CHELSA Climate Data: What It Is, How It Works, and Where It Breaks

CHELSA — Climatologies at High resolution for the Earth's Land Surface Areas — is a global, ~1 km terrestrial climate dataset produced by WSL, the Swiss Federal Institute for Forest, Snow and Landscape Research, with the University of Zurich's Department of Geography. It was introduced by Karger et al. in a 2017 Scientific Data paper and is now on version 2.1.

The dataset that gets cited constantly is the 1981–2010 climatology: 30-year monthly normals for temperature and precipitation, downscaled to 30 arc-seconds — roughly 1 km at the equator. CHELSA also publishes monthly time series (1979–2021), a daily series, and projections out to 2100, but the climatology is the workhorse: it's the base layer for a large share of published species distribution models (SDMs), because it's the highest-resolution global climatology that covers every landmass with no gaps.

GeoRank uses CHELSA v2.1 directly — it's the source for our temperature and rainfall map layers, and for a derived comfort index. We're not affiliated with CHELSA or WSL. This page exists because after integrating it into a production pipeline, we couldn't find a plain-language writeup of what actually matters when you use this data for real — so we wrote the one we wished existed — as we did for ERA5, SARAH-3, and sunshine duration measurement. Full technical detail on how GeoRank specifically processes it is on the methodology page.

CHELSA downscales coarse global reanalysis — ERA5 — using a mechanistic, terrain-informed model rather than pure statistical interpolation. That distinction matters more than it sounds like it should.

For temperature, CHELSA computes dynamically-calculated lapse rates from the vertical air-temperature profile at each location, instead of applying one fixed global lapse rate (the textbook ~6.5°C per 1,000 m) everywhere. Real lapse rates vary by season, humidity, and inversion conditions, and a single constant gets deserts and humid tropics wrong in different directions.

For precipitation, CHELSA redistributes rainfall using physical orographic predictors: a wind exposure index (windward slopes get more rain, leeward slopes get less), planetary boundary layer height (which caps how far upslope the orographic effect can reach), and a valley exposition index (which accounts for rain-shadowing inside individual mountain valleys, not just at the range scale). This is a genuinely different approach from statistically interpolating between weather stations — it lets CHELSA produce physically-motivated precipitation gradients in places with zero nearby gauges, which is most of the world's mountains.

Coarse reanalysis products like ERA5 run at roughly 25–31 km per grid cell. At that scale, an entire mountain range, a whole river valley, and the ridge above it can all fall inside the same pixel and get reported as identical. CHELSA's ~1 km grid is fine enough to separate the valley floor from the ridge 500 m above it, and to give a coastal town a different (and usually damper, more moderate) profile than a location 20 km inland at the same latitude.

In our own pipeline, CHELSA's temperature layer typically tracks ground stations within ±1–2°C in well-instrumented regions — a real, usable improvement over what a ~25 km reanalysis cell can offer for anything smaller than a country. That's the whole value proposition: resolution fine enough that "the climate where I actually am" and "the climate GeoRank shows me" start to converge, rather than both places 15 km apart on a mountainside sharing one number.

It is not perfect at that scale, though — see the two sections below for where it degrades.

CHELSA does not model urban heat islands. A 1 km grid cell over a city centre reports the climatological response of that patch of land generically — it is not measuring what a thermometer bolted to a building in the middle of the asphalt and concrete actually reads. Cities run warmer than their surroundings, sometimes by several degrees, because of reduced vegetation, heat-retaining materials, and waste heat from buildings and traffic — none of which is in CHELSA's downscaling model.

In our own validation against city-centre observations, this shows up as a real, directional bias: CHELSA-derived summer temperatures can underestimate dense city centres by 1–3°C. For ecology and species-distribution work, this rarely matters — CHELSA was built for wildland and vegetated land cover, which is most of the land surface most SDMs care about. But if you're using CHELSA (or anything built on it) to answer "what does it actually feel like to stand in this city," the grid-cell number is a systematic underestimate, not noise that averages out.

CHELSA's wind-exposure and valley-exposition indices are a real improvement over naive elevation-only downscaling, but complex terrain is still the hardest case for any gridded climate product — mechanistic or statistical. Rain shadows compound across successive mountain ranges, valley orientation can flip the sign of an effect between two valleys ten kilometres apart, and thin data coverage in remote high-altitude regions means the model has less to calibrate against exactly where the terrain is most extreme.

