GCRI · Climate and Hazards · Foundations
Measuring climate baselines: methods, evidence and limits
Measuring climate baselines turns imperfect observations into reference values through completeness rules, homogenization, gridding, reanalysis and inter-comparison, and each step changes the result. This methods article explains how those steps work, what blind tests and dataset comparisons show about them and what limits users of any baseline should state.
- As of
- 28 October 2026
- पठन
- 12 minutes
- साक्ष्य
- 15 sources, cited in order
- पद्धति
- A review of WMO guidance on normals and homogenization, the documentation of major temperature datasets and reanalysis, a blind benchmark of homogenization methods and the 2025 inter-comparison of global temperature estimates, organised by step of the processing chain.
- डेटा
- Guidance comes from WMO; station collections from NOAA's National Centers for Environmental Information; dataset documentation from the Met Office, NOAA, Berkeley Earth and the Copernicus Climate Change Service; the benchmark from Climate of the Past; each is openly available from its custodian.
पद्धतियाँ और साक्ष्य
- प्रश्न
- How are climate baselines calculated from observations, how is each step tested, and what uncertainty remains?
- दृष्टिकोण
- A review of WMO guidance on normals and homogenization, the documentation of major temperature datasets and reanalysis, a blind benchmark of homogenization methods and the 2025 inter-comparison of global temperature estimates, organised by step of the processing chain.
- साक्ष्य
- 15 sources, cited in order and listed under References
- Data availability
- Guidance comes from WMO; station collections from NOAA's National Centers for Environmental Information; dataset documentation from the Met Office, NOAA, Berkeley Earth and the Copernicus Climate Change Service; the benchmark from Climate of the Past; each is openly available from its custodian.
- सीमाएँ
- The article compares published methods and values and does not recompute them; national homogenization practice varies and is not fully documented; and dataset versions and reanalysis releases change past values.
01दायरा
What measuring climate baselines involves
The methods that turn observations into reference values, and the evidence that tests them.
Measuring climate baselines is the work of turning raw observations into reference values, such as a 30-year station normal or a global average for 1850–1900, that can be compared across places and decades. It involves deciding which data are complete enough, removing changes that come from the observing system rather than the climate, filling or representing gaps, merging land and ocean records and stating the uncertainty that remains.
The concepts and stakes of baselines, including why different reference periods exist, are set out in the platform's foundational article on climate baselines, reachable from the climate baselines topic. This companion article goes one level down, into the methods and the evidence behind each step and the limits that a user of any baseline should know.
Each method answers a specific threat to comparability. Completeness rules guard against averages biased by missing data. Homogenization guards against steps from station moves and instrument changes. Gridding and interpolation guard against uneven coverage. Ensembles and inter-comparison guard against overconfidence in any single method.
02Stakes
Why the methods matter
Method choices move baselines by tenths of a degree, enough to change records and thresholds.
Method choices are not technicalities. The Intergovernmental Panel on Climate Change (IPCC) reports that methodological advances and new datasets, including a more complete representation of the Arctic, increased its estimate of global surface warming by about 0.1 degrees Celsius between assessments, and that this increase does not represent additional physical warming 5. For a target of 1.5 degrees, a tenth of a degree is a large share.
Datasets built by different teams give different values for the same year. For 2025, the Copernicus Climate Change Service reports 1.47 degrees Celsius above 1850–1900 in ERA5 and 1.39 degrees in HadCRUT5 4, NOAA reports 1.34 degrees 3 and Berkeley Earth 1.44 degrees, plus or minus 0.09 6. The World Meteorological Organization (WMO) consolidates the datasets to 1.43 degrees with an uncertainty of plus or minus 0.13 7. Whether a year is reported as above or below a threshold can depend on the dataset chosen.
Local baselines carry the same issue at higher stakes for users. A station normal that includes an uncorrected relocation can misstate the local climate, and WMO notes that the need for guidance on homogenization was re-emphasized by its members during the recent collection of climatological standard normals 8.
03प्रसंस्करण श्रृंखला
From observation to baseline: the processing chain
The steps from raw readings to a published reference value.
