In the 1980s, an average place in Delhi spent less than one day a year above 120 micrograms of PM2.5 per cubic metre. From 2010 to September 2026, it spent 89.5 such days a year. CPCB calls a day between 121 and 250 micrograms very poor, and a day above 250 severe.
PM2.5 is the mass of particles smaller than 2.5 thousandths of a millimetre in each cubic metre of air. LongPMInd is a peer-reviewed daily map of PM2.5 for India, by Wang and co-authors, and it ends in 2022. The office rebuilt it by the paper's method and carried it to 30 September 2026. This report calls that series the rebuild, and calls the published map the original.
The rebuild runs on model AI from 1980 to 2022, and on model BI from 2023 to September 2026. Every chart marks the point where one model hands over to the other. The averages are taken over 10 km cells of land, never over people.
Source: ledger l17b_cell_days.csv, rebuild, all cells, Delhi; facts.json days and levels; CPCB, National Air Quality Index, 2014, Table 3.11, page 35; read 21 November 2026. Analysis in-house.
1. Very poor days in Delhi went from almost none to about 90 a year. At an average Delhi cell, days above 120 numbered 0.3 a year in the 1980s, 3.3 in the 1990s, 83.5 in the 2000s and 89.5 from 2010 to September 2026. The original agrees on the size of the change, with 1.4 days in the 1980s and 82.7 from 2010 to 2022. At Delhi's monitors in 2023 to 2025, 29.5 per cent of station-days were above 120, against 26.4 per cent in the rebuild, so the rebuild's count is, if anything, low.
2. Delhi's average PM2.5 nearly doubled, to 91.7. The rebuild gives Delhi 49.3 micrograms in the 1980s, 54.8 in the 1990s, 88.6 in the 2000s and 91.7 from 2010 to September 2026. That last figure is 18.3 times WHO's guideline of 5 and 2.3 times India's standard of 40. Of that period, 2023 to September 2026 can be read at CPCB's monitors: on their own station-days the monitors read 93.9, and the rebuild, model BI, reads 88.2.
3. Delhi's winter more than doubled. Delhi's average from December to February was 60.7 in the 1980s and 143.5 from 2010 to September 2026. The smog season, November to February, averaged 153.0 in that last period. The monsoon months rose far less, from 35.1 to 41.5.
4. The share of India's land above India's own standard doubled. In the 1980s, 28.0 per cent of India's cells had an average above 40. From 2010 to September 2026, 59.1 per cent did. The all-India average rose from 35.7 to 45.1, which is 9.0 times WHO's guideline.
5. No decade since 2000 was cleaner than the one before it. Delhi averaged 88.6 in 2000 to 2009 and 92.9 in the full years 2010 to 2025. NCR averaged 75.1 and 76.1, and all India 45.6 and 45.6. Every one of the 20 NCR districts stayed above India's standard of 40 in all four periods.
6. The original LongPMInd tells the same story. On the cells both maps cover, the rebuild's level is within 5 per cent of the original's in every period, for Delhi, NCR and India. The original alone puts Delhi at 50.8 in the 1980s and 89.2 in 2010 to 2022. The two differ more on counts of days than on levels.
Source: ledgers l17_levels.csv, l17b_cell_days.csv and l17_seasons.csv; facts.json levels, days, seasons, cpcb_beside and cpcb_pools; the rebuild is model AI to 2022 and model BI from 2023; read 21 November 2026. Analysis in-house.
Table 1. Average PM2.5 in micrograms per cubic metre by period, for Delhi, NCR, NCR outside Delhi and all India, with multiples of WHO's guideline and India's standard.
