Climate Risk Data: Global Climate Databases for Risk Assessment

No single dataset measures physical climate risk everywhere. An assessment stacks several data layers, then needs a method to turn them into ratings and a careful reading of the scores that come out.

This guide organizes widely used public climate risk data sources by hazard type, with the one limit to know before you use each. Whether you are a risk manager building in-house capabilities, an ESG analyst sourcing data for IFRS S2-aligned disclosure, or a developer integrating climate data into applications, you will find the datasets you need. For a complete walkthrough of how these data sources fit into an end-to-end workflow, see our climate risk analytics guide. For an API-first comparison of climate data providers, see our climate data APIs guide.

What Makes a Good Global Climate Database?

Before choosing a dataset, check six things. Together they decide whether the climate risk data fits the question you are asking.

• What it describes. A flood map describes set floods, a surface elevation model includes roofs and trees, and an event database lists only what was reported.
• Resolution. Downscaled climate projections come at about 25 km, which holds up for temperature trends. Flood and sea level exposure need terrain-scale data (10 m to 90 m) at each site.
• Baseline period. Projections are measured against a reference period. The IPCC’s latest assessment uses 1995 to 2014 as its recent baseline and 1850 to 1900 for warming since pre-industrial times.
• Scenarios and models. Current projections use the Shared Socioeconomic Pathways (SSPs). Gather a low, a middle and a high scenario, such as SSP1-2.6, SSP2-4.5 and SSP5-8.5, and note which climate models sit behind each number.
• Version and date. Projection datasets are reissued with corrections, and observed records keep growing. Record the version and the date you pulled each layer.
• Coverage and gaps. Check the latitude limits, the years covered and how the dataset marks missing data, so a gap is never read as zero risk.

Climate Projection Datasets

Climate projections form the foundation of any forward-looking risk assessment. These datasets provide temperature, precipitation, and wind projections under different emissions scenarios.

NASA NEX-GDDP-CMIP6 (see our in-depth NEX-GDDP-CMIP6 guide) is the workhorse dataset for climate risk assessments. It provides bias-corrected, statistically downscaled daily projections from 35 CMIP6 climate models at about 25 km. Variables include temperature, precipitation, wind, humidity and radiation. It covers four scenarios (SSP1-2.6, SSP2-4.5, SSP3-7.0 and SSP5-8.5) from 2015 to 2100, plus each model’s historical run from 1950 to 2014, though not every model carries every scenario and variable. Limit: each value averages a 25 km cell, so city heat, valleys and coastlines inside the cell are smoothed out, and coverage stops at 60°S.

The CMIP6 archive on the Earth System Grid Federation (ESGF) holds the raw output of the global climate models. Limit: raw output is coarse, often 100 km or more, and is not bias-corrected, so its values cannot be compared directly with observations.

The Copernicus Climate Data Store hosts ERA5 reanalysis, which gives hourly historical conditions from 1940 to the present at about 31 km. ERA5 is essential for establishing the baseline conditions that future projections are compared against. Limit: a reanalysis blends a weather model with observations, and at this grid size it underestimates short, intense rainfall.

CORDEX provides regionally downscaled projections at about 12 to 50 km for set regions such as Africa, Europe and North America. Limit: coverage is regional and the runs differ by region, so check which global models drive the region you need.

Climate Models Disagree, So Read the Spread

NEX-GDDP-CMIP6 carries 35 climate models, and they give different answers for the same place. One reason is climate sensitivity: how much a model warms when carbon dioxide doubles. Models in this set range from under 2°C to above 5°C. The IPCC’s best estimate is 3°C, with a likely range of 2.5°C to 4°C and a very likely range of 2°C to 5°C, and it found that some of the most sensitive models warm faster than the wider evidence supports.