The chain starts with observations: daily station readings on land, ship and buoy measurements at sea, and, for reanalysis, observations of many other kinds. The Global Climate Observing System lists 55 Essential Climate Variables whose records are meant to be sustained for this purpose 9.
Quality control removes errors such as transcription mistakes and failed sensors. Completeness tests decide whether a month or a normal can be computed at all. Homogenization then detects and corrects non-climatic breaks. For global products, station anomalies are gridded, merged with sea surface data and averaged with area weights; HadCRUT5, for example, uses a 5-degree grid from January 1850 10.
Reanalysis follows a different route through the chain. ERA5 combines model data with observations from across the world to produce an hourly record on a 0.25-degree grid from 1940, updated daily with a delay of about five days 11. It fills gaps with physics rather than statistics, which gives complete fields but introduces the model's own biases.
Every step leaves a trace in the published product, so the evidence for a baseline includes its documentation: which data entered, which version was used and what was adjusted. Users comparing two baselines should check that each step was done the same way.
04Completeness
Completeness: how much data is enough
The WMO rules for missing data and why they exist.
The WMO guidelines on the calculation of climate normals set two levels of completeness. A monthly mean from daily values should not be calculated if observations are missing for 11 or more days in the month, or for 5 or more consecutive days 1. A monthly normal needs values in at least 80 per cent of the years, 24 of 30 for a standard normal, and should not be calculated if values are missing in 3 or more consecutive years 1.
The rules rest on evidence about error. The guidelines summarise a study that estimated the error of a monthly mean from missing days, and note that consecutive gaps matter more than scattered ones because weather is persistent 1. Stricter national rules are allowed, and some services continue to use older, tighter criteria 1.
Completeness rules decide what can be compared, not whether it is correct. A complete series can still contain a relocation or an instrument change, which is why homogeneity is treated as a separate step in the same guidelines 1.
05Homogenization
Homogenization: removing non-climatic breaks
How breaks are detected and corrected, and what a blind benchmark showed.
A homogeneous series varies only because of the weather and climate. In practice, stations move, instruments and shelters change, observation times shift and surroundings grow, and each change can introduce a step or a spurious trend. Homogenization detects such breaks and adjusts the series so that the whole record reflects the current observing conditions.
Relative methods dominate because a neighbour shares the regional climate but not the station's local history. NOAA's Global Historical Climatology Network monthly dataset, version 4, uses a Pairwise Homogenization Algorithm that detects abrupt non-climatic shifts through automated comparisons of a target station with correlated neighbours and derives adjustments from the change in their differences 12. Version 4 draws on about 26,000 monthly temperature stations, against about 7,200 in earlier versions, and publishes both unadjusted and adjusted data so users can see the effect 12.
Methods have been tested blind. The HOME benchmark inserted artificial breaks, outliers, gaps and local trends into simulated networks and received 25 homogenised solutions in the blind phase and 22 more after the inserted breaks were revealed; relative homogenization generally improved temperature data, only the best algorithms improved precipitation, and the best performers were modern methods designed to work with an inhomogeneous reference 2. The study also found that automatic algorithms can perform as well as manual ones and that training users is very important 2.
Practice is still uneven between services, which is why WMO's Expert Team on Climate Data Development and Stewardship began online consultations on homogenization good practice in late 2025 8.
06Inter-comparison
Comparing datasets: what the spread tells us
Stated uncertainty, structural uncertainty and why consolidated values are wider.
Two kinds of uncertainty sit in any baseline. Stated uncertainty is what a team estimates for its own method: HadCRUT5 expresses it as an ensemble of 200 realisations 10, and Berkeley Earth gives plus or minus 0.09 degrees Celsius for 2025 6. Structural uncertainty is the difference between teams that made different reasonable choices, and it can only be seen by comparing independent datasets.
