Where Zagreb runs hottest — and who lives there

RQ-A — Composite Heat-Vulnerability Index

Weighted HVI across all 218 mjesni odbori: UHI, elderly share, density, NDVI gap, and heatwave days.

Map layer: Heat vulnerability index (HVI) under Findings.

Summer 2025 index. The HVI has not been recomputed for 2026 — two of its five inputs are unavailable for that summer — so the scores and the district anomalies quoted here are the 2025 values. The hot-and-dense screening it feeds has been redone for 2026; see RQ-K.

Method

Composite HVI weights:

• UHI anomaly — 30%
• Elderly share — 25%
• Population density — 20%
• NDVI gap — 15%
• Heatwave days — 10%

Headline numbers

• Max HVI: 0.762 — Nadbiskup Antun Bauer (Donji Grad)
• Citywide mean HVI: 0.417 across 218 MOs
• 32 high-priority hot + dense neighbourhoods
• 12 of those top-priority MOs are in Donji Grad

Top neighbourhoods (illustrative)

1. Nadbiskup Antun Bauer — HVI 0.76, ΔT +9.1 °C, elderly 29%, density ~20.5k/km²
2. Pavao Šubić — HVI 0.75
3. Travno (Novi Zagreb — Istok) — HVI 0.74
4. Hrvatski narodni vladari — HVI 0.74, elderly 46%
5. Matko Laginja — HVI 0.73, density ~24k/km²

Note

MOs without direct Landsat NDVI sampling use an NDVI gap estimated from UHI anomaly (regression ΔT ≈ 14.7 − 18.7×NDVI, R²=0.77). Ranking is robust; absolute gap values are refined in RQ-C.

Open related layers on the map

RQ-B — Sensitive Facilities in Hot Neighbourhoods

Schools, kindergartens, healthcare, sport sites and fountains inside the 32 hot + dense mjesni odbori.

Map layer: Sensitive facilities under Vulnerability (points in hot / high-HVI MOs highlighted).

Counts are against the summer 2025 hot-and-dense frame of 32 neighbourhoods. That frame is 31 in 2026 — four Stenjevec neighbourhoods left, three in Peščenica – Žitnjak entered — so these facility totals have not been recounted for the newer set.

Method

Facilities from City of Zagreb GeoHub / ArcGIS (pharmacies, kindergartens, elementary and high schools, sport facilities, drinking fountains) assigned to nearest mjesni odbor (≤4 km).

Headline numbers

• 242 facilities inside the 32 hot + dense MOs
• 62 health / pharmacy
• 71 kindergartens
• 32 elementary schools
• 29 high schools
• 29 of 96 high schools citywide (30.2%) — the most over-exposed facility type, against 21.3% of kindergartens and 18.1% of elementary schools
• 30 / 32 of those MOs have no drinking fountain — but see the caveat below

The fountain figure needs context

The city dataset contains only 10 drinking fountains across all 17 districts, two of them inside the 32 hot MOs. “30 of 32 have zero” is therefore as much a statement about dataset coverage as about a deficit targeted on hot neighbourhoods — Zagreb has more public water points on the ground than this source records. The defensible version of the finding pairs it with need: 20 of the 32 hot-and-dense MOs have no water point within a 10-minute walk and an above-average elderly share, which is a concrete siting shortlist rather than a coverage artefact.

Why it matters

Heat risk is not only rooftop temperature — it is also where children, patients, and outdoor users concentrate on hot days. Use the dashboard to overlay facilities on HVI or summer LST.

Open related layers on the map

RQ-C — Tree-Cover Deficit and Cooling Potential

NDVI gap against a realistic greening target, and the modelled cooling if that gap were closed.

Map layer: Tree cover deficit under Greenery (all 218 neighbourhoods).

Method

Compare current NDVI to a realistic greening target — Stenjevec-style district NDVI of 0.422 — and convert the gap into modelled cooling using the study’s regression, ΔT = 14.7 − 18.73 × NDVI (R² = 0.77).

