SmogRank / How we rank

How we rank cities

Correct beats complete: a city is ranked only when enough good measurements back its number.

We rank cities by PM2.5, the fine particles in the air most tied to harm to health. Below: which monitors count, how a city gets its number, and the limits. The pollutant and the scales come last.

Which monitors count

  • Reference monitors first. Government and research stations with regulatory-grade instruments. When a city has any, only they set its value.
  • Low-cost sensors count only in cities with no reference monitor and at least three sensors, with at least two (and a third of them) reporting in an hour. These cities carry a "low-cost sensors" label. We measured how they compare: over 30 days to 6 October 2026, 248 low-cost sensors within 5 km of a reference monitor read a median 5% below it, but in humid air (relative humidity of 70% or more) AirGradient sensors read 15 to 25% high. (by maker, humidity and level) We don't correct for this: the US EPA's published correction for these sensors left them about a third below the reference monitors next to them. A sensor placed indoors reads low. So we leave out sensors that are indoor models or labelled as indoors, and any that average under 35% of the nearest reference monitor (within 50 km) over the same hours of the last three days while that monitor reads 8 µg/m³ or more. Many low-cost sensors also read far too low in clean air, so we check every one against our model estimate for its city: a sensor whose last three days average under 35% of the estimate for the past week (when the estimate is 2.5 µg/m³ or more) is left out, and a low-cost city whose own value does the same isn't ranked and shows the estimate instead. The reverse is a fault too: a low-cost sensor averaging 100 µg/m³ or more over the last three days at ten times or more what the monitors around it read in the same hours (its city's other monitors and the nearest reference monitor within 50 km) is left out. Real smoke raises the monitors around it as well. The 7- and 30-day boards make the same check on the days they average: a low-cost city under 35% of the estimate for that span is left off that board. Applied to cities with reference monitors, that rule would catch about 0.7% of them. And low-cost cities are never shown as the cleanest (cleanest air), nor is a city averaging under 1 µg/m³ there: few monitors measure that low, and a mean so near zero is more often a fault than air that clean.
  • Sensor.Community. We also read the hourly file of Sensor.Community, a network of low-cost sensors that people put up themselves, about 9,400 of them outdoors, most in Europe and most of them the Nova Fitness SDS011. Its database is under the Open Database License (ODbL), which asks anyone who builds a database from it to share that database on the same terms. So we keep its sensors apart from every other source: a value never mixes the two. Reference monitors still come first wherever a city has one. Its sensors go through every check above: three or more in a city, the comparisons with the nearest reference monitor and the model estimate, and the stuck, frozen and spike rules below. We use only sensor models that report PM2.5 in µg/m³, leave out the ones marked indoors, and file each reading at the hour the file was made. Every month we publish the values its sensors set, each city-hour and city-day, in ranked cities and in the cities with only one or two sensors alike, with the sensors we left out and why, under the ODbL (the files); this page is the method by which those values are made from Sensor.Community's readings. We measured them as we did the other low-cost sensors: over 30 days to 12 October 2026, 231 of its sensors within 5 km of a reference monitor read a median 44% below it (the 170 that pass the checks above, 30% below; the SDS011 about half), in dry and humid air alike, where the other low-cost sensors read 5% below. We don't correct for it. So its sensors set a city's value only where nothing else can: in a city with no reference monitor and fewer than three other low-cost sensors. Where other low-cost sensors could rank the city, they do, and Sensor.Community's aren't counted. A city ranked on its sensors likely had dirtier air than its number says.
  • Official and open data, credited to who measured it. Every source is in one of three tiers. Open: the publisher's licence allows commercial use and redistribution without share-alike terms, or, where our data source lists no licence, the publisher's own terms allow it (India's GODL, for example). Where a government publishes the same readings under a more open licence than our data source lists (the ACT and Queensland, CC BY 4.0), we go by the publisher's own licence. Sensor.Community's share-alike licence we meet on its own terms (above). Public, no terms: a government network, or a named private or non-profit publisher, outside the EU and the UK that publishes its readings with no reuse terms for them (a site-wide copyright notice alone, such as "All Rights Reserved" in the footer, covers the site's own text and design, not the readings), checked on the publisher's own site first: Vietnam's, Iceland's, Nepal's (the Department of Environment), Uzbekistan's (Uzhydromet, its hydrometeorological service), Kuwait's (its Environment Public Authority) and Bogotá's (RMCAB, the city's Secretaría Distrital de Ambiente) government monitors, Thailand's Air4Thai monitor in Laos, the low-cost sensors of the City and County of Denver and of India's Central Pollution Control Board, and Aurassure's sensors in Burkina Faso and Nigeria. A network with no publisher named stays out: there is no one to credit or to ask us to stop. We use these readings, credit the publisher by name with "no reuse terms published" wherever they show, read them gently, and take them out within a day if the publisher asks (how). Out: data whose terms forbid reuse, EU and UK publishers that reserve their rights, and anything behind a block or a captcha. A login counts as a block unless the publisher gives a free account to anyone who asks and we use it only as its own pages do (São Paulo's CETESB). Where a network stopped reaching our data source, we read it from its publisher under the same rule: the UK's national network comes straight from Defra's UK-AIR (Open Government Licence v3.0), from 1 January 2026 on, and its days from 2023 to 2025 from UK-AIR's own archive of hourly data, in place of the copy our data source carried, which sometimes held one value for a whole day. Where our data source is missing hours that AirNow has for the same US monitors, we take those hours from AirNow's own hourly files. In the same way, where it is missing hours that the European Environment Agency has for the same European monitors (our data source relays them from the EEA), we take those hours from the EEA's own up-to-date data (CC BY 4.0), and those monitors' earlier days from the EEA's verified and up-to-date files, which hold hours our data source never had (verified values can differ a little from the preliminary ones first published). sensors.AFRICA's low-cost sensors in Kenya, Nigeria and Ghana come from its own hourly file (CC BY); their days before we started reading it, from 2023, come from its monthly archives on openAFRICA (only the cities whose archive is published under CC BY), each sensor's raw readings averaged into hours of six or more readings, then days checked by the same rules as every other history day. A sensor that has moved counts only for the readings taken at its present site. EPA Victoria's stations come from its own website (CC BY 4.0) from October 2026; before July 2026 their days are from the copy OpenAQ carried, and the weeks between have none. New South Wales's stations come from the state's own air quality data service (CC BY 4.0) from October 2026; they stopped reaching our data source in November 2022, so earlier days have none. EPA South Australia's stations in Adelaide come from its own hourly feed (CC BY 4.0) from October 2026, with no earlier days from them. Queensland's stations come from the state's own air data download (CC BY 4.0), each hour's mean, for every day from 1 January 2024, and their days in 2023 from the state's yearly hourly files on its open data portal (CC BY 4.0). Our data