• Use several models that span the sensitivity range, with at least one lower and one higher model. Name each model.
• Treat the hottest models as a high-warming stress case, which is how the IPCC suggests using them.
• Keep each model’s result and show the spread. Report how many models agree on each rating.
• Combine models with a rule one outlier cannot drag, such as the median, and state the rule.
• Expect disagreement on heat and on rainfall. Every model warms, but by different amounts, so heat ratings can differ by a band or two. Rainfall ratings spread wider, and up to about 2040 the scenarios warm almost the same amount, so natural swings in rainfall can rank a low scenario above a high one.

Running several models is heavy work. Each model, scenario and variable is a separate download of daily data, and every hazard has to be computed for each model before any combining starts.

Temperature Hazard Data

Temperature hazards include heat waves, cold stress, and long-term warming. These drive impacts on labor productivity, energy demand, and infrastructure.

For future projections, NEX-GDDP-CMIP6 provides daily mean, maximum and minimum temperature (tas, tasmax, tasmin) and humidity. Heat wave assessments typically count days above a threshold set from the local baseline, such as a high percentile of daily maximum temperature. For health and outdoor work, add humidity, because humid heat is harder on the body.

Landsat 8 and 9 surface temperature, from USGS and NASA, maps how hot roofs, roads and open ground get. It measures at 100 m, delivered on a 30 m grid, about every 8 days with both satellites, in daytime passes only. Limit: land surface temperature is how hot a surface would feel to the touch, which differs from the air temperature in a weather report, and it is captured only on clear-sky passes.

Precipitation and Drought Data

Precipitation hazards include drought, extreme rainfall, and changes in seasonal patterns. These affect water availability, agriculture, and flood risk.

NEX-GDDP-CMIP6 provides precipitation projections, but converting raw precipitation to drought indices requires additional processing.

CHIRPS (Climate Hazards Group InfraRed Precipitation with Station data) blends satellite and station rainfall at about 5 km from 1981, which is valuable where gauges are sparse. Version 3 covers 60°N to 60°S. Limit: it gives daily totals, so it cannot show the short bursts behind flash floods.

The SPEI Global Drought Monitor provides monthly Standardized Precipitation-Evapotranspiration Index values from 1955, a drought indicator that accounts for both rain deficits and temperature-driven evaporation. Limit: a coarse 1° grid (about 110 km) showing observed drought, with no projections.

Rainfall Frequency Curves for Surface-Water and Flash Flood

Surface-water and flash flood analysis starts from rain: how much falls in 15 minutes, an hour or a day in a storm of a given rarity. National agencies publish this as rainfall frequency curves, also called intensity-duration-frequency (IDF) curves. Examples are NOAA Atlas 14 in the US, the Flood Estimation Handbook rainfall model (FEH22) in the UK, the Bureau of Meteorology design rainfalls in Australia, and Environment and Climate Change Canada’s station IDF files. Applying a curve to one site is where the work gets heavy.

• Find the right curve. Some countries publish a national grid, others publish station curves or regional equations, and many publish none.
• Check the durations. Flash floods need storms of minutes. Some curves start at an hour or more, so the shortest values can be extrapolations.
• Check the climate allowance. Most national curves assume a stable climate and leave the uplift to separate guidance, while a few build it in. Adding an uplift on top of one already built in counts it twice.
• Turn the curve into a storm. A model needs rain over time, so the curve has to become a design storm with a shape and a length before anything can run.

Limit: most curves describe rainfall at a gauge or a grid point from past records. See how rainfall intensity is measured in our rainfall intensity guide.

Continuuiti sources the best available published rainfall frequency curve for each site, turns it into design storms, and names the source in the report. That includes storms shorter than an hour for flash flood analysis.

Water Stress and Flood Data

Water stress represents long-term supply-demand imbalances, while flood risk captures acute inundation events. These require different data sources.

WRI Aqueduct 4.0 reports present-day water stress, depletion, variability, groundwater decline, drought and flood risk by sub-basin. Projections of water stress, depletion, variability, supply and demand cover 2030, 2050 and 2080 under three pathways: SSP1-2.6, SSP3-7.0 and SSP5-8.5. It is a widely used dataset for the water stress hazard in physical climate risk assessments. Limit: values average a whole sub-basin, and its middle pathway is SSP3-7.0, so it does not line up with an SSP2-4.5 column from your climate projections.