The dashboard layer now covers all 218 mjesni odbori. The original research table still emphasises the most heat-exposed neighbourhoods; the map no longer hides the rest of the city.

What the map shows

• Realistic NDVI gap, expressed as a percentage (max(0, 0.422 − district_mean_NDVI) × 100)
• Zero means the district is already at or above the target (not “no heat”)
• Modelled cooling if the gap closed: ≈ 18.73 × NDVI gap (°C)

Headline numbers (citywide layer)

• 218 neighbourhoods mapped across 17 districts
• Largest gaps sit in the least green districts (notably Donji Grad and Trešnjevka – sjever on the latest summer sample)
• Greener outer districts sit at 0% deficit on this target

Two limits on how far this can be read

NDVI is measured per district here, not per neighbourhood. Every neighbourhood in Donji Grad shares one NDVI value; within a district the map does not separate local greenness. Ranking districts is reliable; ranking neighbourhoods inside a district is not a vegetation ranking.

The hectares-of-canopy figures from the first public draft were withdrawn (dimensional error). The layer now reports neighbourhood area and a simple gap × area order-of-magnitude field — not a validated planting target. A properly calibrated NDVI-to-canopy conversion remains outstanding.

Planning takeaway

Use the layer to see which districts sit below the greening target, then pair with heat, elderly share, and fountain/refuge access when shortlisting sites. Dense historic-core neighbourhoods have far less room for new canopy than large outer MOs even where the modelled °C gain looks similar.

Open related layers on the map

RQ-D — Cool Refuge Accessibility

Walk time from high-HVI neighbourhood centroids to the nearest cool refuge (park, shade, indoor cool space).

Map layer: Cool refuge access under Greenery.

Walk times cover the summer 2025 hot-and-dense frame of 32 neighbourhoods; that frame is 31 in 2026. The access figures themselves do not vary by summer.

Method

For priority high-HVI neighbourhoods, measure network / crow-fly access to cool refuges and green areas. Points on the map are MO centroids coloured by minutes to the nearest refuge.

How to read categories

• ≤5 min — good access
• 5–10 min — moderate gap
• >10 min — priority for new cool spaces or shaded routes

This one does not change by summer

Walk times are computed from fixed neighbourhood centroids to fixed amenity locations, so RQ-D carries no temperature input and is identical for 2025 and 2026. It does not need a per-year re-run; only a change to the park, refuge or fountain datasets would move it.

Planning takeaway

Even well-greened cities fail residents if cool spaces are not walkable during heatwaves. Pair this layer with sensitive facilities (RQ-B) to see where schools and clinics lack nearby refuge.

Open related layers on the map

RQ-E — Multi-Year UHI Trend 2015–2025

Sentinel-2 NDVI greening slopes by district, plus multi-year LST context for Donji Grad and comparison districts.

Map layer: NDVI greening trend under Greenery.

Method

Summer (Jun–Aug) mean LST (Landsat) and NDVI (Sentinel-2 from 2017 onward). District slopes show where canopy is improving or declining over time.

What the map shows

District polygons coloured by ndvi_slope (NDVI change per year). Positive slopes = greening; negative = browning / densification pressure.

Planning takeaway

Use together with tree-cover deficit (RQ-C): a district can still be hot today while slowly greening — or already cool but losing canopy. Investment tables for 2017–2025 greenery spend sit in the analysis outputs for follow-up reporting.

Open related layers on the map

RQ-F — Lived Heat Burden (Citizen Survey)

What 7,329 residents report about coping with heat, and where their experience diverges from the satellite maps.

Map layer: Citizen survey heat burden under Vulnerability & census.