source carries the state's 24-hour running average in place of each hour, so we don't use its copy of these hours; an hour the state's download hasn't given stays empty. A station with no PM2.5 in a year's file keeps that copy's days for that year: daily means of a running average, close to the real ones but smoothed. The Australian Capital Territory's stations come the same way from the ACT Government's own air quality monitoring data (CC BY 4.0), each hour's mean, for every day: our data source carries its 24-hour rolling average in place of each hour too. The ACT publishes an hour that read below zero as 0, and we take it as it is. Bogotá's network (RMCAB) comes from its own website from October 2026, with no history before we started reading it. Serbia's network (SEPA, the Environmental Protection Agency) comes from SEPA's own open-data API (Serbian Open Data Licence) from November 2026; it keeps only 30 days, so earlier days are from the copy our data source carried until December 2025, and the weeks between have none. Where one of its stations is also a European Environment Agency station we already use, it is counted once. Brazil's state networks outside São Paulo (Minas Gerais, Rio de Janeiro city, Paraná, Espírito Santo, Bahia, Maranhão, Pará, Pernambuco, Mato Grosso do Sul, Rio Grande do Sul, Santa Catarina and the Federal District), gathered by the Ministry of the Environment's MonitorAr system, come from its yearly open data files (CC BY) for their days from 2023 to the end of its latest file (2025 for now). Its live service gives only an air quality index, not the concentration, so these stations have no later days. Only readings the network marks valid are used, error codes (985) are left out, and so is a station's day averaging under 1 µg/m³, a unit at fault. Catalonia's stations (the Generalitat's network, XVPCA) take their past days from the Generalitat's own open hourly data (its open licence, with credit), which holds every hour of the stations whose copies in our data source and in the European Environment Agency's files stopped or thinned out in 2026 (Barcelona's among them). Castilla y León's stations take the days of 2026 that their copies in our data source and in the European Environment Agency's files lack (most of February to May) from the Junta de Castilla y León's own open daily means (CC BY 4.0, with credit), which match those copies' days to within rounding where both have them. Its stations in Zamora, Segovia and Ávila, which we read hour by hour from the Junta from November 2026, take their earlier days from the same daily means (from late December 2025 to mid-August 2026, later for Segovia and Ávila), and Zamora's from March to December 2025 from the European Environment Agency's verified files. Berlin's stations take the days their copies in our data source and in the European Environment Agency's files lack (nine days in August 2026 and five in September) from the Berlin Senate's own hourly data (the Datenlizenz Deutschland, with credit), whose days match those copies' to within a tenth of a µg/m³ where both have them whole. Sweden's stations take the days their copies in our data source and in the European Environment Agency's files lack (most of Stockholm's from July to October 2026) from the national air-quality data host's hourly data (Naturvårdsverket and SMHI, CC0), whose hours match the European Environment Agency's to within a hundredth of a µg/m³ where both have them. Hong Kong's stations take the days their copy in our data source lacks (June and July 2026) from the Environmental Protection Department's own hourly readings as DATA.GOV.HK's archive keeps them (DATA.GOV.HK's terms of use, with credit). São Paulo state's network (CETESB) comes from CETESB's own data system, QUALAR, through a free account CETESB gives anyone, under CETESB's open-data terms (unrestricted use, with credit): each station's valid hourly means, its days from 2023 (from April 2023 where our data source carried the station until then). MonitorAr's copy of these stations isn't used: it places each at the centre of its municipality and has far fewer of their hours. Nepal's Department of Environment stations come from its own air quality portal from October 2026, with no history before we started reading them: one reading a minute, averaged into hours with 45 or more minutes, and an hour averaging under 1 µg/m³ left out (a faulty unit there reads zero or close to it). Because some of its units read near zero when faulty, its stations also get the low-cost sensors' check against the model estimate for their city. Uzhydromet's automatic stations in Uzbekistan come from its own monitoring website from October 2026, each station's latest hour as its page shows it, with no history before we started reading them; they get the same check against the model estimate, and an hour under 1 µg/m³ is left out. Kuwait's stations (its Environment Public Authority) come from the authority's own monitoring service, live and for past days, each station's hours as it publishes them in whole µg/m³; an hour under 1 µg/m³ is left out. Singapore's National Environment Agency publishes one hourly reading for each of the island's five regions (north, south, east, west and central) rather than each station's; we read them from data.gov.sg under the Singapore Open Data Licence, from 2023 on, and count the five regions as the city of Singapore's five monitors, so its hour is their median and low-cost sensors there are held as in any city with reference monitors. For South Korea, Thailand, Chile and South Africa we have asked the publishers to confirm that city averages with credit are fine, and show their monitors while we wait; if one says no, we take its data out. We don't read the government monitoring networks of China, Turkey, Iran and Georgia: they publish live readings only on their own websites. The networks of Turkey and Iran don't answer requests from outside their countries, so we can't read them. A public mirror of China's release platform gives us its national stations' hourly readings a day late, too late for live boards: past boards from January 2023 on rank its cities on them. Past boards from January 2023 to December 2025 rank Turkey's cities on its verified hourly readings, which the European Environment Agency publishes. Georgia's network publishes running 24-hour means, not hours, so we take its daily means as its agency publishes them, a day late, too late for live boards: past boards from January 2023 on rank its cities on them. On live boards, the few of their cities that appear are measured mostly by community low-cost sensors, tagged as such. They are not the only gap. Across most of Brazil, and in Indonesia, most of Vietnam, Egypt, Nigeria, Iraq and most of the Gulf states, official readings aren't published in a form we read, or aren't published at all, and few other monitors fill in. Most cities in these places are missing from SmogRank, not clean, and have a model estimate instead of a rank. How much of the world we measure. We would rather leave a city out than show data we may not use.
  • Mobile sensors and stations with no PM2.5 reading for 30 days are left out of current rankings, even when they still report other pollutants, and so are monitors that report only one PM2.5 value a day, which never makes an hour; a reference monitor that has since closed still counts in past rankings for the time it reported. Readings below zero or above 1,500 µg/m³ are treated as faults, and a monitor that reports the exact same value 24 times in a row is treated as frozen and left out until it changes. Any run of 18 or more identical readings in a row is left out of the hours it covers, so a monitor stuck at one value all day and another the next sets no city days. The exception is a reference monitor's run at 5 µg/m³ or less, the way some networks write clean air (Belgium's 0, Spain's whole numbers, the value Germany's states give a reading too low to detect): it counts while another of the city's monitors, or with none the measured cities within 60 km, read 6 or less over those hours. Any run of 6 or more readings of 100 µg/m³ or more that stay within 0.5 µg/m³ of the first is left out too: real air that dirty moves more than that from hour to hour, and a monitor pinned near its ceiling doesn't.

