed in Table 1. A compression level of 1 (DeflateLevel = 1) is used  when creating NetCDF files to provide a good balance between compression and read/write performance.

This, as well as the new indices, resulted in 2.1 TB with a total of 1057 files for the present version (v2), as compared with the 1.3 TB and 412 files of the first version (v1).

Several technical issues were identified in the first version of the C3S Atlas Dataset and were corrected in the latest (present) version:

• Problems with _FillValue and missing_value attributes: These two attributes are used to identify gridboxes with no information and are mapped to NaN/missing-values by standard software frameworks. The latter is used to specify a code for NaN values (missing data), where the dataset does not provide a numeric value (e.g. regions not covered by the dataset, removed outliers, etc.). The former is used to identify the gridboxes that should be filled during the generation of the archive (either with a value or with a NaN) and is used to flag potential problems in the writing process. In some cases, datasets are generated using the same value for both attributes. In the C3S Atlas Dataset we intentionally use different values for the "_FillValue" and "missing_value" properties to ensure that all data has been written to the file, including values representing "missing_value" where data is missing and needs to be interpreted by the user's application (e.g. SST on land-points or data outside the CORDEX domain). This caused issues for the WPS subsetting system (which is based on Python and uses xarray), as xarray cannot handle multiple types of missing values when saving a file, although reading such files poses no problem. To resolve this, the WPS adopted a solution of merging the two missing value types and retaining a single 'missing_value' when saving files.
• The use of character coordinate variables in the C3S Atlas Dataset, specifically the character-based ensemble attribute used to describe the models in the ensemble, led to incompatibilities with certain versions of netCDF4, particularly regarding compression options. However, this issue has been addressed in recent versions of the libnetcdf library. 
• Originally a high compression level (level 9) was used to achieve maximum data reduction. However, this significantly slowed down data access. After some discussion, an analysis was carried out to find a better balance between file size and performance, and it was found that a compression level of 1 offers a more optimal trade-off between compression and read/write speed. We adopted this change for the present version of the C3S Atlas Dataset (v2). 
• CDO cannot directly process the C3S Atlas dataset due to issues with the ensemble attribute. Specifically, CDO has a legacy limitation requiring 'time' to be the first dimension, which requires pre-processing before applying CDO operators. Today, it is widely accepted that users must pre- and post-process data when working with CDO, rather than feeding it raw files. It is also important to note that CDO is designed to handle ensemble members distributed across separate files; placing “time” as the first dimension prevents the use of CDO functions tailored for ensembles.

3.2. Spatial subsetting

The CDS form (and API) allows users to spatially subset the data for download by selecting any latitude–longitude region. If the selected area exceeds the dataset’s spatial coverage, the full dataset will be returned. If the area overlaps partially, only the data points within the covered region will be provided. However, requests that do not intersect the dataset’s region at all will fail.

The BERKEARTH, CPC, SST-SCI, ERA5, ERA5-Land, CMIP5 and CMIP6 datasets cover the entire Globe, while E-OBS, CERRA and CORDEX-EUR-11 are for Europe (see Figure 7) and CORDEX-CORE is a mosaic covering mostly the continental parts of the World (see Figure 4).

Figure 7: Illustrations of the extent of the E-OBS (left), CORDEX-EUR-11 (center), and CERRA (right) European domains.

In the longer term, we intend to introduce temporal subsetting too easing the selection of the time period and allowing to choose shorter time periods. 

4. Fit-for-purpose quality control 

Some basic quality control procedures were implemented to check the consistency of the data produced for the different indices and variables across datasets. This section describes the tests implemented to analyze and fix issues in the dataset. 

4.1. Tests implemented

A comprehensive quality control process was conducted on all variables and indices to identify potential issues in both the original data (e.g., units, coordinates) and throughout the index generation workflow. This process enabled the detection of metadata errors and data inconsistencies during data harmonisation, and fixing or excluding the involved simulations. The quality control involved two qualitative tests:

• Spatial maps: These were used to identify spatial misalignments relative to reference features, such as coastlines (see Figure 8). This analysis was particularly useful in spotting interpolation errors and coordinate issues in the raw data. Only a few problems were found in the original datasets. Notably, the IPSL-CM5A2-INCA_r1i1p1f1 simulation showed irregular native longitude and latitude coordinates for sst and siconc, and was therefore excluded from the dataset. In addition, mosaics of different ensemble members were used to assess ensemble dispersion. This analysis helped identify anomalies in the spread or spatial patterns of climate projections. Notably, the simulation EUR11_MPI-M_MPI-ESM-LR_r1i1p1_IPSL_WRF381P_v1 was found to be a clear outlier, showing physically inconsistent values (when compared with the rest of members) across most variables. As a result, it was excluded from the dataset.

Figure 8: Spatial map for near surface air temperature (t) for Euro-CORDEX.

• Climate Stripe Plots: Stripe plots were used to detect various issues, including simulations that deviate significantly from the ensemble range for particular years or contain extended periods of missing data. These plots are generated by aggregating variables annually over the entire domain—globally for CMIP and regionally for CORDEX. Figure 9 presents a stripe plot for near-surface air temperature (tas) from CMIP6 under the SSP5-8.5 scenario. Each row represents a different model in the ensemble (see Section 2.4), and each column corresponds to a year, showing spatially aggregated annual means. This analysis identified several models with prolonged data gaps that were excluded from the dataset. Specifically, CNRM-CM6-1_r1i1p1f2 (for variable sst under scenario ssp370) and IITM-ESM_r1i1p1f1 (for sst under all scenarios, and all variables under ssp370). Note that in many cases, the ensemble shows large uncertainty, with a large spread with low- and high-end models. The qualitative inspection of the stripe plots was used to detect outliers (models with values which are significantly lower/higher than the ensemble spread, indicating wrong units, or similar problems). Although the spread was high in some cases, the outliers were physically consistent (with the exception of EUR11_MPI-M_MPI-ESM-LR_r1i1p1_IPSL_WRF381P_v1, already mentioned) and, therefore, they were not removed from the dataset.

Figure 9: Example of climate stripe plot for CMIP6 showing annual mean values (in columns) for the different GCMs forming the ensemble (in rows). Note that different members exhibit different temperature magnitudes (i.e. systematic biases if they were compared to an observational dataset).

5. Known issues 

No issues reported.

6. Versions of the C3S Atlas Dataset 

The following table details the different versions of the C3S Atlas Dataset (and dates, in chronological order), with main changes and new elements. Note that in the CDS only the latest version (v2 at the moment) is available. 

The latest version (v2) introduced several modifications to some variables/datasets already existing in v1. They are summarised below in Table 5.

Table 5. Description of changes introduced in the latest version of the C3S Atlas (v2) for variables and datasets already existing in the first version (v1). 

7. Versions of the document 

The table below outlines the changes made to this document, listed in chronological order. Document versions follow the Vn.m format, where n refers to the dataset version and m denotes the document version within that version cycle.

8. Related articles

• Page:
  CORDEX: Regional climate projections
• Page:
  CMIP6: Global climate projections
• Page:
  CMIP6: Decadal climate predictions
• Page:
  CMIP: Global climate projections
• Page:
  C3S Climate projections