Method

City-run heat survey, hand-cleaned export of October 2026: 7,329 respondents across all 17 districts (smallest sample Brezovica, n=34; largest Trešnjevka – jug, n=751). The previously published figures were on the 7,289-response export; the 40 added responses move no headline by more than 0.2 points. 110 questions cleaned and scored into a heat-burden index across 14 life domains, 18 coping behaviours, 8 climate attitudes and 28 proposed interventions. Aggregated to district and mjesni odbor level. 184 neighbourhoods have at least 5 respondents (cells below that are suppressed for privacy), with a median of 32 and 99 clearing 30 — so read the neighbourhood-level layer as indicative and the district figures as the reliable cut.

All district rankings and correlations below exclude Brezovica (n=34) — the next-smallest district sample is 143, and a 34-response district acts as a leverage point on a 16-point correlation.

Headline numbers

• 67% find hot weather “hard” or “very hard” to cope with
• 75% avoid going outdoors at peak heat “often” or “always”
• 91% agree the City should act on climate adaptation
• 85% report a working air-conditioning unit
• 38% report difficulty covering monthly household costs

Heat costs people the outdoors, not the indoors

The hardest-rated life domain citywide is outdoor leisure away from home (3.35/4), then sport and recreation (3.33) and leisure in the neighbourhood (3.29). Being at home rates 2.54–2.68 and being at work only 2.32. Residents report losing going out, not staying in — which points at street shade, shaded routes, and public space rather than building retrofit alone.

Where the survey agrees with the satellites

Moderate-to-strong across the board: the survey is measuring something real, and physical exposure is a reasonable first-pass proxy for lived burden — but not a substitute for it.

The burden ranking is flatter than it looks

Across the 16 reliable-sample districts the burden index spans only 2.77 to 3.07 (mean 2.96). Fifteen of the 16 sit between 2.93 and 3.07. Donja Dubrava now tops the list at 3.07 — 0.005 ahead of Peščenica – Žitnjak and 0.011 ahead of Trešnjevka – sjever (which topped the ranking at wave 1) — 0.10 above the city mean. Do not treat that ordering as a priority list.

The survey question with real discriminating power is street shade, which spans 2.37 to 3.57 (mean 2.99) — four times the spread of the burden index:

Policy support is lopsided

Tree measures are close to unanimous — tree-lined paths 4.83/5, greening squares and roundabouts 4.82, preserving existing trees 4.82, greening school yards and planting/maintaining trees both 4.81. Anything that removes parking is far more contested: temporary parking greening is the single most-opposed measure at 3.49/5, with parking-to-pocket-forest conversion at 3.78. Expect friction on that specific lever, not on planting.

Note

The sample skews younger and better-connected than the city. Respondents aged 65+ are 509 of 7,329 (6.9%), against 21% of the actual population — the survey under-samples exactly the group the rest of this analysis identifies as most at risk. Self-reported AC (85%) also sits far above the census-modelled citywide figure (~60%). Read the survey for pattern and preference, not for population-level prevalence.

Open related layers on the map

RQ-G — Adaptive Capacity Versus Exposure

Education, income proxy and air-conditioning prevalence against heat exposure — testing whether the hottest districts are also the least able to cope.

Map layer: Higher education share under Vulnerability & census. Air-conditioning prevalence and the composite socioeconomic index are district-level tables, not yet published as map layers.

Method

Census 2021 education, occupation and unemployment per district, plus a modelled net-wage estimate and a modelled air-conditioning prevalence, combined into a composite socioeconomic index (0–1) for each of the 17 districts. Cross-referenced against this study’s own district UHI anomaly and against the citizen survey (RQ-F).

The wage column is a modelled proxy on a 2023 basis, derived from education structure — not measured income. Read it for ranking districts against each other, never as current absolute pay.

The counter-intuitive result

Richer districts in Zagreb run hotter, not cooler — socioeconomic index against district UHI anomaly gives r ≈ +0.40 (n=17). The mechanism is visible in the data rather than hidden in the correlation: the affluent historic core is the dense, low-vegetation fabric that produces the heat island.

