Correlations between the map’s fields
The map paints 28 modelled pollution fields and 3 weather fields on the globe. This study asks which of them rise and fall in the same places, and what that says about where the air's contents come from.
Single snapshot. Every number here comes from the snapshot valid at . It is daytime over Asia and Australia, night over Europe, Africa and the Americas. Correlations are across space, not over time.
- Partial ρ
- the link that remains after holding every other field fixed
- 95% interval
- the range ρ would likely fall in if the map were resampled in 15° tiles
Local correlation maps
Each dot is a 5° tile with about 150 grid cells. Its area and colour show how strongly the two fields rank together inside that tile, from violet (one rises as the other falls) through grey to red (they rise together). Faint dots are land.
Global correlation matrix
Rank correlation (Spearman ρ) between every pair of the 15 global fields and 3 weather fields, weighted by area. Rows are ordered by clustering, so related fields sit together; each label carries the field's globe colour. Dot area and colour show ρ; hover or focus a dot for the numbers.
The 16 strongest pairs. The 95% interval for ρ comes from resampling the map in 312 tiles of 15° (300 times), which allows for neighbouring cells being alike.
Joint distributions
Joint distributions, weighted by area. Each dot is a bin; its size shows how much of Earth's surface falls there, its colour whether that surface is land or ocean. Concentration axes are logarithmic.
Correlation by region
Spearman ρ on land cells inside each region box. The same pair can be strong in one place and absent in another, because different sources dominate.
Day and night
Land cells split by local solar time at the snapshot. Daylight drives photochemistry, so some links flip sign across the terminator.
PM2.5 share of PM10
How much of PM10 is fine particles. Dot: median. Thick bar: middle half of cells. Thin line: 10th to 90th percentile. Cells with PM10 above 0.5 µg/m³, weighted by area.
Spatial variability by field
How much each field changes from place to place. Fields in the shaded band (below 0.5) look nearly uniform, so the globe draws them fainter.
Europe correlation matrix
The European ensemble adds ammonia, VOCs, carbon fractions, wildfire smoke and pollen. Land cells only. Alder, birch, olive and ragweed pollen are left out because they are zero everywhere in October.
Scroll sideways to see every field.
Global vs European model
Zooming into Europe swaps the global model for the European ensemble. Rank agreement shows whether they put pollution in the same places; the ratio shows whether they agree on how much.
What explains these patterns
A correlation does not show a cause. For each strong pattern we wrote down the most likely physical explanation and ran a check that would fail if the explanation were wrong. Every field here is model output, so a passing check shows the model behaves as the physics predicts. Confirming it in the real atmosphere needs measurements.
Conclusions
About the data
What the globe shows
A globe of modelled air composition. Each dot is a point on Earth, sized by the concentration of the layers you switch on: fine particles (PM₂.₅), ozone, nitrogen dioxide, dust and more. Only PM₂.₅ is on at first; open Layers to add others. Hover or tap anywhere to read the values underneath, or search for a city.
The data
- Global: CAMS global atmospheric composition forecasts, 15 fields on a 0.4° grid (about 45 km).
- Europe: the CAMS European ensemble on a 0.1° grid (about 11 km) with 25 fields, 13 of them Europe-only. It replaces the global dots when you zoom in over Europe.
- Weather: temperature, humidity and wind from ECMWF IFS at 0.25°, for the same hour.
Searched places
Choosing a place opens the same readout as tapping the globe: the layers you have on, with every other field under More fields, for the hour on the time bar, so it always matches the dots. Search uses a built-in list of cities, so what you type never leaves your device, and choosing a place sends no request.
How to read the colours
Each layer has its own pigment, all at the same perceived lightness, and colours mix where layers overlap. Dot size encodes the value. The strength slider changes only how strongly a layer is drawn, never the data.
Mixed colours are an artistic encoding, not a combined index, and nothing here is a health index. WHO 2021 guideline values appear beside readouts as context only.
Limits
- Values are modelled estimates, not sensor readings. Each stands for a grid cell, not street-level air or personal exposure.
- PM₁₀ includes PM₂.₅, and dust and other aerosols overlap, so do not add layers together.
- Aerosol optical depth describes the whole atmospheric column at 550 nm, not air at ground level.
- Correlations in this study come from shared model inventories and chemistry as well as the real atmosphere. They show co-location, not cause.
- Values are stored at fixed precision (0.1 µg/m³ for most gases), so the lowest bins are coarse.
Method
405,900 global cells at 0.4°, weighted by cos(latitude); weather resampled to the same grid. Spearman ρ is the headline because concentrations span orders of magnitude; Pearson on log₁₀ values and partial correlations come from the same cells. Land and ocean use the globe's own water mask. The analysis script is analysis/correlations.py in the project repository.
Credits & licence
Earth: NASA Blue Marble Next Generation, July 2004 composite. Static imagery, not current satellite conditions.
Data: Copernicus Atmosphere Monitoring Service / ECMWF, via Open-Meteo, CC BY 4.0. Weather: ECMWF IFS via Open-Meteo AWS Open Data, CC BY 4.0.
Photos in the explanations come from Wikimedia Commons; each is credited beneath it. Two are CC BY-SA (Pdekyvere, Derek Harper), the others CC0 or public domain.