| Period | Area | Average PM2.5 | Times WHO's 5 | Times India's 40 | Cells above 40, per cent | CPCB beside (Delhi, from 2023) |
|---|---|---|---|---|---|---|
| 1980 to 1989 | Delhi | 49.3 | 9.9 | 1.2 | 100 | |
| 1990 to 1999 | Delhi | 54.8 | 11.0 | 1.4 | 100 | |
| 2000 to 2009 | Delhi | 88.6 | 17.7 | 2.2 | 100 | |
| 2010 to September 2026 | Delhi | 91.7 | 18.3 | 2.3 | 100 | 2023 to Sep 2026 at the monitors: CPCB 93.9, rebuild 88.2 on the same station-days |
| 1980 to 1989 | NCR, all of it | 47.1 | 9.4 | 1.2 | 96 | |
| 1990 to 1999 | NCR, all of it | 49.8 | 10.0 | 1.2 | 100 | |
| 2000 to 2009 | NCR, all of it | 75.1 | 15.0 | 1.9 | 100 | |
| 2010 to September 2026 | NCR, all of it | 75.1 | 15.0 | 1.9 | 100 | |
| 1980 to 1989 | NCR outside Delhi | 47.1 | 9.4 | 1.2 | 96 | |
| 1990 to 1999 | NCR outside Delhi | 49.7 | 9.9 | 1.2 | 100 | |
| 2000 to 2009 | NCR outside Delhi | 74.8 | 15.0 | 1.9 | 100 | |
| 2010 to September 2026 | NCR outside Delhi | 74.8 | 15.0 | 1.9 | 100 | |
| 1980 to 1989 | All India | 35.7 | 7.1 | 0.9 | 28 | |
| 1990 to 1999 | All India | 36.4 | 7.3 | 0.9 | 31 | |
| 2000 to 2009 | All India | 45.6 | 9.1 | 1.1 | 55 | |
| 2010 to September 2026 | All India | 45.1 | 9.0 | 1.1 | 59 |
READINGEach row is one area in one period. The last column puts CPCB's monitors beside Delhi for the years they can be read, 2023 to September 2026. CAUTIONThe last period lacks October to December 2026, which are the worst months of Delhi's year. The CPCB figure covers 2023 to September 2026 only, and no monitor figure exists for 2010 to 2017. SOURCESource: ledger l17_levels.csv, rebuild, all cells; facts.json cpcb_beside; the last period is model AI to 2022 and model BI from 2023; read 21 November 2026. THE POINTEvery area stood higher from 2000 onward than in the 1980s and 1990s, and Delhi stood highest of all. |
Delhi's average PM2.5 rose from 49.3 micrograms in the 1980s to 54.8 in the 1990s. It then jumped to 88.6 in the 2000s, and stood at 91.7 from 2010 to September 2026. Every Delhi cell was above India's standard of 40 in every period.
For the years from 2023, CPCB's Delhi monitors stand beside the rebuild. On the monitors' own station-days in 2023 to September 2026, the monitors read 93.9 and model BI reads 88.2. For 2018 to 2022, on the same footing, the monitors read 102.9 and model AI reads 99.2. In both spans the rebuild reads below the monitors.
Source: ledger l17_levels.csv, rebuild, all cells, Delhi; facts.json cpcb_beside and cpcb_2018_2022, which pool the rows of l17_cpcb_check.csv by station-days; read 21 November 2026. Analysis in-house.
NCR, the 599 cells of the National Capital Region's 20 districts, rose from 47.1 to 49.8, then to 75.1 in the 2000s and 75.1 from 2010 to September 2026. All India rose from 35.7 to 36.4, then to 45.6 and 45.1. In the full years 2010 to 2025, the three stood at 92.9, 76.1 and 45.6.
The largest step in every series came between the 1990s and the 2000s. Part of that step may come from the inputs, because NASA's MERRA-2 record took in no satellite aerosol data before 2000. So the size of that step should be read with care, while the levels after 2000 do not depend on it.
Source: ledger l17_levels.csv, rebuild, all cells, NCR and All India; Wang et al. 2024, section 4.3, page 3570; the last period is model AI to 2022 and model BI from 2023; read 21 November 2026. Analysis in-house.
READINGIn every full year from 2001 to 2025, Delhi's rebuilt level was 80.4 or more, about twice India's standard. The highest year was 2018, at 107.2, and in Delhi the black squares of the monitors sit above the rebuild in every year. CAUTIONThe line changes from model AI to model BI at 2023, and the step there is the change of model. On the grid in 2018 to 2022, AI reads above BI by 11.2 to 12.4 per cent over all India, so the all-India fall from 46.4 in 2022 to 41.7 in 2023 is the model. The 2026 points cover January to September only. SOURCESource: ledgers l17_annual.csv and l17_cpcb_check.csv; CPCB station-days through CREA's mirror; read 21 November 2026. THE POINTDelhi's air has stood near twice India's standard in every year since 2001, and the monitors put it higher still. |
The rebuild gives Delhi 91.5 in 2023, 93.0 in 2024 and 86.6 in 2025, on model BI. At the same places and on the same days, CPCB's monitors read 101.4, 104.0 and 98.7. In each of these full years the monitors read above the rebuild.