Global river flood hazard maps from the EU’s Copernicus Emergency Management Service, produced by the Joint Research Centre, show river flood depth at about 90 m for seven return periods, from the 1-in-10-year to the 1-in-500-year flood. Limit: they cover rivers draining more than 500 km² and today’s climate only, and they leave out rain-driven and coastal flooding.

WRI Aqueduct Floods maps river and coastal flood depth at about 1 km for return periods up to 1-in-1,000 years, for today and for 2030, 2050 and 2080 under older emissions scenarios. Limit: at about 1 km it shows exposure at district scale and cannot place water on a single building.

Groundsource, released by Google Research in 2026, lists about 2.6 million flood events reported in the news across more than 150 countries. Limit: each event carries the outline of the place a report names, from a street to a whole district, so it shows where flooding was reported. One flood can appear as many records, places with more news coverage hold more records, and no flood cause is recorded.

Surface-Water Flood Needs a Model Run

River and coastal flood maps leave out rain that pools before it reaches a river. Mapping it takes a rain-on-grid model, which pours a design storm over a bare-earth terrain grid, treats buildings as obstacles, and tracks where the water flows and collects in small time steps. That makes it compute heavy: one storm over a few square kilometres can take minutes or longer to run, and a portfolio needs several storms at every site. You can watch rain-on-grid runs on worked-example sites in our pluvial flood simulator, and run it at your own locations with an account.

Flood Zones Are a Regulatory Starting Point for Flood Data

Official flood zone maps, such as FEMA’s flood maps in the US, are often the first flood data people reach for. A flood zone is a regulatory line. Maps go through public review, owners can appeal them, and boundaries are revised as new evidence arrives. The high-risk zone is drawn for one flood, the one with a 1 percent chance each year, and US maps largely leave out flooding from heavy rain. Each map also shows conditions in the year it was made. Record the zone and the map’s effective date, because lenders, insurers and planners use them. Then run a flood risk assessment for water depth at several return periods, including surface-water flooding.

Sea Level Rise Data

Sea level rise affects coastal properties, infrastructure, and entire communities. Assessment requires both sea level projections and high-resolution coastal elevation data, such as the bare-earth models listed under terrain below.

IPCC AR6 Sea Level Rise Projections provide the authoritative global projections by scenario. By 2050, median global sea level rise is projected at 0.19 m under SSP1-2.6, 0.20 m under SSP2-4.5 and 0.23 m under SSP5-8.5, relative to 1995 to 2014, so the scenarios differ little before mid-century. By 2100 the likely range under SSP5-8.5 is 0.63 to 1.01 m. Limit: these are global averages, local rise differs where land sinks or rises, and higher outcomes driven by ice sheets cannot be ruled out.

The NASA and IPCC Sea Level Projection Tool gives the same projections for points along the coast and at tide gauges, including land movement, by decade to 2150. Limit: it gives sea level, and flood depth at a site still needs coastal elevation and storm surge.

NASA Sea Level Change provides observed sea level trends from satellite altimetry, essential for validating model projections against actual measurements. Limit: satellites measure the open ocean since the early 1990s, so trends at a shoreline need tide gauge records.

Fire and Compound Hazard Data

Wildfire risk depends on temperature, humidity, wind, and vegetation. The Fire Weather Index needs temperature, relative humidity, wind speed and precipitation, and NEX-GDDP-CMIP6 carries all four as daily values.

The Copernicus Emergency Management Service publishes historical fire danger indices, including the Fire Weather Index, computed from ERA5 from 1940 to the present. Limit: it measures fire weather and says nothing about fuel or what stands near a site.

NASA FIRMS (Fire Information for Resource Management System) provides near-real-time active fire detections at 375 m to 1 km. Use it for historical fire occurrence and current monitoring. Limit: it records fires seen burning at each satellite pass.

GlobFire, from the EU Joint Research Centre, turns satellite burned-area maps into individual fire perimeters, and the Global Wildfire Information System (GWIS) shows them worldwide from 2001. Limit: perimeters come from coarse burned-area data, so small fires can be missed, and the record lags the present by a few years.