How sure each figure is

Every city on our boards and pages carries a small three-bar mark for how much stands behind its figure. It says nothing about how clean or dirty the air is, only how many independent instruments measured it.

  • Three bars, monitors 3 km+ apart: government-grade (reference) monitors at two or more locations in the city, 3 km or more apart, so no single location speaks for it. Monitors closer than 3 km count as one location. A board counts only the locations that reported for it: on the 24-hour board, those with 12 or more hours in the window; on a past day, those that reported that day (12 or more of its hours, or a full day in our history from before the hourly record); on a past week, month or year, the count on most of its days. A city whose other locations were silent shows two bars there.
  • Two bars, one monitor location: government-grade monitors at one location only, one monitor or a few side by side. Its readings are reliable, but they describe one part of the city.
  • One bar, low-cost sensors: three or more consumer sensors and no reference monitor. On the whole they read close to reference monitors, though AirGradient's read 15 to 25% high in humid air, and Sensor.Community's, which rank a city only where the others can't, read well below them.
  • An outlined bar, 1–2 sensors, not ranked: one or two low-cost sensors and no other monitor. They pass every check the other sensors do, and with two, an hour they read more than 2 times and 5 µg/m³ apart counts no value. One cheap unit can still be well off, so these cities take no rank. Each board can show them with a switch that is off by default: they then sit at their place by value, marked –, and ranked cities keep their numbers. Past weeks, months and years from 8 October 2026 have the same switch, with the board's minimum of days measured. They never appear on share cards, citations, past days, episodes or the race. Against the nearest reference city within 25 km, their 7-day values read a median 1.01 times it (73 cities, 14 October 2026). Where those one or two sensors are Sensor.Community's, the city shows with Sensor.Community's ODbL credit.

What a city is

Each monitor belongs to the nearest city in its own country from GeoNames' list of places with 15,000 or more people, so a monitor across a border or a river never counts for the city on the other side. A monitor's country is the one its coordinates fall in when the country it is filed under is across a border from it (two Californian stations filed under Mexico count for a Californian city, not Tijuana). A large city takes in its boroughs and suburbs within its reach (about 4 km for a town, up to 25 km for the largest cities), but never across a state or province line. A state or province that is one city, such as Seoul, Berlin, Bogotá or Tashkent, takes every monitor well inside its boundary and within 25 km of its centre, even where a suburb across the line is nearer, and so does a capital's own territory (Delhi, Mexico City, Washington DC) whatever the distance. A place inside such a city counts for it too rather than ranking on its own: Najafgarh, Narela and New Delhi count for Delhi, and their pages open with Delhi's reading. A city that is a whole country or territory of its own, Hong Kong, Macau or Singapore, takes every monitor in it, so the Hong Kong government's stations in Mong Kok, Tuen Mun and the other districts count for Hong Kong, and the districts' pages open with its reading. The reverse holds for a capital's territory: a monitor well outside it never counts for a place inside it, so Gurugram's and Ghaziabad's stations just over Delhi's line count on their own side. Where a station's own name names a nearby city of the same state or province, at least twice the size of the place the nearest-city rule picked, it counts for that city: "Teri Gram, Gurugram" counts for Gurugram, not the village of Bādshāhpur nearer to it. Near a capital's territory the name can also cross the line, since our outline of the territory is simplified: "Sector 125, Noida, UP" counts for Noida, not Delhi. Where a feed puts a station far from the city its publisher names, we check it by hand and place it in that city: Indore's Chhoti Gwaltoli station is filed 45 km out in the fields by Mānpur, and counts for Indore. A far name alone moves nothing, since it is often a metro's label for a station outside the city ("IITH Kandi, Hyderabad" is 42 km from Hyderabad, and stays where it is). When a rule like these moves a monitor, its past days move with it: the history is rebuilt so the place it left keeps none of them. A road or a building named after a city doesn't count as naming it, and a city keeps a monitor named after one of its own neighbourhoods. A government or research monitor that no such city reaches goes to a town of 5,000 or more nearby, which then gets a page and a place on the boards like any city; low-cost sensors only count for cities of 15,000 or more, since their checks lean on a model estimate we don't make for towns. Monitors more than 12 km from any such place in their country aren't ranked. A city carries the name its monitors' publisher gives it where GeoNames lists an older one: Chandrapur, not Chānda; Sagar, not Saugor. A few are set by hand from the city's own council or board: Sivasagar, not Sibsāgar; Kurukshetra, not Thānesar; and Tiruppur, as its council writes it. A place GeoNames doesn't list where a government monitor stands is added by hand, with its location and population from published sources, and is then placed like any other town: Byrnihat, the industrial town on the Meghalaya-Assam line, from Wikipedia's location and the Meghalaya pollution control board's 5,444 people for it and its neighbouring villages, so its station no longer counts for Dispur.