A “target the poorest districts” framing would miss the two hottest districts in the city. This does not overturn the study’s equity findings — it means income is a separate, complementary dimension, not a substitute for the exposure-and-demography lens in RQ-A.

Two districts are hot and under-equipped

AC prevalence tracks heat exposure citywide (r ≈ 0.70) — the hot, dense core also has newer housing stock. Two districts break the pattern: Peščenica – Žitnjak (ΔT +5.8 °C in 2026, AC 58.7%) and Donja Dubrava (ΔT +4.2 °C, AC 56.2%) sit above the citywide heat threshold with below-average cooling capacity, and rank #13 and #17 of 17 on the socioeconomic index. Three independent signals — exposure, economic standing, and home cooling — converge on the same two districts.

District averages hide the households that matter

Affluence at district level does not mean affluence at household level, and this is where a “Donji Grad is rich, so it can cope” reading breaks down:

At the district level, Donji Grad’s own household income-hardship rate (37.6%) now tracks the citywide mean (38.0%) almost exactly — at wave 1 it ran slightly above it. The subgroup that matters is still there: Donji Grad’s residents aged 65+ report hardship at 47.5%, ten points above the district’s own average, alongside the lowest AC ownership and the highest vulnerability score of any Donji Grad subgroup. The affluent hot core still contains a least-buffered subgroup — the district-wide average just isn’t the number that reveals it; the age cut is.

Podsljeme: cool and green does not mean wealthy

Podsljeme is the coolest district in Zagreb by a wide margin (−3.6 °C, NDVI 0.81) and ranks only #10 of 17 on the socioeconomic index — solidly mid-table, with the second-lowest modelled air-conditioning prevalence in the city (50.3%). Living somewhere environmentally desirable and being affluent are not the same thing here. Worth remembering before assuming that green, cool neighbourhoods are the well-off ones, or that the affluence–heat correlation runs cleanly in every direction.

Planning takeaway

Keep RQ-A’s exposure-and-demography composite as the primary targeting lens. Use capacity as a second filter that changes what you build, not where: in high-capacity hot districts the gap is public-realm shade for the people who cannot retreat indoors; in low-capacity hot districts (Peščenica – Žitnjak, Donja Dubrava) the gap includes cooling capacity itself.

Open related layers on the map

RQ-H — Which Lever to Pull, Where

Heat Risk Population Index and geographically weighted regression across all 218 neighbourhoods, splitting risk into climate-driven and demography-driven, plus surface-versus-air decoupling.

Map layers: Heat Risk Population Index (HRPI), GWR local β, GWR local R², Heat-risk dominant driver and Thermal-risk intervention priority under Vulnerability & census; Surface–air coupling zones and LST vs TAIR hot-spot overlap under Baseline.

Method

Adaptations of Bečić & Gašparović’s Atmosphere 17, 466 and Earth 7(3), 72 models to Zagreb’s 218 mjesni odbori, using Census 2021 rather than the papers’ 2001–2024 change series. HRPI is the mean of four min–max normalised components per neighbourhood: 2 m air temperature, land-surface temperature, demographic sensitivity (DSI), and population density. A geographically weighted regression (27-neighbour adaptive kernel) then estimates, locally for each neighbourhood, how much of its heat risk is attributable to demographic sensitivity rather than to the climate signal itself.

Separately, Getis-Ord Gi* identifies statistically significant surface- and air-temperature hotspots, and a coupling classification flags where the two diverge.

HRPI distribution — all 218 neighbourhoods

Mean 0.512, median 0.515, range 0.141 (Gračani) – 0.756 (Hrvatski narodni vladari). This part of the model covers every neighbourhood, and it agrees directionally with the RQ-A vulnerability index: Hrvatski narodni vladari tops both, and Podsljeme’s hillside neighbourhoods anchor the bottom of both.