The 2025 spread of 1.34 to 1.47 degrees above 1850–1900 3,4 is larger than the stated uncertainty of some single datasets, which shows why consolidation matters. The NOAA value of 1.34 degrees is NOAA's own; the Copernicus values for other centres are computed with the ERA5 offset of 0.88 degrees 3,4. WMO's range of plus or minus 0.13 degrees around its consolidated value of 1.43 degrees 7 is wide enough to include the structural spread.
Reanalysis adds a time dimension to uncertainty. The most recent months of ERA5 are a preliminary release (ERA5T), which may differ from the final data published two to three months later if serious flaws are found 11, so a baseline comparison made early in a year may move slightly when the record is finalised. Copernicus notes that its 2025 figures and associated text were updated on 28 January 2026 to reflect the full availability of global datasets 4, a reminder that annual values have publication dates as well as data periods.
Converting between baselines needs care. The Copernicus Climate Change Service uses an offset of 0.88 degrees between 1850–1900 and 1991–2020 for ERA5 and applies the same offset to other datasets when ranking 2025 4. An offset estimated from one dataset applied to another is itself a source of error, so a careful comparison states the dataset, version, period and offset together.
| चरण | Error addressed | Evidence or reference |
|---|---|---|
| Completeness tests | Bias from missing days and years | WMO guidelines on normals 1 |
| Homogenization | Steps from relocations and instrument changes | Pairwise algorithm 12; blind benchmark 2 |
| Gridding and merging | Uneven coverage of land and ocean | HadCRUT5 ensemble 10 |
| पुनर्विश्लेषण | Gaps filled with a physical model | ERA5 11 |
| Inter-comparison | Structural differences between methods | WMO consolidation 7 |
07साक्ष्य
साक्ष्य कहाँ से आता है
The collections and datasets that make baselines testable, and their custodians.
Station baselines are collected and published openly. For 1991–2020, WMO collected normals from 141 countries and territories, and NOAA's National Centers for Environmental Information publishes more than 6,500 station files, computed by countries following the WMO guidelines 13. National collections such as the United States normals add gridded fields and 15-year normals for 2006–2020 calculated with the same methods 14.
A user of any baseline can therefore report a short set of facts that make it testable: the dataset and version, the reference period, the completeness rule applied, whether the data were homogenized and by which method, any offset used to convert between periods and the uncertainty stated by the producer. Each of these is documented by the custodians cited here 1,12,10,11, and leaving any of them out makes a comparison between two baselines impossible to check.
Global datasets are documented and versioned. HadCRUT5 is in version HadCRUT.5.2.0.0 10; GHCN monthly is in version 4 with adjusted and unadjusted data 12; ERA5 publishes a preliminary release that can differ from the final data two to three months later 11. The platform's source and data register records which version it uses and when.
08Non-stationarity
Baselines that change: non-stationary methods
Methods for baselines in a climate that will not stand still.
A fixed 30-year average assumes a stable climate within the period. Two families of method relax that assumption. Shorter normals give more weight to recent conditions; NOAA's 15-year normals for 2006–2020 are intended for uses that need more recent climate information 14. Non-stationary statistical models let the parameters of a distribution, such as the location of an extreme-value distribution, change with time or with a covariate.
NOAA Atlas 15 applies the second approach to national design rainfall. Its first volume accounts for temporal changes in historical observations through non-stationary methods, and its second projects estimates into the future with downscaled climate model information 15. Such methods need longer and cleaner records than a simple average, which brings homogeneity and completeness back to the centre. The platform's extreme value statistics page covers the statistics involved.
09साधन
Standards and guidance for baseline methods
The WMO rules, guidance and consultations, and the IPCC conventions, as of 28 October 2026.
WMO's technical regulations define period averages, normals and standard normals, and its 2017 guidelines on calculating normals set the completeness and homogeneity practice used for the 1991–2020 collection 1,13. In October 2025 WMO opened consultations on homogenization good practice through its services commission 8.
The IPCC fixes the conventions for global assessment, including the use of 1850–1900 as an approximation of pre-industrial conditions because it is the earliest period with sufficiently complete global observations 5. National standards such as NOAA Atlas 15 set how baselines enter design 15. How such standards are governed is covered by the Global Risks Forum's climate risk governance topic, and the international data cooperation they depend on by its Diplomacy platform.