From readings to a ranking

  • Each hour: the city's value is the median of its monitors, if at least a third of them reported that hour (three are always enough). So when most of a city's network goes quiet, the one or two monitors still reporting don't speak for the whole city; its page says how many reported. With three or more monitors, a reading over three times the others' median (and 50 µg/m³ above it) is dropped, so one faulty sensor can't make a city the world's worst. At any number of monitors, a reading over 100 µg/m³ and more than four times the same monitor's readings in the hours on either side is dropped as a spike; a burst of up to six hours in a row over 100 is judged the same way, each of its hours against the readings either side of the burst. Where the monitor was silent just before or after, the side is its first reading across the silence (within a day), so a reading can't escape this by sitting next to a gap; a monitor's first readings after a day or more off are judged against the readings after them. A burst still going at the newest hour, with no reading after it yet, is held out while it is over four times the reading before it, until it ends (then judged as above) or passes six hours (then counted), unless a measured city within 50 km reads a third of it. Any of these stays when another of the city's monitors rose that hour too: real smoke, dust and fireworks last longer or show on several monitors. A monitor's sudden jump to five times its reading before or more, into a run over 100 of up to twelve hours that then ends, is left out too, when something else measured the air there and none of it rose to twice its own reading before the jump: another of the city's monitors (a low-cost sensor's calm readings can clear a reference monitor's run this way) or a measured city within 50 km, with no satellite fire detection within 100 km that day or the day before, fewer than 50 within 300 km, and no mapped smoke or volcanic ash. With nothing else measured nearby, the run counts. However long it lasts, an hour a single monitor puts over 100 µg/m³ and at ten times the city's usual or more (the median of its days over the last four weeks, or its typical reading at that hour of the day over the last 30 days when that is higher, so a town whose evenings are smoky isn't judged on one) is left out too when something else measured the air that hour and all of it read far less: the city's other monitors, of either kind, under a fifth of it, and the measured cities within 150 km under a third and none of them at four times its own usual (and 25 µg/m³) or more, with no satellite fire detections, mapped smoke, volcanic ash or model dust near, as above. It counts again as soon as a neighbour rises with it or that evidence arrives, and with nothing else measured nearby it counts. An hour left out this way stays out once it is older than the last 24 hours, in the city's charts and its daily means too, and a city page whose newest hours were left out says so instead of showing the hour before them as its latest. A newest hour still waiting for a reading after it (one monitor location, over 100 µg/m³ and over four times the hour before its run) is marked unconfirmed, and the health advice goes by the 24-hour average until it is judged. A monitor that reports the same value for a whole day is stuck (often at an error code such as 985), and is left out until its value moves again; in history, such a day is left out, and so is a monitor's day with any reading above 1,500 µg/m³, since one error code lifts the whole day's average.

One real day, worked through

Aguascalientes, Mexico, on 28 March 2026, local time. The city has three government monitors; each mark is one monitor's reading for one hour, in µg/m³. The line is the city's value for each hour: the median of the readings that count.

  • Instituto Educativo
  • Centro
  • Secretaría de Medio Ambiente, stuck: left outcounted
  • City, each hour
  • Day mean

35.7 µg/m³ Unhealthy for sensitive groups Both rules applied, as on SmogRank: the mean of 24 of 24 hours (18 needed).

050100150160+00:0006:0012:0018:0023:00day 35.7day 55.7city 506.6day 22.0day 22.0987.0 at 09:00, left out987.0 at 09:00, counted987.0 at 09:00, left out987.0 at 09:00, left out050100150160+00:0006:0012:0018:0023:00day 35.7day 55.7city 506.6day 22.0day 22.0987.0 at 09:00, left out987.0 at 09:00, counted987.0 at 09:00, left out987.0 at 09:00, left out
Chart of Aguascalientes's 72 hourly readings on 28 March 2026 and the city's value for each hour, from 23.5 to 89.5 µg/m³ with the rules applied; the day's mean is 35.7. The full readings are in the table below.
  • A spike. At 09:00, Instituto Educativo read 987.0, while Centro read 26.1. Hours over 100 in a row are judged together, against the same monitor's readings either side of the run: here 987.0 at 09:00 and 150.0 at 10:00, against 61.0 the hour before and 43.0 the hour after. An hour of such a run more than four times both, with no other monitor rising, is left out, so the city's 09:00 value is 26.1. Counted, that hour would have been 506.6 and the day 55.7. The 150.0 at 10:00 isn't four times them, so it counts.
  • A stuck monitor. Secretaría de Medio Ambiente read exactly 18.0 at every reading from 5 March to 28 April, 1,302 in a row. Real air doesn't hold one value for most of a day, so a run of 18 or more identical readings is left out: it sets none of this day's hours. Counted, it would have been the city's value for 8 hours and pulled the day to 22.0.
  • The day. The mean of the city's 24 hourly values (18 of 24 are needed): 35.7 µg/m³.
All 72 readings
HourInstituto EducativoCentroSecretaría de Medio AmbienteCity
00:0099.026.918.0 (stuck, left out)63.0
01:0076.023.118.0 (stuck, left out)49.6
02:0063.026.718.0 (stuck, left out)44.9
03:0048.019.418.0 (stuck, left out)33.7
04:0047.029.818.0 (stuck, left out)38.4
05:0038.023.918.0 (stuck, left out)31.0
06:0037.021.918.0 (stuck, left out)29.5
07:0039.018.518.0 (stuck, left out)28.8
08:0061.033.718.0 (stuck, left out)47.4
09:00987.0 (lone spike, left out)26.118.0 (stuck, left out)26.1
10:00150.029.018.0 (stuck, left out)89.5
11:0043.025.818.0 (stuck, left out)34.4
12:0037.016.018.0 (stuck, left out)26.5
13:0045.018.018.0 (stuck, left out)31.5
14:0037.019.018.0 (stuck, left out)28.0
15:0042.020.918.0 (stuck, left out)31.5
16:0037.015.018.0 (stuck, left out)26.0
17:0036.012.018.0 (stuck, left out)24.0
18:0033.014.018.0 (stuck, left out)23.5
19:0039.016.018.0 (stuck, left out)27.5
20:0039.015.018.0 (stuck, left out)27.0
21:0047.017.218.0 (stuck, left out)32.1
22:0037.028.018.0 (stuck, left out)32.5
23:0035.027.018.0 (stuck, left out)31.0

Struck through: left out by the rules. City: the median of the readings that count.

Readings: Mexico's national air quality information system (SINAICA), through OpenAQ. The Aguascalientes page shows this day from OpenAQ's daily averages per monitor (see History below), so its number can differ. Aguascalientes today.