Terrain and Context Data

Terrain data provides critical context for hazard calculations. First check whether an elevation model is a surface model, which includes treetops and roofs, or a bare-earth model. Flood and sea level work needs bare earth.

NASA SRTM (Shuttle Radar Topography Mission) provides global elevation at 30 m from a radar mission flown in February 2000, covering 60°N to 56°S. Limit: it is a surface model, so trees and buildings raise the ground level, which can understate flood and sea level exposure in cities and forests.

Bare-earth models such as GEDTM30 (30 m, global, built with satellite lidar and machine learning) and DeltaDTM (low-lying coasts, built with satellite lidar and a terrain filter) remove trees and buildings. Limit: the corrections are statistical, and different terrain models can put the same spot several metres apart, so name the model you used.

ESA WorldCover provides global land cover classification at 10 m. Limit: it exists for 2020 and 2021 only.

Dynamic World, from Google and WRI, gives 10 m land cover for each Sentinel-2 image with less than 35% cloud since mid-2015. Limit: single-date maps are noisy, so combine several images over a season.

Building outlines come from Microsoft, Google Open Buildings and Overture Maps, which merges these with OpenStreetMap. Limit: machine-detected outlines can miss or merge buildings, and they carry no ground-floor height above the ground, which flood damage estimates need.

Road centrelines come from OpenStreetMap, also packaged by Overture Maps. Limit: volunteers map it, so completeness varies by country.

Historical Disaster Databases

Historical disaster data helps validate risk models and understand past impacts. These databases complement forward-looking climate projections.

EM-DAT (Emergency Events Database), maintained by CRED at UCLouvain, records over 27,000 disasters worldwide since 1900, including economic losses, fatalities, and affected populations. Limit: an event enters only with at least 10 deaths, 100 people affected, or an emergency declaration or call for international aid, so small local events are missing, and its maintainers advise leaving pre-2000 records out of trend analysis.

The NOAA Storm Events Database provides detailed records of storms and severe weather in the United States, including county-level impacts and narrative descriptions. Limit: it covers the US only, and its damage figures are reported estimates.

IBTrACS, from NOAA, collects tropical cyclone tracks and intensities from agencies worldwide, from 1841 to the present, updated several times a week. Limit: early decades miss storms, and it gives a storm’s track and wind speed, so wind at a site needs a wind field model.

NASA’s Global Landslide Catalog lists rainfall-triggered landslides reported from 2007. Limit: it has had few new entries since 2019, and it holds only what was reported.

Aggregated Indices and Portals

Sometimes you need a ready-made answer for a place. Each index or portal below answers a different question, so pick the one that matches yours.

The Germanwatch Climate Risk Index ranks 174 countries by deaths, people affected and economic losses from extreme weather, for the latest year and over three decades, using EM-DAT data. The latest edition is the CRI 2026. Limit: it looks back at past impacts by country and says nothing about future exposure or a single site.

FEMA’s National Risk Index gives US risk scores by county and census tract. Limit: it reflects today’s hazard and loss history and carries no climate projections.

The ND-GAIN Index (Notre Dame Global Adaptation Initiative) measures country vulnerability and readiness to adapt to climate change, combining exposure, sensitivity, and adaptive capacity indicators. Limit: country scores cannot separate two sites in the same country.

The World Bank Climate Change Knowledge Portal provides country-level climate risk profiles combining historical data, future projections, and vulnerability indicators. Limit: profiles summarise countries and regions, so use them for context.

Climate Risk Data Needs a Method to Become an Assessment

A method turns a stack of climate risk data into a view of which sites are at risk. It sets how each layer is read, which scenarios and years are compared, and where a value becomes a rating. Most sound methods share these steps.