The intervention-targeting columns — read the coverage caveat first

The dominant driver and priority class attributions are the reason this analysis exists: read together they say which lever to pull, where. A critical hotspot that is demography-driven needs interventions aimed at the people already living there — shade, water, cooling refuges, outreach timed to the local age structure. A climate hotspot that is climate-driven needs the vegetation and surface work the rest of this study argues for.

But both attributions come from the GWR, and the GWR fits only 192 of the 218 neighbourhoods. The reason is upstream of the model: those 26 neighbourhoods have no Landsat land-surface sample in summer 2025, so they have no LST value, no local regression coefficients, and no fit. Until September 2026 all 26 were labelled climate-driven by default rather than left unclassified, which inflated that class by nearly half:

The 21 critical hotspots and the 34 demographic-sensitive neighbourhoods are all fitted, so those two classes are usable as they stand. The climate-hotspot class is 29, not 34.

The data now carries the fitted figures. The classification was corrected in September 2026: unfitted neighbourhoods return no label, they are flagged as unfitted, and they render in neutral grey on the dashboard rather than taking the colour of a class they were never assigned to. On that basis roughly six in seven neighbourhoods with a usable local fit have their heat risk attributed mainly to demographic sensitivity rather than to the climate signal.

What this cannot support

The 26 unfitted neighbourhoods are not peripheral — they include Andrija Medulić (Donji Grad), August Cesarec (Gornji Grad – Medveščak), Cvjetnica (Trnje), Donje Svetice and Folnegovićevo naselje. They are disproportionately inner-city, which is what a Landsat sampling gap in the dense historic core would produce.

Until September 2026 they also carried no district label — a merge artefact, since the district was present in the source all along under a different column. That made any district-level breakdown of drivers unusable: crosstabs silently dropped 26 central neighbourhoods and gave the rest a default label. An earlier reading of this data concluded that Donji Grad has no climate-driven neighbourhoods; that was an artefact, not a finding. The district join is now repaired, but the underlying sampling gap remains: those 26 neighbourhoods still have no local fit, so a driver breakdown covers 192 of 218 and should say so. Recovering LST for the inner-city core is the outstanding work.

Surface versus air: exactly one hotspot on both

This part uses a separate model and covers all 218 neighbourhoods. Of them, 21 are significant surface-temperature hotspots and 13 are significant air-temperature hotspots. Exactly one — Rudeš, in Trešnjevka – sjever — is a significant hotspot on both at once, matching the ≈0.5% share the source paper found.

The actionable class is the 20 surface-hot-only neighbourhoods: surface-hot but air-decoupled, where the satellite sees a hot surface that the local air model does not. Surface treatments — cool roofs, lighter pavement, shading a specific hot surface — are likely to matter more there than general ambient cooling. They include Antun Mihanović, Dr. Ante Starčević, Ciglenica and Donja Kustošija (Trešnjevka – sjever / Stenjevec), Kralj Zvonimir and Kralj Petar Svačić (Donji Grad), and Marin Držić and Martinovka (Trnje) — largely the same inner-core neighbourhoods RQ-A already flags, but reclassifying why.

Coupling zones citywide: coupled 124 (57%), air above expected 51 (23%), air below expected 43 (20%). Citywide surface–air correlation is r = 0.589, close to the source paper’s ≈0.54.

Caveats

• GWR local fit: mean R² 0.827, median 0.893 across the 192 fitted neighbourhoods. Where local R² is low the specification does not explain local variation well, and the priority class should not be used for targeting at those specific locations.
• HRPI itself covers all 218 neighbourhoods, including the 26 without an LST sample — so for those, the index rests on three of its four components. Treat their HRPI values as less certain than the rest.
• Air temperature comes from a ~9 km reanalysis grid, so “decoupled” partly reflects model resolution rather than only physical decoupling.
• Census 2021 is a single snapshot; the source papers used a 2001–2024 change series, so this is a cross-sectional adaptation, not a replication.
• HRPI and the RQ-A index are two legitimate but different ways of ranking the same 218 neighbourhoods. They agree at the extremes and have not been reconciled in between.