  • Broken days: smoke, dust and winter smog cover a region, so a city's day of 150 µg/m³ or more is left out when the air around it says otherwise: when the measured cities within 150 km that day read under a tenth of it (with two or more of them, their median, and none above a fifth; with one, it must be within 50 km, or the day must also be five times the city's own usual, the median of its days in the 15 before and after, or before, for a day just ended), when it is 900 or more, the level of a monitor's error code, and no city within 300 km reaches a third of it, or, for a city with no measured neighbour within 300 km, when it is 500 or more and 25 times its usual. A city whose nearest measured neighbours are 150 to 300 km away isn't judged below 900, since smoke from a nearby fire can miss a city that far off. In a city with two monitors, a monitor's day of 150 or more and over ten times the other's is left out. Fort McMurray's smoke of 2023 and 2025 stays; a monitor stuck near 985 in Mexico, one drifting at 350 beside a clean neighbour 3 km away, and Reykjavík at 466 while the next town read 5 don't. A day just ended that no measured city within 150 km can judge is checked against the Copernicus (CAMS) model as history days are (below): a reference monitor's day of 150 or more, ten times the model's recent week or more while the model shows nothing unusual and five times the city's own usual, is left out, unless it was a fireworks night, fire or thick smoke was near it, or a measured city within 300 km read a third of it.
  • Low-cost readings far too high: a low-cost city's day, or its 24-hour, 7- or 30-day value, of 100 µg/m³ or more is left out when it is five times our model estimate for the city or more and no measured city within 300 km reaches a third of it; before the estimates began, two weeks or more of days at five times the city's usual, and at five times its same weeks in other years, that end in a sudden drop, with no measured city within 300 km reaching a third of them, are left out too. A low-cost city's 24-hour value of 100 µg/m³ or more is also left out of that ranking when it is ten times its usual or more (the median of its days over the four weeks before) and no measured city within 300 km reaches a third of it (a low-cost city that is itself ten times its usual doesn't count while a reference monitor nearby reads), with no satellite fire detections, mapped smoke, volcanic ash or model dust near it and no fireworks night: several cheap sensors can fail together, as when fog wets them. Real smoke and smog show in the model or in the cities around. A Nova Fitness SDS011 sensor's reading of 999.9 µg/m³, the top of its range, where a wet or failing unit sticks, is an error code and is left out wherever it appears.
  • Each day: the mean of the city's hourly values over its own local calendar day, if at least 18 of the 24 hours were measured. The exception is Georgia: its government network publishes running 24-hour means, not hours, so a city's day there is the median of its stations' daily means as its agency publishes them (the mean stamped at the next midnight, read once it has stopped changing), a day late. None of the hourly checks (spikes, stuck or frozen hours, bursts) can run on them, a day whose means sat at one value all day, read under 1 µg/m³ (or fall within a week of days that did: a broken unit), or include a run of zeros is left out, and the broken-day, model and neighbour checks still apply. They count on past boards only: a city they measure stays off the 7-day, 30-day and year-to-date boards.
  • Boards: the last 24 hours needs 18 measured hours; 7 days needs 5 measured days; 30 days needs 21; the year to date needs 70% of days. Ranks are worst first; equal values share a rank.
  • Past periods: each day, week, month and year gets a board once it ends (a day ends when every time zone has finished it). Weeks, months and years need measurements on 70% of their days.
  • The WHO guideline on year boards: each finished year's board counts its cities against the WHO's 2021 annual guideline for PM2.5 (5 µg/m³) and its four interim targets (35, 25, 15 and 10 µg/m³). Only the cities on that board count: those measured on at least 70% of the year's days. A year's average from fewer than all its days can differ from a full year's: a city measured mostly in its cleaner months reads lower than its whole year would.
  • History before a city's first hourly measurement comes from OpenAQ's daily averages per monitor (days with 75% of hours measured), combined with the same rules. A monitor's day whose average, lowest, median, upper-quarter and highest readings, spread and hour count all repeat another of its days exactly is a feed sending old hours again under new dates, and every copy is left out. A monitor's day at one value all day is left out, except a reference monitor's at 5 µg/m³ or less, which is checked from its hours as below. A monitor's day with one hour far above the rest (over 100 µg/m³, four times the day's median, and lifting the day's average by 5 µg/m³ or more on its own), or next to a day it spent at one value, is recomputed from its hours by the hourly rules: spikes of up to six hours and runs of 18 or more identical readings left out, and no value with fewer than 18 hours left. Some networks write clean air as one value: China's national network publishes any reading under about 6 µg/m³ as 3, Belgium's writes 0, Spain's report whole numbers, and Germany's states put a reading below what the instrument detects at a fixed value such as 1.5 or 2. So a clean site can read one low value for days, and a reference monitor's run at 5 µg/m³ or less counts as clean air rather than a frozen monitor on a day when another of the city's monitors read 6 or less, or, with no other monitor that day, the measured cities within 60 km did. A history day of 150 µg/m³ or more that stands out (five times the city's usual, or from a monitor that reports too rarely to have one) is recomputed from its monitors' hours without error codes, hours stuck at one value and spikes, unless a city within 150 km read a third of it that day without such faults and no monitor held one value for half the day. A monitor's day still at 150 or more after that, or whose hours show no fault, is checked against the Copernicus (CAMS) model for that place: a day ten times the model or more while the model shows nothing unusual is left out (five times, when the day is also ten times the city's own usual; three times, when OpenAQ's archive holds fewer than 6 of its hours to check), unless it was a fireworks night in that country (as above), far more fires than that week's usual were detected close to the city, or thick smoke was mapped over it that day or the day before, another of the city's monitors at its own site read it too, a measured city within 300 km read a third of it, or the city has come near it in the same weeks of other years. A day written from a monitor's hourly readings, for which the daily averages hold no copy, gets the same check from its stored value.
  • Days missing from OpenAQ's archive: it has almost no readings for 57 days, most of them between 11 January and 25 April 2023 (all of 16 February to 18 March), plus 14 July 2023 and 1 February 2024. The monitors did report: for those days, cities in the United States and Europe use the US EPA's Air Quality System daily summaries and the European Environment Agency's verified hourly data, from the same monitors we use on other days, with the same rules (75% of hours, or 18 hours of the city's day). Cities elsewhere have no value for those days, so they are missing from those days' boards, and from months and weeks where the gap leaves them under 70% of the days. On some days only one or two countries' feeds are missing from it: the United States and Canada on March 20, 2024 and Japan on March 21, 2024. Cities in the United States use the EPA's records for their day; cities in Canada and Japan have no value for theirs and are missing from that day's board, not clean.
  • Days a network went quiet: sometimes most of a country's monitors have no readings for a day in the history we read, as India's national network had on 24 to 27 October and 3 and 4 November 2026. When a country with 10 or more measured cities on the days around it (the median of the 14 days either side) has fewer than a fifth of them on a day, that day's past board for the country, and its weeks, months and years, say so by name; the world's and each continent's boards, and the home page's "Who led" chart, name such countries when they usually have 50 or more measured cities. The cities missing that day are not clean; on longer boards they are averaged over their other days, or left off when that leaves them under 70% of the days.
  • Days without a value: a city's day needs 18 of its 24 local hours measured. City charts show a day that falls short as a grey square, and pointing at it gives the reason: no readings from the monitors that day (a monitor or its network was down, or its data hadn't reached us), readings for only part of the day, readings the checks on this page left out (a stuck monitor, error codes, a day the cities around contradict), the day missing from OpenAQ's archive (above), or, where no count of hours was kept, too few hours or readings that failed those checks. Days before a city's first reading are left empty. Data that arrives late fills a day in later, up to a week after it.