For January to September 2026, the rebuild gives Delhi 67.9. That figure leaves out the three worst months of the year, so it is never set beside a full year in this report.
Source: ledgers l17_annual.csv and l17_cpcb_check.csv, Delhi; facts.json ai_minus_bi_grid; read 21 November 2026. Analysis in-house.
India's 24-hour standard for PM2.5 is 60 micrograms. At an average Delhi cell, days above 60 numbered 96.2 a year in the 1980s, 138.6 in the 1990s, 232.2 in the 2000s and 227.0 from 2010 to September 2026. In the full years 2010 to 2025, the count was 230.0.
Days above 120 in Delhi numbered 0.3 a year in the 1980s and 3.3 in the 1990s. They rose to 83.5 in the 2000s and 89.5 from 2010 to September 2026. Days above 250, CPCB's severe band, rose from none in the 1980s to 9.0 a year in the last period.
Source: ledger l17b_cell_days.csv, rebuild, all cells, Delhi; facts.json days; India's National Ambient Air Quality Standards, 2009; CPCB, National Air Quality Index, 2014, Table 3.11; read 21 November 2026. Analysis in-house.
CPCB's monitors stand beside these Delhi counts for the years they can be read. In 2023 to 2025, 29.5 per cent of station-days at Delhi's monitors were above 120, against 26.4 per cent in the rebuild. Above 250, the monitors recorded 6.7 per cent and the rebuild 3.5 per cent.
So the rebuild undercounts the worst days. Its counts above 250, and to a lesser degree above 120, are low. Its count above 60 runs slightly high, at 61.0 per cent of station-days against the monitors' 59.4.
Source: facts.json cpcb_pools, Delhi, 2023 to 2025, full years; ledger l17c_cpcb_windows.csv; model BI; read 21 November 2026. Analysis in-house.
Across NCR, days above 60 rose from 78.1 a year in the 1980s to 196.5 from 2010 to September 2026. Days above 120 rose from 0.4 to 51.6. Across all India, an average cell spent 33.2 days a year above 60 in the 1980s and 82.3 days in the last period, counted over all of India's cells.
These counts are made at each 10 km cell and then averaged over the area's cells. An average of the whole area would hide the polluted places. All India's average was above 120 on no day in any period, while an average cell was above 120 on 9.7 days a year from 2010 to September 2026.
WHO's 24-hour guideline of 15 is not counted here. Even Delhi's average was above 15 on more than 364 days a year in every period, so the count cannot show any change.
Source: ledgers l17b_cell_days.csv and l17_days_above.csv, rebuild, all cells, NCR and All India; the last period is model AI to 2022 and model BI from 2023; read 21 November 2026. Analysis in-house.
READINGIn Delhi the bars above 120 rise from almost nothing in the 1980s and 1990s to more than 80 days from the 2000s. The red bar is the rebuild from 2010 to September 2026, on model AI to 2022 and model BI from 2023. CAUTIONOn these cells all India shows 86 days above 60 for the last period, while the headline over all cells is 82. The last bar lacks October to December 2026, and the model undercounts the worst days. SOURCESource: ledger l17b_cell_days.csv, cells the original also covers; facts.json days; read 21 November 2026. THE POINTBoth the rebuild and the original show the very poor days in Delhi arriving from the 2000s onward. |
READINGIn the 1980s most of the plain from Punjab to Bengal fell between 30 and 120 days a year above 60. From 2010, most of it fell between 150 and 240 days, and the average cell of India moved from 33 days to 82. CAUTIONThe right map joins model AI to 2022 and model BI from 2023, and it lacks October to December 2026. Western Rajasthan is dark in both maps, most likely from desert dust, which was not tested input by input. SOURCESource: ledger l17b_cell_days.csv and the per-cell store l17b_cell_days.npz; read 21 November 2026. THE POINTThe days above India's daily standard spread across the whole Indo-Gangetic plain. |
Table 2. Days a year above 60, 120 and 250 micrograms at an average cell, by area and period, for the rebuild and the original, on the cells the original also covers.