Open related layers on the map

RQ-I — What the Tree Cadastre Can and Cannot Tell Us

162,676 individually registered trees joined to districts: a maintenance register rather than a canopy layer, and a stock too young to shade much yet.

Map layers: Green Cadastre (Trees) and Green Cadastre (Tree Density) under Greenery.

Method

Zagreb’s zeleni katastar records 162,676 individual trees across 240 species, each with a height class, crown and trunk diameter class, and a coordinate. Every tree was joined to its city district by point-in-polygon — all 162,676 matched, none dropped — then normalised by district area and compared against this study’s own heat anomaly and NDVI.

The register is not a canopy layer

Cadastred-tree density correlates positively with heat and negatively with greenness:

The hottest, least-green districts have the most registered trees. Podsljeme — the greenest (NDVI 0.81) and coolest (−3.6 °C) district in Zagreb — holds just 577 cadastred trees, 10 per km², of which 85% are under 5 m. Brezovica has 745. Trnje has 2,287 per km².

There is no contradiction: the cadastre registers municipally managed street and park trees, so it maps where the city plants and waters, not where canopy exists. Podsljeme’s cooling comes from hillside forest and private gardens that no municipal register contains.

This settles an open question. The project had wondered whether the cadastre was already the canopy-cover layer it planned to request from the city’s geoportal service. It is not — and substituting it for canopy cover would invert the relationship the rest of this study measures. What it is good for is different and still valuable: it is the inventory of the trees the city already owns, and therefore of what can be protected, pruned, watered through a heatwave, or replaced.

The stock is young

Only 15.6% of registered trees — 25,371 of them — are 16 m or taller, and over a third are saplings under 5 m. Shade scales with crown size, so a third of the register delivers almost none today. The citywide mean height is 9.0 m.

Where the mature street canopy actually is

Two things follow. First, the inner ring already has the city’s densest mature street canopy and is still the hottest part of Zagreb — street trees alone have not closed the gap in dense historic fabric, which points at total canopy fraction and built form rather than tree count. Second, Stenjevec is the clearest anomaly: 934 trees per km², the fourth-densest register in the city, but only 4.0% mature and a mean height of 6.9 m. It has been planted recently and at scale, and will deliver far more shade in fifteen years than it does now.

Planning takeaway

Use the cadastre for stewardship, not for measuring canopy: which trees to protect from removal, which to prioritise for watering during a heatwave, and where recent planting still needs a decade to pay off. Measure actual canopy with NDVI (RQ-C) or a dedicated canopy layer. And treat “we planted N trees” as a statement about future shade — with 35% of the register under 5 m, planting counts and delivered cooling are years apart.

Open related layers on the map

RQ-J — Where Every Line of Evidence Agrees

Counting how many of six independent axes flag each district, and what changes when heat is no longer the only one.

Uses the layers behind RQ-A, RQ-F, RQ-G, RQ-H and the heatwave module together.

Method

This study now has six independent axes of evidence, built from different data by different methods:

1. Daytime heat — district anomaly at or above the citywide hot threshold (+5.8 °C in 2026), from Landsat
2. Heatwave days — the four districts with the most tropical-night heatwave days (RQ11)
3. Lived burden — the five worst self-reported burden districts among those with a reliable survey sample (RQ-F)
4. No street shade — the five worst street-shade ratings in the survey (RQ-F)
5. Low socioeconomic standing — the five lowest on the composite index (RQ-G)
6. Low cooling capacity — below-average air-conditioning prevalence while above the baseline for heat (RQ-G)

Plus a seventh marker: two or more HRPI critical hotspots inside the district (RQ-H).

Each district is scored by how many of these place it in the worst tier.

