Published: 8 August 2026 | Author: Team JanVayu | Reading time: 6 min
Earlier today we published a result that bothered us. We had put surface temperature on all 70,417 of India's municipal ward records (68,596 distinct wards, as corrected below), and green cover alongside it, and the two barely moved together: a national correlation of −0.07. Within a single city it was negative in 55% of them (680 of 1,247 cities with twenty or more distinct wards), a coin toss. We wrote the narrow version of the claim and left it there.
The narrow version was wrong. The arithmetic was fine. The variable was the problem.
The map now has a Tree cover layer, canopy alone, for every level from ward to state. Against the same heat data, on the same wards:
| Against ward surface heat | Green cover | Tree cover |
|---|---|---|
| National correlation | −0.069 | −0.429 |
| Within-city median correlation | −0.107 | −0.412 |
| Cities where the relationship is negative | 680 / 1,247 (55%) | 1,087 / 1,240 (88%) |
| Coolest fifth vs hottest fifth of wards | 0.8 °C | 4.8 °C |
The two city counts differ slightly (1,247 and 1,240). Both are cities with twenty or more distinct wards, and we have not traced why seven drop out of the tree-cover count. The heat-island effect was there all along. We were looking at it through the wrong lens.
What "green cover" counts
ESA WorldCover classifies every 10 m of the planet. Its vegetation classes are tree cover, shrubland, grassland, cropland, wetland and mangrove. "Green cover" adds them all together.
In a city that is a reasonable stand-in for greenery: parks, scrub, roadside trees. Outside a city, it is farmland. The median Indian gram panchayat is 97% green, and 87% of them are above 90%. The number is arithmetically correct and tells you almost nothing, because it really measures "not built and not bare", which describes nearly all of rural India.
It is worse for the heat question. A harvested or fallow field in Vidarbha in May counts as green by this definition and is one of the hottest surfaces in the country, reaching 57 °C. So the layer mixed "shaded by canopy" with "bare dirt that used to be a crop" in one number, and we then asked it which places are cool.
Tree canopy on its own (WorldCover class 10) separates places properly. The median ward is 9% treed, the median panchayat 12%, the median state 38%. The values spread across the full range instead of piling up at 97%.
The number that matters
The least-treed fifth of India's wards averages 43.9 °C at the surface. The most-treed fifth averages 39.1 °C.
That is a 4.8 °C difference, measured across roughly 13,500 wards on each side, in the same season, from the same satellite. Green cover's version of the comparison is 0.8 °C, which is nothing.
It holds nearly everywhere. In 88% of the 1,240 Indian cities with twenty or more wards, the wards with more canopy are the cooler ones. Bengaluru moves from −0.17 on green cover to −0.44 on tree cover. Savanur, the town we singled out last time as a baffling reversal at +0.83, goes to −0.03. The reversal was cropland all along, not a quirk of arid urban form.
Delhi makes it concrete. Naraina, in west Delhi, has 12% tree cover, 87% built-up land and a surface temperature of 41.5 °C. A ward in New Delhi's Lutyens area has 45% tree cover, 53% built-up and 39.9 °C. It is less built-up (53% against 87%), has much more canopy, and is 1.6 °C cooler, so the gap is not a clean canopy effect.
Two corrections to this morning's post
The earlier framing was narrower than the data warranted. This morning's post said the defensible statement was "within a humid or temperate Indian city, green cover tracks cooler ward surfaces; across India, and inside arid cities, it does not". That is true of green cover. But we presented it as a limit on what the data could show, when it was a limit on the variable we had chosen. We should have asked what green cover was made of before drawing conclusions from it.
The ward counts have been revised down. While chasing an unrelated bug we found the ward atlas carries 2,541 exact-duplicate geometries, bunched in a few cities. Patna's "628 wards" are 115 distinct shapes. Mangalore's "540" are 60. "Ward 1" appears twenty-three times in Patna. Every correlation in this post is computed on deduplicated wards. The earlier post's table has a correction note attached, with the wrong figures left visible. The correlations barely moved (the national one went only from −0.054 to −0.069), but the counts were wrong and we had printed them.
We found the bug only because a total went down. Our new land-cover calculation puts every polygon of a level into one grid, so each pixel belongs to exactly one ward. That works when wards tile the map and fails when they overlap. Patna came back with 556 of 628 wards having no data at all. It looked like ordinary missing data, and we investigated only because the previous method had managed more.
What to do with it
Open the map, pick any boundary level, set Colour to tree cover. Every level except villages has it (ward, block, panchayat, city, district, state) at 99.93% coverage or better. Tap anything for all five figures at once: annual air, surface heat, tree cover, green cover, built-up.
Update, 9 August 2026: villages carry tree cover too (v26.6.153), so tree cover now exists at every level.
If you are arguing for planting in a specific ward, its canopy percentage and surface temperature are now a click away, next to the same figures for the ward next door.
Two cautions from earlier still apply, and both are on the map itself. Surface temperature is not air temperature. It is how hot the ground gets under a clear pre-monsoon sky, well above the shade forecast. And this is a seasonal figure, taken from the hottest, clearest stretch of the year. It answers "how hot does this place get", not "how hot is it usually".
We are also not claiming canopy causes the difference. Wards with more trees differ in other ways: density, building materials, water, wealth. What the data supports is that canopy is the vegetation measure that tracks surface heat in India, and cropland-inclusive "green cover" is not.
Method, per-level coverage and caveats: The Boundary Map. Land cover from ESA WorldCover 2021 (CC BY 4.0); heat from Landsat 8/9 via Microsoft Planetary Computer; annual PM2.5 from SatPM2.5 V6GL03.