| Period | Area | Above 60, rebuild | Above 120, rebuild | Above 250, rebuild | Above 60, original | Above 120, original | Above 250, original |
|---|---|---|---|---|---|---|---|
| 1980 to 1989 | Delhi | 96 | 0.3 | 0.0 | 108 | 1.4 | 0.0 |
| 1990 to 1999 | Delhi | 139 | 3.3 | 0.0 | 143 | 3.2 | 0.0 |
| 2000 to 2009 | Delhi | 232 | 84 | 5.8 | 227 | 76 | 6.6 |
| 2010 to September 2026 | Delhi | 227 | 90 | 9.0 | not covered | not covered | not covered |
| 2010 to 2022 | Delhi | 233 | 93 | 9.0 | 225 | 83 | 8.2 |
| 2010 to 2025 | Delhi | 230 | 92 | 9.3 | not covered | not covered | not covered |
| 1980 to 1989 | NCR | 78 | 0.4 | 0.0 | 88 | 0.9 | 0.0 |
| 1990 to 1999 | NCR | 99 | 1.1 | 0.0 | 106 | 1.8 | 0.0 |
| 2000 to 2009 | NCR | 208 | 49 | 2.0 | 206 | 55 | 3.1 |
| 2010 to September 2026 | NCR | 196 | 52 | 2.5 | not covered | not covered | not covered |
| 2010 to 2022 | NCR | 207 | 57 | 2.9 | 203 | 58 | 3.6 |
| 2010 to 2025 | NCR | 200 | 54 | 2.6 | not covered | not covered | not covered |
| 1980 to 1989 | All India | 35 | 1.1 | 0.0 | 41 | 0.4 | 0.0 |
| 1990 to 1999 | All India | 39 | 1.0 | 0.0 | 47 | 0.6 | 0.0 |
| 2000 to 2009 | All India | 88 | 11 | 0.2 | 98 | 13 | 0.2 |
| 2010 to September 2026 | All India | 86 | 10 | 0.2 | not covered | not covered | not covered |
| 2010 to 2022 | All India | 93 | 12 | 0.2 | 105 | 14 | 0.3 |
| 2010 to 2025 | All India | 88 | 11 | 0.2 | not covered | not covered | not covered |
READINGEach row is one area in one period, with the rebuild's three counts and then the original's. Counts under ten days carry one decimal. CAUTIONWhere a count is under ten days, read the difference in days and never as a per cent. In Delhi in the 1980s the rebuild counts 0.3 days above 120 and the original 1.4, which is 1.1 days apart. The original ends in 2022, so it has no figure for the later periods. SOURCESource: ledger l17b_cell_days.csv, cells the original also covers; the periods running past 2022 are model AI to 2022 and model BI from 2023; read 21 November 2026. THE POINTThe rebuild and the original agree that Delhi now spends most of each year above India's daily standard. |
On the cells both cover, the two series agree within a few per cent on Delhi's days above 60 from 2000 onward. In 2010 to 2022 they give 233.1 and 225.5. Above 120 the rebuild counts more in Delhi, with 92.8 against 82.7. Across India the rebuild counts fewer days above 60, with 92.8 against 105.3.
Source: facts.json days, rebuild_common and original_common; ledger l17b_cell_days.csv; read 21 November 2026. Analysis in-house.
Delhi's winter, December to February, averaged 60.7 micrograms in the 1980s, 72.1 in the 1990s, 127.8 in the 2000s and 143.5 from 2010 to September 2026. The post-monsoon months, October and November, rose from 57.6 to 146.0. The smog season, November to February, rose from 61.6 to 153.0.
The monsoon months, June to September, rose far less, from 35.1 to 41.5. The pre-monsoon months, March to May, rose from 51.4 to 74.0. So the rise in Delhi is above all a rise in the cold months.
For the years from 2023, the CPCB monitors stand beside these Delhi figures only as a whole-year pool. On their own station-days in 2023 to September 2026 they read 93.9, against 88.2 for model BI. The ledgers hold no season split of the monitors.