This counts breadth, not severity. A district with one flag can still carry the most extreme single value in the city — Donji Grad shows two flags and still holds five of the twenty-one critical hotspots. Some axes are also correlated: heatwave days and low AC coverage both track the eastern lowlands. Read the count as how many different kinds of evidence point here, never as a risk score.

The result (summer 2026)

What this changes

2026 lines up better than 2025 did. Trešnjevka – sjever moved from three flags to four and is now both the hottest district and the most broadly flagged — last summer those were different places. Peščenica – Žitnjak gained the daytime-heat flag as it entered the hot five.

But two of the four most-flagged districts still sit outside the daytime ranking entirely: Donja Dubrava at +4.2 °C and Sesvete at 0.0 °C. Both dropped from five flags in 2025 to four in 2026, and both stay in the top tier on evidence that has nothing to do with daytime surface heat.

The reason is mechanistic: the eastern lowlands lose heat slowly at night (RQ11’s tropical nights), have the weakest cooling capacity and the lowest incomes (RQ-G), and their residents report among the highest lived burden (RQ-F) — while their daytime surface temperatures are unremarkable because the fabric is lower-density.

Two distinct priority cases, not one ranking:

• Intensity, inner ring. Donji Grad, Trnje, Trešnjevka – sjever and Stenjevec carry the extreme daytime anomalies and 13 of the 21 HRPI critical hotspots between them.
• Breadth, eastern lowlands. Peščenica – Žitnjak, Donja Dubrava, Sesvete and Gornja Dubrava are flagged by more axes each, driven by night-time heat retention, low adaptive capacity and lived experience.

Trešnjevka – sjever is the only district in the top tier of both — a stronger argument for its pilot status in 2026 than it was in 2025.

The four unflagged districts

Maksimir, Gornji Grad – Medveščak and Podsljeme clear all six axes in 2026. Podsljeme is genuinely the cool green reference. The other three warrant a caveat rather than an all-clear: Gornji Grad – Medveščak has four neighbourhoods above 30% elderly share, 7,563 residents aged 65+, and declining greenness, District-level axes miss neighbourhood-level concentrations — see the coverage caveat in RQ-A.

Planning takeaway

Report both cases and say which one a given intervention serves. Daytime-surface measures — shade, cool pavement, canopy — belong in the inner ring where the anomaly is extreme. Night-time measures — permeable surfaces, green roofs, reducing stored heat — plus cooling-capacity support belong in the east, where five of six axes converge and the daytime map shows almost nothing.

Open related layers on the map

RQ-K — Summer 2025 Versus Summer 2026

Both measured summers side by side: 2026 was warmer in the air, flat at the surface, and its extra surface heat landed outside the historic core.

Dashboard: most heat layers carry a Summer year toggle (2025 / 2026), defaulting to 2026.

Which numbers are comparable, and which are not

Three district-level figures for 2025 are in circulation. They differ because of sample basis, not because the weather was re-measured:

None of these is “wrong.” The sample gives Donji Grad only 12 pixels, so its anomaly carries a 95% bootstrap confidence interval of [+8.23, +9.88] °C — and the originally published +9.64 sits inside it. The two valid figures are statistically indistinguishable; the difference is sampling noise on a very thin district, not a change in the weather or a corrected error. The site was moved onto the 8,000-pixel basis in September 2026 so that the narrative, the charts and the dashboard all agree and 2025 can be compared with 2026 directly. Quote it as “about +9 °C”.

For scale, the sampling intervals on the other hot districts:

The ranking is robust — Donji Grad and Trešnjevka – sjever are clearly the two hottest in 2025, and clearly ahead of Trnje. The second decimal is not, and any statement that two districts differ by less than about half a degree is inside the noise.

The third value briefly appeared in the 2025 summary data, but it had been computed from the 2026 intermediate rather than the 2025 one, and that intermediate reconciles with neither summer’s measurements (about 2.4 °C from both). It was a sourcing error rather than a corrected measurement, and it also put Maksimir into the 2025 hot-district set in place of Stenjevec. The 2025 figures have since been rebuilt from the 8,000-pixel 2025 sample using the identical method as 2026 — anomaly against the rural-class mean, hot threshold at the 75th percentile — and the rebuild restores the originally published five hot districts.