Ways to rank

Boards rank cities by their average PM2.5, worst first. A "Rank by" row offers other orders where they mean something, because an average can hide how often the air was bad, how bad it got and how many people breathed it: unhealthy days and people on the 30-day and year-to-date boards and on past months and years; the checked peak on the 7-day, 30-day and year-to-date boards and on past weeks, months and years. A city needs its place on the average board, with the same measured days. We also tried ranking by the hours people are awake and out (7:00 to 23:00 local time): the order came out almost the same as the average's, so instead each city page shows how its air changes over the day.

  • Unhealthy days: the share of a city's measured days whose daily mean was over 35.4 µg/m³, the top of the US EPA's "moderate" band: Unhealthy for Sensitive Groups or worse. A city with a few terrible weeks and a city that is bad most days can share an average; this puts the one with more bad days first. Cities with the same share are ordered by the number of unhealthy days (30 of 30 above 22 of 22, since more measured days back it), then by their average.
  • Peak: a city's worst 6 hours in a row, confirmed by a second monitoring location. Monitors within 3 km of each other count as one location, and a location's hour is the median of its monitors after the checks above (stuck units, frozen readings and spikes left out, and a reading the city's own hour dropped as far above its other monitors), only in hours the city has a value. A reading of 100 µg/m³ or more that a monitor at another location gives exactly in the same hour, and that held from the hour before or after, is a value both units are stuck at, so it doesn't confirm anything. A day's peak more than 12 times that day's average (and 50 µg/m³ above it) isn't counted: the day's own readings rule it out. Over any 6 hours, each location with readings for at least 5 of them gets its own average, and the city's figure is the second-highest of those averages: two locations 3 km or more apart both read that much or more, so one faulty monitor, a fire beside one monitor or smoke over one corner of a city can't set it. A day's peak is the worst 6 hours ending on it (local time); a board's is its worst day's, counted on the city's measured days only. A city needs a checked peak on 3 or more days; where that is fewer than the average board needs, the board shows the count, since missing days can only hide a higher peak. Cities measured at one location aren't ranked this way. A location agrees with a peak when its own average over those hours is at least three quarters of it (the 25% uncertainty EU rules allow a single PM2.5 measurement); city pages say how many locations agreed. Peaks are worked out from each monitor's hourly readings as they come in, so none go back before the day we began; a board offers this ranking once its whole span is covered.
  • People: a city's population times its share of unhealthy days, the number of people living there who breathed unhealthy air on an average day. Population is the city itself from GeoNames, not its wider metro area, so big metros are undercounted. Only cities measured at two or more locations at least 3 km apart are ranked (monitors closer than that count as one location): one monitor shouldn't speak for millions of people, and a city measured at one location can't show how widely bad air spread. Country and region boards add up their cities.

"Why it ranks here"

City pages compare recent days with the city's own history: the last 24 hours, or a full measured week, far above or below its usual for this time of year (its days within three weeks of the date in earlier years; its typical day when it has too few of those), a typical level far above the WHO guideline, or a clear seasonal peak. These are patterns in the numbers. When we name a cause, it's the usual cause for that pattern (smoke, dust, heating season), labelled as likely, not a confirmed source. When a city reads 10 times its usual or more and nothing backs it (satellites saw no fires, smoke or dust near it, no measured city within 150 km reads even a third of it, and it is measured at one location), the page says exactly that and names no cause: a single monitor that far from its usual can be faulty. When a city nearby rose too, the page names it.

For fire smoke we also show evidence: on a bad day in the last week, the number of satellite fire detections within 100 and 300 km (NASA FIRMS, VIIRS on NOAA-20; at least 50 within 300 km to be mentioned), leaving out spots that burn nearly every day, such as gas flares, refineries and steel works, as the map does, and for North America whether NOAA's Hazard Mapping System drew medium or heavy smoke over the city. We don't model the wind, and mapped smoke can be high above the ground, so this stays "likely". On a night of fireworks in the city's country (New Year, Diwali in India and Nepal, Bonfire Night in Britain and New Zealand, Lunar New Year and the Lantern Festival where they are celebrated with fireworks, Halloween in Ireland, 4 July in the US, Canada Day, Bastille Day), a bad day that also stands out from the city's usual for that time of year is named as likely fireworks, and a single monitor's short burst between 5 pm and 5 am that night isn't left out as a spike. The map of fires and smoke also outlines where the Copernicus model put thick dust (dust optical depth 0.5 or more, heavy at 1.0) and heavy smoke from burning (organic matter optical depth 0.3 or more, only near 10 or more fire detections and outside North America, where NOAA traces smoke): drawn textured as a model, never counted toward a ranking.

For volcanoes, on a bad day in the last week we name any volcano within 300 km that had a volcanic ash advisory that day, and whether its ash was seen in satellite images. Advisories come from the Tokyo VAAC (Japan, the Philippines, Kamchatka) and the Washington VAAC (Mexico, Central America, the Caribbean, northern South America); volcanoes elsewhere, such as Etna or those in Indonesia, aren't covered yet. Advisories are written for aircraft and the ash is often far above the ground, so this too stays "likely".

For dust, on a bad day in the last week we look at the Copernicus Atmosphere Monitoring Service (CAMS) global forecast: how much its dust dims sunlight through the whole air column over the city that day (dust optical depth at 550 nm, a daily mean). Clean air is under 0.1 and the outbreaks that fill the air at street level with desert dust reach 0.5 to 1 or more. We name dust when it is 1 or more, or 0.5 or more and at least one and a half times the city's usual for the time of year (its days within 15 of the date in earlier years, or its last 30 days until it has those), so a desert city's everyday haze doesn't count. The model sees the whole column, and dust can pass high overhead, so this too stays "likely".