Source: ledger l17_seasons.csv, rebuild, all cells, Delhi; facts.json seasons and cpcb_beside; read 21 November 2026. Analysis in-house.
NCR's winter rose from 56.9 to 110.8, and its smog season from 57.2 to 117.6. All India's winter rose from 44.9 to 65.1, and its monsoon from 24.5 to 25.9. A season's average is not held against any limit here, because India's limits apply to a year or to a day.
The season figures are hardly touched by the missing end of 2026. Each season averages only its own months. The post-monsoon figure for 2010 to September 2026 therefore covers 2010 to 2025.
Source: ledger l17_seasons.csv, rebuild, all cells, NCR and All India, with its months column; the last period is model AI to 2022 and model BI from 2023; read 21 November 2026. Analysis in-house.
READINGIn Delhi and NCR the winter and post-monsoon bars more than double from the 1980s, while the monsoon bars barely move. The hollow diamond is the rebuild from 2010 to September 2026, on model AI to 2022 and model BI from 2023. CAUTIONThe darkest bar is 2010 to 2022 on model AI alone, so that it matches the original's last period. The dashed line at 40 marks India's annual standard for scale only. SOURCESource: ledger l17_seasons.csv, cells the original also covers; read 21 November 2026. THE POINTDelhi's cold months now carry more than twice the PM2.5 they carried in the 1980s. |
Table 3. Average PM2.5 in each of five seasons, by period, for Delhi, NCR and all India, with the rebuild and the original for 2010 to 2022.
| Area | Season | 1980s | 1990s | 2000s | 2010 to Sep 2026 | Rebuild, 2010 to 2022, cells the original covers | Original, 2010 to 2022 |
|---|---|---|---|---|---|---|---|
| Delhi | Winter, Dec to Feb | 60.7 | 72.1 | 127.8 | 143.5 | 145.5 | 137.3 |
| Delhi | Pre-monsoon, Mar to May | 51.4 | 55.1 | 75.2 | 74.0 | 75.8 | 73.8 |
| Delhi | Monsoon, Jun to Sep | 35.1 | 35.7 | 42.6 | 41.5 | 41.9 | 41.1 |
| Delhi | Post-monsoon, Oct to Nov | 57.6 | 66.7 | 141.7 | 146.0 | 144.8 | 136.3 |
| Delhi | Smog season, Nov to Feb | 61.6 | 72.7 | 137.6 | 153.0 | 153.6 | 144.5 |
| NCR | Winter, Dec to Feb | 56.9 | 63.4 | 104.8 | 110.8 | 114.9 | 115.9 |
| NCR | Pre-monsoon, Mar to May | 50.4 | 51.5 | 67.4 | 65.0 | 67.3 | 67.4 |
| NCR | Monsoon, Jun to Sep | 34.6 | 34.2 | 40.0 | 38.5 | 39.2 | 39.2 |
| NCR | Post-monsoon, Oct to Nov | 52.6 | 58.3 | 112.5 | 113.5 | 115.5 | 115.8 |
| NCR | Smog season, Nov to Feb | 57.2 | 63.7 | 111.3 | 117.6 | 121.2 | 122.1 |
| All India | Winter, Dec to Feb | 44.9 | 47.8 | 64.8 | 65.1 | 68.5 | 72.5 |
| All India | Pre-monsoon, Mar to May | 40.0 | 39.5 | 45.9 | 44.8 | 47.4 | 47.3 |
| All India | Monsoon, Jun to Sep | 24.5 | 24.4 | 26.9 | 25.9 | 26.5 | 26.2 |
| All India | Post-monsoon, Oct to Nov | 38.0 | 38.6 | 53.6 | 55.3 | 57.3 | 59.8 |
| All India | Smog season, Nov to Feb | 44.1 | 46.7 | 64.1 | 65.5 | 68.6 | 72.6 |
READINGEach row is one season in one area. The last two columns set the rebuild beside the original on the same cells and the same years. CAUTIONThe smog season overlaps winter and post-monsoon, so the rows are not independent. All India's rebuild runs lower than the original in winter and the smog season, at 68.5 against 72.5 in winter, 2010 to 2022. SOURCESource: ledger l17_seasons.csv; the column for 2010 to September 2026 is model AI to 2022 and model BI from 2023; read 21 November 2026. THE POINTEvery area's winter stood well above its 1980s level in each period from 2000. |
Against the original, Delhi's rebuild reads a little higher in the cold months of 2010 to 2022. It gives 145.5 for winter against 137.3, and 153.6 for the smog season against 144.5.
