तरीक़ा

How the analysis works

Short method, updated 19 October 2026. Each report identifies its data, period, calculation and coverage. Historical reports keep their own dated methods. The detailed method paper linked here describes the current implementation and its limits.

Three measures, three meanings

Measured concentration describes the pollution recorded at monitoring stations. AQI converts pollutant concentrations into an index. Weather-removed concentration is a model estimate under a specified reference weather distribution. These measures have different units and must not be treated as interchangeable.

The daily report labels measured and weather-removed figures separately. Its daily weather-removed figures are displayed only when the city's weather-removed year-on-year change is positive. A pollutant's daily weather-removed figure also requires that pollutant's change to be positive. An omitted figure is not a zero estimate, and the displayed cells are not a complete distribution of changes. Annual summaries and special studies state their own inclusion rules. The station count, comparison period and weighting belong to the particular measure. A city summary does not describe every neighbourhood.

The readings and their coverage

The analysis uses hourly CPCB and DPCC monitoring records compiled by the Centre for Research on Energy and Clean Air (CREA). When a report uses an alternative feed, its source and period must be identified. This office runs no monitoring instruments. Errors in our analysis remain our responsibility.

Station identifiers are used to match records. Duplicate records cannot create extra hours. Quality checks and units are pollutant-specific. A daily pollutant value requires at least 16 distinct valid hourly readings within its 24-hour window. Missing readings are not zero pollution.

For carbon monoxide and ozone, the documented-formula daily calculation uses the highest qualifying eight-hour mean wholly inside that window. An eight-hour block needs at least six valid hourly readings. The daily 16-hour requirement also applies. The calculation does not join observations across a gap to invent a continuous block.

The daily window and AQI formula

The current daily AQI window runs from 4 PM IST on the previous day to 4 PM IST on the report date. The start is included and the end is excluded. The daily weather producer uses the opposite endpoint convention. Its effect on shipped figures has not been quantified; the detailed paper explains this limitation. Earlier reports may use another stated window. A calendar-day result and a rolling-window result must not be compared as if their periods were identical.

For the public AQI calculation, each qualifying pollutant concentration is converted with CPCB's documented formula. A station AQI requires at least three qualifying pollutants, including PM2.5 or PM10. The highest sub-index sets the station's AQI, and the responsible pollutant is named. The daily city AQI is the mean of qualifying station AQIs. Separate concentration or population summaries state their own basis.

An AQI displayed at 500 is an index ceiling. It does not establish that a monitoring instrument cannot measure a higher concentration. Calculating an index from mean concentrations can also differ from averaging hourly indices. Our documented-formula result should be identified as our calculation from the recorded concentrations.

Sources: CPCB's AQI calculation note and CPCB's AQI calculator. The AQI 500 investigation explains the distinction between the index and measured concentration.

Weather inputs and their sources

The saved files checked on 19 October 2026 contain ERA5 reanalysis supplied through Open-Meteo. Some 2026 hours instead carry separate labels for Open-Meteo Best Match or gap fills using weather station observations. These are not all the same weather product. A complete hourly record does not prove every hour came from ERA5.

The historical file also contains ERA5 gap filling for the height of the boundary layer from Google Earth Engine for part of 2024. Its clock alignment was chosen using an overlap check. A comparison depends on the reference weather used. Its uncertainty range does not measure every source of error in the reference weather.

The temperature-inversion series uses selected Safdarjung balloon profiles. Observation times vary across dates and belong in the comparison record. A missing inversion height is not zero height. A single profile is not an all-day measurement across Delhi.

Each dated analysis should identify when its inputs were saved, where gaps were filled and when observations were taken. This source check does not quantify their effect on shipped pollution figures. Sources: Open-Meteo historical weather, ERA5 on Google Earth Engine and the University of Wyoming sounding archive.

What weather removal can tell us

Separate random-forest models estimate concentrations for seven pollutants using station, time and weather information. The standard daily specification uses 64 prediction repetitions and 1,000 bootstrap replicates. When spread fields are absent, those repetitions use the same reference weather. The detailed paper explains this source-conditional limit. Model versions, training periods and validation results belong to dated assessments, available in the model-performance archive.

The model asks what concentration it predicts under the chosen reference weather. The answer depends on the model, its training data and that reference. Weather removal does not measure emissions directly. A remaining difference does not by itself establish a policy effect, identify a source or rule out every weather-related explanation.

Training scores and comparisons with another model do not guarantee accuracy on a new period. Results overlapping a model's training years are not independent forecasts. Gas measurements require particular attention to units, quality and atmospheric chemistry. Ozone changes cannot be treated as a simple emissions indicator.

Comparisons and uncertainty

The daily year-on-year baseline uses the corresponding date last year, plus or minus seven days. Comparisons use matched station identifiers and disclose their eligible network. Special investigations can use different, explicitly stated panels and periods. Their estimates must not be combined with a daily comparison merely because both are percentages.

A bootstrap interval describes the variation generated by its resampling design. It does not automatically cover instrument error, missing observations, model error or every form of temporal dependence. The station panel, resampling unit and assumptions matter. A range crossing zero does not establish a direction of change. Sensitivity results that change the conclusion must be disclosed.

Weather-draw variation, sampling uncertainty and model validation answer different questions. None alone turns an observational comparison into a causal estimate. Every numerical finding needs its own coverage and uncertainty checks.

Population near monitors

Population estimates describe residents within the stated proximity threshold of eligible monitors. They are not personal exposure measurements and do not mean a monitor represents only those residents. The current population surface, official annual total and allocation method are described in the population methodology linked on this page.

The population headline counts people near monitors in the Poor or worse band, effective AQI above 200. A day with no qualifying Poor monitor is not relabelled using Moderate monitors. Station categories and city summaries retain their actual values. Historical population estimates keep their dated basis.

The long record and the limits of attribution

The 1980 to July 2026 long-record report separates earlier model-based estimates from the later monitoring record. The observing-system change around 2000 is retained and discussed. A smooth chart does not remove that change in evidence, and a historical association does not establish its cause.

Checks, wording and human decisions

R and Python code compute the figures. Language models can assist with code and wording, but their prose is not a measurement or a validation result. Numerical, source, coverage and public-content checks are required. These checks reduce errors; they do not make errors impossible.

Ajay Maken selects the daily film's narratives and approves its final version. Automatic review panels are not part of the daily run. A separate expert review requires his specific request. The full written record can contain findings omitted from his chosen film.

The public record must retain results and limitations needed to interpret its conclusions. Detailed working logs may be kept separately, but a material caveat must not disappear merely to make a shorter report.