• Place each asset precisely. Convert addresses to building-level coordinates and check the quality of each match. A town-level match can work for heat and is too coarse for flood.
• Pick one measure per hazard that links to damage or disruption, such as days above a local heat threshold or flood depth at several return periods.
• Compare like with like. Measure future change against the same model’s own historical run, and keep baseline periods and time windows consistent across hazards where the data allows.
• Set the thresholds and write them down. The point where a value becomes High or Severe is a judgment, so a reviewer should be able to see and test it.
• Add exposure and vulnerability to get money. A loss figure needs the asset’s value, use and floor height, read through depth-damage curves. FEMA’s Hazus curves cover US building types, and the EU Joint Research Centre publishes global curves. Limit: Hazus curves are built for US construction, and the global curves are averages by continent and land-use class.
• Check against what happened. Compare projections with the observed event record near the site. A mismatch is a finding to explain.

A Composite Climate Risk Score Can Hide the Hazard That Matters

Most methods end in a score. A composite climate risk score blends every hazard at a site into one rating, often on a 1 to 5 scale. It is useful for ranking a portfolio, and it can hide the reading that matters most. Averaging lets one Extreme hazard disappear under several quiet ones, so a site can read Moderate while one hazard sits at the top of the scale. A flat average also counts every hazard equally, though a foundry depends on water and heat far more than a warehouse does. Before trusting a score, ask to see the hazard readings under it, check whether hazards are weighted to the operation, and confirm whether a zero means safe or no data.

Why Climate Risk Data Is Hard to Assemble at Portfolio Scale

Pulling one dataset for one site takes a morning. Doing it for hundreds of sites, across a dozen hazards, three scenarios and several years, is a different job, and the problems show up between datasets.

• Grids and dates do not line up. A 25 km climate cell, a 90 m flood map and a 30 m terrain model each describe a different patch of ground, and their baseline periods and future windows rarely match.
• Scenario names do not match. A water dataset’s middle pathway can be a different and hotter scenario than the middle scenario in your climate projections, so two columns that look aligned can describe different futures.
• Codes look like values. Datasets use special numbers for no data, arid land or a fill value. Read as values, they turn an empty cell into Extreme risk or a dry site into a flooded one.
• Checks grow with every site. Each location, flood depth and coastal reading needs a sense check against its surroundings. A quick script returns a number for every site without saying which numbers are wrong.

This is the work Continuuiti takes on: its climate risk API returns the hazard layers aligned across scenarios and years for every site, and our team runs the checks across the whole portfolio, so you can focus on what the results mean. To compare platforms, see how to evaluate climate risk assessment tools.

Choosing the Right Data for Your Assessment

The right data depends on your use case:

• For IFRS S2-aligned disclosure: IFRS S2 requires climate-related scenario analysis to assess your climate resilience, using an approach that fits your circumstances (paragraph 22). You disclose your inputs: which scenarios you used and their sources, whether they cover a diverse range, the time horizons, and the scope of operations, such as the operating locations covered (paragraph 22(b)(i)). The standard sets no number of scenarios. Using years that match how you define short, medium and long term (paragraph 10(d)) keeps the disclosure consistent. Record each dataset’s name, version and models so you can explain your inputs. For the amount and percentage of assets or business activities vulnerable to physical risks (paragraph 29(c)), use all reasonable and supportable information available without undue cost or effort (paragraph 30). Hazard data at each asset makes that figure easier to support. IFRS S2 does not define “vulnerable”, so set your own cut-off and explain it.
• For portfolio screening: about 25 km climate projections work for heat, rainfall and drought across hundreds or thousands of assets. Flood and coastal exposure still need flood maps and terrain at each site, even at the screening stage. Country indices rank countries and cannot separate two sites in one country, so screen at site level and drill down where the flags are.
• For site-specific analysis: combine climate projections with a bare-earth terrain model, building outlines, river and coastal flood maps, local rainfall frequency curves and the observed event record. Consider local factors that global models miss, such as drainage and flood defences.

Frequently Asked Questions

A climate risk assessment needs the right climate risk data layers, a method that turns them into ratings, and a careful reading of the scores. The datasets in this guide are the building blocks. Know what each one describes and where it stops, and the assessment you build on them will stand up to review.