Bad-air episodes

An episode is a run of days when many cities close to each other had air far dirtier than their own usual. A city counts on a day when its daily mean is at least 25 µg/m³ and at least twice its usual for that time of year: the median of its daily means within three weeks of the date, in every year we have, leaving out the week either side of the day itself so a long episode doesn't raise its own baseline. Days a monitor reported as zero are left out of it, and a city needs 20 days to have a usual. Each day, counting cities within 350 km of each other form a group; a group of four or more that touches a group from the day before (a city within 350 km) continues the same episode. An episode gets a page when it reached eight or more cities spread over at least 150 km, so one metro area's districts on their own don't make one. Names come from where the cities were and the months it lasted, and cover every city counted: a region is named only when all the cities are in it, and a part of a country (such as north-western) is the one where most of them were and holds the episode's worst city and the one furthest above its usual; otherwise the name widens. Two episodes in the same place and month are told apart by their dates. An episode's figures (each city's days, highest reading and how far above its usual) come only from the days that city counted in that episode, each against its usual for that date. Episodes are found again every day for the last two months, so a live episode grows as days complete, and late data can join or split an older one. A page says when an episode ended; when its cities' readings stop the day after its last day, it says that instead, since the episode may have run longer, and names the monitor network that went silent when our data shows it.

A likely cause is named only from the date (New Year's night, the Fourth of July in the United States, Diwali in India and Nepal for an episode that began on the festival's night or the day after, with smoke from fires named beside it when the fire evidence is there too) or from evidence on the counted days: NASA FIRMS fire detections within 300 km of the city, NOAA smoke maps over it, or heavy dust over it in the CAMS forecast (the rule above, on 40% or more of the counted days). Satellites see crop, field and forest fires alike, so we say smoke from fires, not wildfire smoke. That evidence is kept every day from 24 September 2026; for earlier episodes it comes from NASA's FIRMS archive and NOAA's smoke map archive for the episode's own days. NOAA's smoke maps cover North America, and count there on days without fire detections. An episode with none of these names no cause.

Cleaner or dirtier than a year before

Cleaner or dirtier compares each measured city's mean PM2.5 with the same calendar days a year earlier, counting only the days it was measured in both years, so a city whose monitors were quiet for a month in one year isn't judged on a different season. A year runs from 1 January to 31 December, or to yesterday while it runs, and 29 February is left out. A city needs 55% of the year's days measured in both years, and at least 30. Cities are ranked by the change in µg/m³, biggest rise first, so a city going from 2 to 4 µg/m³ doesn't outrank one going from 40 to 60; the change in per cent is shown beside it. Weather swings air quality from one year to the next as much as emissions do, so a change says how the air was, not why. A change is only as fair as its monitors: of a city's monitors with a start date on file, when under half reported through both years whole (a monitor counts when it started within 30 days of the earlier year's 1 January and was still reporting within 30 days of the end), its row is marked “its monitors changed between the years”. It stays on the list, but it isn't named as the city that improved most or rose most, and it isn't on the card's top 10. A city whose monitors have no start date on file isn't judged.

Time of day

A city page's time of day takes the city's hourly values over the last 30 days and, for each hour of its local clock (daylight saving included), the median of that hour's readings. Only days with 18 or more hours measured count, and a city needs 20 such days, so every hour is read over the same full days. The worst and cleanest hours named are the highest and lowest of those medians; when they are less than 20% apart, the page says the day barely changes. Traffic, cooking and heating, and the shallow night-time air that holds smoke near the ground all shape the day, so it says when the air is usually worse, not why.

How the map is drawn

The map of the world and of each continent shows the air near the ground as one coloured field. Every measured city spreads its reading over about 150 km, fading out with distance, and where several cities' reaches meet, their readings are blended on the scale's own steps. Past about 150 km from any measured city the field stops: we don't fill the gaps between cities with a guess. The field is a picture of the readings, not a measurement of the places between them, and it never counts toward a ranking. Cities with only one or two low-cost sensors stay out of it. Where a country's government network isn't read, or has gone quiet, the map hatches the country and says why; where no monitor we use stands at all, a fainter hatch says so.

Beyond the measured cities the map shows the Copernicus model's surface PM2.5, faint and hatched so it can't be mistaken for a reading, scaled by the same calibration against nearby measured cities that our estimates use. The model's smoke and haze (organic matter) and desert dust are soft fields that follow its 3-hour steps, so they move through the day rather than jump. None of the model's layers ever count toward a ranking or a cause.

The earth underneath is NASA's Blue Marble, and the night side is worked out from the sun's position for the moment shown, with NASA's Black Marble city lights. Clouds come from the GOES and Himawari weather satellites every 10 minutes, and over Europe, Africa and the Indian Ocean from EUMETSAT's Meteosat every hour. Fires glow from satellite detections: by the day for a day, and in hourly play from the last six hours of geostationary scans, brightening while a fire is seen and cooling after. Wind is NOAA's GFS model every 6 hours, near the ground and at the jet stream's height, and tropical cyclones are NOAA's and the US Navy's tracks. Wind, clouds and cyclones are there to explain the picture; they never touch a ranking. When our updates stop for some hours, the next daily catch-up fills those hours back in from the networks' own records, so the gap closes within a day.

Estimates for cities without a monitor

Cities with no monitor we can use get a model estimate instead of a ranking: the Copernicus Atmosphere Monitoring Service (CAMS) global forecast, a daily mean of its 3-hourly surface PM2.5 at the city's location (a grid about 44 km across). Models miss local sources, so we scale each city's estimate by how the model compared with cities within 300 km measured by reference monitors over the last 7 days (on 4 days or more, counting cities whose network we read a day late for past boards only, as China's; the median ratio of up to 10 of them; low-cost sensors aren't used). Tested against 2,186 measured cities, this put 65% of single-day estimates within ×1.5 of the monitor, against 45% for the raw model. Cities with fewer than three such neighbours keep the raw model value, and their pages say so. Estimates are always labelled, rounded to whole numbers, listed by population rather than ranked, and never mixed into the measured boards.