In practice, we see the error budget widen sharply in mountainous coastal terrain: precipitation totals that track observations to roughly ±12% in well-gauged, gentler terrain can drift to ±25% where a coastline meets a mountain range head-on. Temperature holds up better but still loosens to roughly ±2–3°C in steep terrain, largely because a grid cell's mean elevation rarely matches the elevation of the specific point you actually care about — a peak, a pass, or a valley floor, all averaged into one number.

Both are free, global, ~1 km climatologies, and both show up constantly in the same ecology and SDM papers — sometimes as sensitivity checks against each other. They are not interchangeable, and the difference is methodological, not just cosmetic.

The practical difference: WorldClim's accuracy is tied to station density — it's strong where stations are dense (much of Europe, North America) and weaker where they're sparse (much of the tropics, high mountains, the Arctic), because there's less ground truth for the spline to interpolate between. CHELSA's mechanistic approach doesn't need a nearby station to produce a physically plausible orographic gradient, which is why it tends to be the preferred choice for topographically complex or data-sparse regions. Neither is strictly "more accurate" everywhere — pick based on your climatology period and your terrain, and check both against local ground truth if the decision matters.

The CHELSA climatology is a 30-year average, not a snapshot of any particular year and not a live feed of current conditions. That's the entire point of a climatology: it smooths out any single year's heat wave, drought, or cold snap to describe the long-run climate regime a location sits in — which is exactly what species-distribution and biogeography work needs, since organisms respond to climate over decades, not to this July.

What it does not tell you is what this year, or next year, will look like. The baseline ends in 2010; it does not capture warming or precipitation-pattern shifts since then. If your question is about present-day or forecast conditions rather than the long-run climatic envelope a place sits in, CHELSA's climatology is the wrong tool — you'd want a current reanalysis or a forecast product instead.

A few things that only become obvious once you're actually pulling the files. CHELSA's monthly climatology tiles are distributed as individual GeoTIFFs — 12 files per variable, one per month, each roughly 80–150 MB depending on the variable. Temperature ships as scaled int16 (raw × 0.1, minus 273.15 to convert from kelvin-tenths to °C); precipitation ships as scaled uint16 (raw × 0.1 → mm/month). Get the scale factor or the nodata sentinel wrong — CHELSA uses -32768 for temperature and 65535 for precipitation — and you'll silently bake corrupted values into a production layer. We've done it; it's an easy mistake to make once and a bad one to make twice.

GeoRank pulls CHELSA's tas, tasmax, tasmin, and pr tiles directly from CHELSA's public object storage, warps them from 30 arc-sec to our 0.05° map grid, and bias-corrects the precipitation layer against roughly 25,000 GHCN-M v4 station normals. The temperature layer feeds the map's temperature tiles directly; temperature and precipitation together feed the Outdoor Hours Index, where CHELSA's ~1 km temperature is combined with coarser ERA5 humidity and wind as regional modifiers. Full detail, including the exact formulas, is on the methodology page.

• Urban heat islands are not modelled. Grid-cell values represent the generic land-cover response, not a city-centre thermometer — expect summer underestimates in dense cities.
• Orographic accuracy degrades in complex terrain. Rain-shadow and windward effects are modelled physically, but error still widens in steep, mountainous, or coastal-mountain terrain relative to gentler ground.
• It's a climatology, not current conditions. The 1981–2010 baseline describes a long-run regime, not this year, and does not reflect warming or precipitation shifts since 2010.
• Encoding gotchas are real. Scaled integer storage with distinct nodata sentinels per variable (temperature vs. precipitation) is an easy source of silent data-corruption bugs if you're writing your own ingestion pipeline.
• Downscaled ≠ locally observed. Even at 1 km, a grid cell is an area average. Microclimates smaller than ~1 km — a sheltered courtyard, a frost hollow, a single exposed ridge — will still diverge from the grid value.
• Total precipitation isn't phase-separated. The pr variable sums rain and snow; if you need liquid-vs-frozen precipitation specifically, CHELSA's base variables don't split it out for you.

Primary sources for everything on this page: chelsa-climate.org (dataset, downloads, model documentation), the Karger et al. (2017) paper above, and WSL, the producing institution. CHELSA's climatologies are released under CC0 1.0 — no attribution is legally required, though citing the paper is standard academic practice and keeps the project fundable.