10Frontiers
What is changing in baseline methods
Three areas where methods are moving.
Shared homogenization practice. WMO consultations 8 aim to bring national practice closer to tested methods 2.
Non-stationary references. Trend-aware design values 15 are moving from research into national standards.
Larger, better documented station archives. Expanded networks with adjusted and unadjusted series side by side 12 make adjustments transparent and testable.
11कंसोर्टियम
Baseline methods across the Consortium
Where method choices meet governance and finance.
Within the Consortium, The Global Centre for Risk and Innovation (GCRI) leads on science, technology and evidence, including the methods behind baselines, supported by the platform's methods and uncertainty and spatial statistics pages. Nexus Foundry is operated at national and regional level by Nexus Consortiums through their host and lead organisations, within a multilateral programme led by these three pillar not-for-profit institutions on a public-private-planet partnership.
The Global Risks Forum (GRF) leads on governance, and its Climate and Nature Governance platform covers the institutions that adopt standards built on baselines. The Global Risks Alliance (GRA) asks how baseline choices change measured exposure. Work moves between them in a cycle: a question is stated, evidence assembled, analysis built and tested, findings published in Reports, competent authorities decide, and lessons reset the next cycle.
What this article does not claim: it explains the methods used to measure climate baselines and the public evidence on it. It is not a forecast, a warning or advice for any decision; national meteorological, hydrological and geological services are the authorities for warnings. Values quoted belong to named datasets and versions and change with new releases. GCRI does not regulate or certify, GRF does not regulate or finance, and GRA does not lend, underwrite or rate. Naming an organisation, framework or dataset is not a claim of partnership, endorsement or deployment.
प्रश्न
प्रश्न और उत्तर
What is homogenization of climate data?
It is the detection and correction of changes in a record that come from the observing system, such as station moves or new instruments, rather than from the climate. Relative methods compare a station with its neighbours; NOAA's monthly network uses a pairwise algorithm of this kind 12.
Do homogenization adjustments change the warming trend?
They change individual station records, and blind testing shows that good relative methods improve temperature data 2. NOAA publishes both unadjusted and adjusted data so the effect can be checked 12.
Why do global temperature datasets disagree?
They make different reasonable choices about coverage, gap filling and merging land and sea data. For 2025 they range from 1.34 to 1.47 degrees Celsius above 1850–1900 3,4, and WMO's consolidated value of 1.43 degrees carries a range of plus or minus 0.13 7.
How many years of data are needed for a climate normal?
WMO guidance recommends data in at least 80 per cent of the years, 24 of 30 for a standard normal, with no gap of 3 or more consecutive years 1.
शर्तें
शब्दावली
- Homogenization
- The detection and adjustment of non-climatic changes in a climate record so that it reflects only weather and climate.
- Relative homogenization
- Homogenization that compares a station with correlated neighbours to separate local changes from the regional climate.
- Structural uncertainty
- Uncertainty arising from different reasonable choices of method between independent datasets.
- Ensemble realisation
- One of many equally plausible versions of a dataset, used together to express its uncertainty.
संबंधित
Further reading across the network
Reuse and cite
Text CC BY 4.0. Cite the page with its date; data tables are linked under each figure.
स्रोत
संदर्भ
15 sources, numbered in the order they are first cited
- World Meteorological Organization (2017). WMO Guidelines on the Calculation of Climate Normals (WMO-No. 1203). www.ncei.noaa.gov. Accessed 2026-10-28.
- Venema et al., Climate of the Past (European Geosciences Union) (2012-01-10). Benchmarking homogenization algorithms for monthly data. Climate of the Past 8, 89–115. doi.org. Accessed 2026-10-28.
- NOAA National Centers for Environmental Information (2026-01-14). Assessing the Global Temperature and Precipitation Analysis in 2025. www.ncei.noaa.gov. Accessed 2026-10-28.
- Copernicus Climate Change Service (2026-01-14). Global Climate Highlights 2025. climate.copernicus.eu. Accessed 2026-10-28.