Source: ledger l17_seasons.csv, cells the original also covers, 2010 to 2022; facts.json seasons; read 21 November 2026. Analysis in-house.
From 2010 to September 2026, the highest district averages in NCR were Delhi at 91.7, Baghpat at 90.0, Gautam Buddha Nagar at 89.9 and Ghaziabad at 88.7. Sonepat followed at 87.5, Panipat at 86.6 and Faridabad at 85.4. The lowest were Alwar and Mahendragarh, at 61.6 each.
In the 1980s every district lay between 41.1, in Muzaffarnagar, and 52.2, in Bhiwani. Every one of the 20 districts stayed above India's standard of 40 in all four periods.
For Delhi from 2023, the monitors stand beside the rebuild. On their own station-days in 2023 to September 2026, CPCB's Delhi monitors read 93.9 and model BI reads 88.2.
Source: ledger l17_levels.csv, rebuild, all cells, NCR districts on FAO GAUL 2015 boundaries; facts.json districts and cpcb_beside; the last period is model AI to 2022 and model BI from 2023; read 21 November 2026. Analysis in-house.
READINGOn both rows, NCR turns from mostly 40 to 60 micrograms in the 1980s to mostly 60 to 100 from the 2000s. The darkest cells lie around Delhi and to its east. CAUTIONThe rebuild's last map runs to September 2026 on model AI to 2022 and model BI from 2023, while the original's ends in 2022. The eastern districts read lower on the rebuild than on the original. SOURCESource: the per-cell store l17_cell_levels.npz; ledger l17_levels.csv; FAO GAUL 2015 districts; read 21 November 2026. THE POINTBoth maps show the whole of NCR well above India's standard from the 2000s onward. |
The districts around Delhi rose most. Ghaziabad went from 45.9 to 88.7, Baghpat from 46.5 to 90.0 and Meerut from 43.6 to 81.4. The south-western districts rose least, with Bhiwani going from 52.2 to 68.9 and Alwar from 46.4 to 61.6.
The two series rank the eastern districts differently. In 2010 to 2022, the rebuild reads above the original on the Haryana side, by 7.5 per cent in Sonepat. It reads below it in western Uttar Pradesh, by 14.5 per cent in Muzaffarnagar and 10.5 per cent in Meerut.
Source: facts.json districts; ledger l17_levels.csv, cells the original also covers, 2010 to 2022; research draft scholar_draft_levels.md, section 3, section 4 claim 5; read 21 November 2026. Analysis in-house.
READINGEvery district's dots lie to the right of the line at 40 in every period. The districts nearest Delhi show the longest bars, from about 45 to about 90. CAUTIONThe dark dot for 2010 to September 2026 joins model AI to 2022 and model BI from 2023, and lacks October to December 2026. The scale starts near 40, not at zero. SOURCESource: ledger l17_levels.csv; facts.json districts and cpcb_beside; read 21 November 2026. THE POINTNo NCR district was below India's annual standard in any decade since 1980. |
Table 4. Average PM2.5 in each NCR district by period, with the rebuild and the original for 2010 to 2022 on the same cells.