Everything below uses the matched 8,000-pixel basis: the two summers share 7,982 of 7,988 sample locations, so the same ground is differenced and Donji Grad rests on the same 12 pixels in both years.

2026 was an air event, not a surface event

Air temperature rose in every one of the 218 neighbourhoods, by +0.82 to +1.47 °C. Surface temperature barely moved, and only 58% of pixels were warmer. The gap between what satellites see and what people breathe closed by about a degree.

That has a delivery consequence: a summer hotter in the air but flat at the surface rewards shade, ventilation and cool refuges more than albedo and pavement treatments.

The heat island narrowed — which is not progress

Built-up-minus-rural contrast fell from +6.42 to +6.19 °C, because the cool surfaces warmed far faster than the built ones:

Water warmed nine times as much as built-up. The rural reference itself rose +0.36 °C. A falling UHI intensity here means the cooling reservoir is losing ground, not that the heat island is weakening — anyone quoting it as an improvement has the sign backwards.

The extra surface heat landed outside the historic core

Trešnjevka – sjever (+8.27 °C) is the hottest district in 2026, overtaking Donji Grad (+8.13 °C). The historic core’s surface actually cooled slightly, while Novi Zagreb – istok (+1.45 °C), Peščenica – Žitnjak (+1.28 °C) and Novi Zagreb – zapad (+1.05 °C) took the warming.

This is the same direction the rest of the study now points in. Those are lower-density, more vegetated districts — consistent with the land-cover breakdown above — and Peščenica – Žitnjak is already flagged on four independent axes in RQ-J. The 2026 data strengthens the case that the eastern and southern districts, not only the historic core, belong in the priority frame.

The hot-district set is a moving target

Membership is a within-year 75th percentile, so a district can cross the line without its own temperature moving much — and the threshold itself rose from +5.59 to +5.80 °C. Stenjevec left; Peščenica – Žitnjak joined. The 2025 set is the same five originally published, which is a useful independent check on the rebuild described above.

Heatwaves: the sharpest change of all

Recomputed for both summers on one basis — station-calibrated ERA5-Land at all 218 neighbourhood centroids. The 2025 figures reproduce the originally published values, which validates the method.

Median heatwave days more than doubled, tropical nights rose by half, and every neighbourhood had at least a week of heatwave conditions in 2026 against five with none in 2025.

This is the finding that ties the rest together. A summer that is +1 °C in the air, flat at the surface, and carries 50% more tropical nights is a nocturnal heat event — which is exactly what the surface−air gap closing by 0.97 °C describes. It also reinforces RQ11’s original conclusion: heatwave burden is set by night-time cooling capacity, and it concentrates in the eastern lowlands.

One caveat on the method: the station bias correction is year-specific by design (+0.91 °C for 2025, +0.53 °C for 2026), so it is calibration rather than a uniform shift. On the uncorrected ERA5 basis the direction is identical and the relative gap is wider still (median 7 → 23 days).

Two year-flips that look dramatic and are not

• Elderly heat-exposure scores fell (0.180 → 0.170; night version 0.145 → 0.089) while air temperature rose everywhere. These are normalised within each year, so the map can look cooler in a hotter summer. Always read them beside an absolute layer.
• Night surface temperature appears to jump +1.8 °C, but rests on 2 usable ECOSTRESS scenes in 2025 against 6 in 2026 — different nights, not two comparable summer means.

Planning takeaway

Two summers is a difference, not a trend. Report 2026 as warmer in the air, flat at the surface, greener nowhere, and warming fastest outside the historic core — and keep the 2025 column alongside it. For direction over time, the multi-year series in RQ-E remains the right instrument.

Open related layers on the map