- Intergovernmental Panel on Climate Change (2021-08-09). Climate Change 2021: The Physical Science Basis. Summary for Policymakers (Working Group I contribution to the Sixth Assessment Report). www.ipcc.ch. Accessed 2026-10-28.
- Berkeley Earth (2026-01-14). Global Temperature Report for 2025. berkeleyearth.org. Accessed 2026-10-28.
- World Meteorological Organization (2026-03-23). State of the Global Climate 2025. wmo.int. Accessed 2026-10-28.
- World Meteorological Organization (2025-10-21). Consultations on good practices for homogenisation of (climate) time-series data (circular 6455061/2025/S/CMP with concept note). extranet.wmo.int. Accessed 2026-10-28.
- Global Climate Observing System (Current ECV list, accessed 2026-10-28). Essential Climate Variables. gcos.wmo.int. Accessed 2026-10-28.
- Met Office Hadley Centre and Climatic Research Unit, University of East Anglia (HadCRUT.5.2.0.0). HadCRUT5 global temperature dataset. www.metoffice.gov.uk. Accessed 2026-10-28.
- Copernicus Climate Change Service (Climate Data Store dataset, updated daily). ERA5 hourly data on single levels from 1940 to present. cds.climate.copernicus.eu. Accessed 2026-10-28.
- NOAA National Centers for Environmental Information (Version 4). Global Historical Climatology Network monthly (GHCNm), version 4. www.ncei.noaa.gov. Accessed 2026-10-28.
- NOAA National Centers for Environmental Information (1991–2020 collection, product page accessed 2026-10-28). WMO Climate Normals (1991–2020 global collection). ncei.noaa.gov. Accessed 2026-10-28.
- NOAA National Centers for Environmental Information (1991–2020 normals, v1.0.1). U.S. Climate Normals. www.ncei.noaa.gov. Accessed 2026-10-28.
- NOAA Office of Water Prediction (Informational page, accessed 2026-10-28). NOAA Atlas 15 informational page. water.noaa.gov. Accessed 2026-10-28.
अभिलेख
Editorial record
Facts and policies are stated as of 28 October 2026. This article is reviewed at least every 180 days and whenever a cited assessment, dataset or instrument changes, including new versions of the cited datasets, the outcome of the WMO homogenization consultations and the publication of NOAA Atlas 15.
Published by the GCRI editorial desk for the Climate and Hazards platform. To report an error or request a correction, use the platform's corrections and concerns route; corrections are dated and listed with the article.
Text is published under the Creative Commons Attribution 4.0 licence (CC BY 4.0): reuse it with credit to GCRI and a link to this page. Third-party data keep their own terms.
Cite this article
GCRI editorial desk (2026, October 28). Measuring climate baselines: methods and their limits. Climate and Hazards. https://therisk.global/climate-hazards/climate-baselines-methods-and-evidence/
@online{gcri2026climatebaselinesmethod,
author = {{GCRI editorial desk}},
title = {{Measuring climate baselines: methods and their limits}},
publisher = {Climate and Hazards},
year = {2026},
month = {10},
url = {https://therisk.global/climate-hazards/climate-baselines-methods-and-evidence/},
language = {en}
} TY - ELEC AU - GCRI editorial desk TI - Measuring climate baselines: methods and their limits PB - Climate and Hazards PY - 2026 DA - 2026/10/28 UR - https://therisk.global/climate-hazards/climate-baselines-methods-and-evidence/ LA - en ER -
Republish
Republish or translate this article free of charge, in print or online, under CC BY 4.0: keep the byline and Climate and Hazards, link to the original and note any change. Figures, maps and twins made by the institutions carry the same licence; third-party data keeps its own credit line.
Attribution to paste
<p>Measuring climate baselines: methods, evidence and limits — GCRI editorial desk, Climate and Hazards. First published 28 October 2026 at <a href="https://therisk.global/climate-hazards/climate-baselines-methods-and-evidence/">https://therisk.global/climate-hazards/climate-baselines-methods-and-evidence/</a>. Republished under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/).</p>