Limits

  • A monitor measures the air where it stands; a city with one monitor next to a busy road may look worse than its average resident's air.
  • Coverage is uneven: the United States, Europe and Japan have many usable monitors, while across most of Asia, Africa and Latin America we have few or none, because official readings there aren't open for reuse or aren't published at all. How much of the world we measure.
  • Whole networks sometimes stop reaching us for a day or more. Their cities leave the 24-hour board until the data returns, and the country page says how many monitors went quiet and since when. After 30 days their monitors leave the current rankings, but the country page and their cities' pages still name the network and the day it stopped, for up to a year, unless other monitors now measure most of those cities.
  • Providers sometimes revise data after publication; our past boards use what was published when they were built, and history refreshes may change them slightly.

The pollutant

PM2.5: particles under 2.5 micrometres, in micrograms per cubic metre (µg/m³). We compare against the World Health Organization's 2021 guidelines (15 µg/m³ for a 24-hour mean, 5 µg/m³ for a yearly mean) and colour values by the US EPA's PM2.5 categories (2024):

  • Good: 0–9.0 µg/m³
  • Moderate: 9.1–35.4 µg/m³
  • Unhealthy for sensitive groups: 35.5–55.4 µg/m³
  • Unhealthy: 55.5–125.4 µg/m³
  • Very unhealthy: 125.5–225.4 µg/m³
  • Hazardous: 225.5 and above µg/m³

Scales

Numbers can be shown on the scale you know: pick one under any chart's colour key. We start from your country's scale where it has one (US AQI in the US, India's AQI in India, China's AQI in China, Malaysia's IPU in Malaysia, Indonesia's ISPU in Indonesia, the European bands in the EU, Iceland, Liechtenstein, Norway and Switzerland, the UK's DAQI, Canada's AQHI+) and µg/m³ elsewhere; your choice is kept in one cookie. Rankings don't change with the scale. Share images, widgets and the citation stay in µg/m³.

Each scale here is its PM2.5 part only. Official indexes take the worst of several pollutants (ozone, NO₂ and others), so the official number for a city can be higher than ours. The scales are defined on 24-hour means (AQHI+ and the European index on hourly ones); we apply them to whatever mean a page shows, such as a 7-day average. Every scale side by side, with a converter.

US AQI

The US EPA's Air Quality Index for PM2.5 (2024 breakpoints). Source. Its bands, what they mean and a converter.

  • Good (0–50): up to 9.0 µg/m³
  • Moderate (51–100): over 9.0 to 35.4 µg/m³
  • Unhealthy for sensitive groups (101–150): over 35.4 to 55.4 µg/m³
  • Unhealthy (151–200): over 55.4 to 125.4 µg/m³
  • Very unhealthy (201–300): over 125.4 to 225.4 µg/m³
  • Hazardous (301+): over 225.4 µg/m³
India AQI

India's National Air Quality Index (CPCB), PM2.5 sub-index. Source. Its bands, what they mean and a converter.

  • Good (0–50): up to 30.0 µg/m³
  • Satisfactory (51–100): over 30.0 to 60.0 µg/m³
  • Moderate (101–200): over 60.0 to 90.0 µg/m³
  • Poor (201–300): over 90.0 to 120.0 µg/m³
  • Very poor (301–400): over 120.0 to 250.0 µg/m³
  • Severe (401+): over 250.0 µg/m³
China AQI

China's Air Quality Index (HJ 633-2026), PM2.5 individual index. Source. Its bands, what they mean and a converter.

  • Excellent (0–50): up to 35.0 µg/m³
  • Good (51–100): over 35.0 to 60.0 µg/m³
  • Lightly polluted (101–150): over 60.0 to 115.0 µg/m³
  • Moderately polluted (151–200): over 115.0 to 150.0 µg/m³
  • Heavily polluted (201–300): over 150.0 to 250.0 µg/m³
  • Severely polluted (301+): over 250.0 µg/m³
Malaysia IPU

Malaysia's Air Pollutant Index (Indeks Pencemar Udara, IPU) from the Department of Environment, PM2.5 sub-index. Source. Its bands, what they mean and a converter.

  • Good (0–50): up to 12.0 µg/m³
  • Moderate (51–100): over 12.0 to 50.4 µg/m³
  • Unhealthy (101–200): over 50.4 to 150.4 µg/m³
  • Very unhealthy (201–300): over 150.4 to 250.4 µg/m³
  • Hazardous (301+): over 250.4 µg/m³
Indonesia ISPU

Indonesia's Air Pollution Standard Index (Indeks Standar Pencemar Udara, ISPU; regulation P.14/2020), PM2.5 sub-index. Source. Its bands, what they mean and a converter.

  • Good (0–50): up to 15.5 µg/m³
  • Moderate (51–100): over 15.5 to 55.4 µg/m³
  • Unhealthy (101–200): over 55.4 to 150.4 µg/m³
  • Very unhealthy (201–300): over 150.4 to 250.4 µg/m³
  • Hazardous (301+): over 250.4 µg/m³
European AQI

The European Environment Agency's air quality index bands for PM2.5 (2024 revision); it has bands, not numbers, so values stay in µg/m³. Source. Its bands, what they mean and a converter.

  • Good: up to 5.0 µg/m³
  • Fair: over 5.0 to 15.0 µg/m³
  • Moderate: over 15.0 to 50.0 µg/m³
  • Poor: over 50.0 to 90.0 µg/m³
  • Very poor: over 90.0 to 140.0 µg/m³
  • Extremely poor: over 140.0 µg/m³
UK DAQI

The UK's Daily Air Quality Index (1 to 10), PM2.5 bands. Source. Its bands, what they mean and a converter.

  • Low (1–3): up to 35.0 µg/m³
  • Moderate (4–6): over 35.0 to 53.0 µg/m³
  • High (7–9): over 53.0 to 70.0 µg/m³
  • Very high (10): over 70.0 µg/m³
Canada AQHI+

Canada's AQHI+ for PM2.5 (1 point per 10 µg/m³), which is defined on hourly values. Source. Its bands, what they mean and a converter.

  • Low risk (1–3): up to 30.0 µg/m³
  • Moderate risk (4–6): over 30.0 to 60.0 µg/m³
  • High risk (7–10): over 60.0 to 100.0 µg/m³
  • Very high risk (10+): over 100.0 µg/m³

Two guides build on this page: how accurate low-cost sensors are, from our own pairs, and choosing and sizing an air purifier for when the air is bad.

Nothing on SmogRank is medical advice. For health guidance during bad air, follow your local health authority.