| NCR district (state, district) | 1980s | 1990s | 2000s | 2010 to Sep 2026 | Rebuild, 2010 to 2022, cells the original covers | Original, 2010 to 2022 |
|---|---|---|---|---|---|---|
| Delhi (the office's Delhi boundary, 12 cells) | 49.3 | 54.8 | 88.6 | 91.7 | 93.4 | 89.2 |
| Uttar Pradesh, Baghpat | 46.5 | 52.0 | 85.8 | 90.0 | 92.5 | 95.3 |
| Uttar Pradesh, Gautam Buddha Nagar | 48.8 | 55.0 | 89.7 | 89.9 | 93.9 | 93.5 |
| Uttar Pradesh, Ghaziabad | 45.9 | 51.5 | 85.2 | 88.7 | 91.9 | 97.3 |
| Haryana, Sonepat | 48.0 | 52.6 | 85.3 | 87.5 | 89.5 | 83.3 |
| Haryana, Panipat | 45.2 | 49.5 | 82.7 | 86.6 | 89.0 | 87.0 |
| Haryana, Faridabad | 49.6 | 54.7 | 86.6 | 85.4 | 89.2 | 87.1 |
| Uttar Pradesh, Bulandshahr | 45.9 | 51.4 | 84.8 | 84.5 | 89.2 | 96.7 |
| Haryana, Gurgaon | 49.9 | 53.9 | 82.9 | 81.7 | 84.8 | 80.5 |
| Uttar Pradesh, Meerut | 43.6 | 48.1 | 77.1 | 81.4 | 84.5 | 94.5 |
| Haryana, Jhajjar | 50.2 | 53.4 | 80.7 | 80.1 | 82.3 | 76.7 |
| Haryana, Rohtak | 50.1 | 52.9 | 79.7 | 78.9 | 81.2 | 76.2 |
| Rajasthan, Bharatpur | 49.6 | 51.3 | 76.5 | 74.2 | 78.4 | 74.8 |
| Haryana, Rewari | 49.9 | 52.0 | 74.6 | 73.2 | 75.5 | 70.4 |
| Haryana, Jind | 47.0 | 49.1 | 73.1 | 71.6 | 73.8 | 72.9 |
| Uttar Pradesh, Muzaffarnagar | 41.1 | 44.3 | 67.3 | 71.6 | 74.1 | 86.7 |
| Haryana, Karnal | 42.0 | 44.9 | 69.1 | 71.6 | 73.7 | 77.8 |
| Haryana, Bhiwani | 52.2 | 52.3 | 72.2 | 68.9 | 71.3 | 70.1 |
| Haryana, Mahendragarh | 48.3 | 47.8 | 64.3 | 61.6 | 63.2 | 60.6 |
| Rajasthan, Alwar | 46.4 | 46.8 | 63.4 | 61.6 | 63.2 | 59.8 |
READINGEach row is one district, ordered from the highest last-period level to the lowest. CAUTIONNewer districts sit inside their parents on the 2015 boundaries, so Palwal is in Faridabad and Nuh is in Gurgaon. Delhi is the office's 12-cell Delhi boundary. SOURCESource: ledger l17_levels.csv; the column for 2010 to September 2026 is model AI to 2022 and model BI from 2023; read 21 November 2026. THE POINTThe whole of NCR, district by district, lies far above India's standard. |
From 2010 to September 2026, the NCT of Delhi had the highest state average, at 92.2. Haryana followed at 71.9, Bihar at 71.4 and Uttar Pradesh at 69.6. Punjab and West Bengal stood at 63.0 and 62.9. The lowest were the Andaman and Nicobar Islands at 19.5 and Arunachal Pradesh at 20.2.
For the NCT of Delhi from 2023, the monitors stand beside the rebuild. On their own station-days in 2023 to September 2026, CPCB's Delhi monitors read 93.9 and model BI reads 88.2. The NCT of Delhi reads 92.2 against 91.7 for the office's 12-cell Delhi, because the two boundaries take in slightly different cells.
Source: ledger l17_levels.csv, rebuild, all cells, states on 2011 census boundaries; facts.json states and cpcb_beside; the last period is model AI to 2022 and model BI from 2023; read 21 November 2026. Analysis in-house.
In the 1980s, 7 of the 34 states and union territories with a figure had an average above 40. From 2010 to September 2026, 16 did. Two of those 16, Daman and Diu and Dadra and Nagar Haveli, hold only a few cells each.
Rajasthan was high from the start, at 56.8 in the 1980s and 59.5 in the last period, most likely from desert dust. Jammu and Kashmir barely changed, from 31.1 to 29.7.
Source: facts.json states and states_above_40; ledger l17_levels.csv; read 21 November 2026. Analysis in-house.
READINGThe states of the Indo-Gangetic plain moved furthest to the right of the line at 40. The black tick for the original sits close to the rebuild's dark dot for most states. CAUTIONThe diamond for 2010 to September 2026 joins model AI to 2022 and model BI from 2023. Daman and Diu and Lakshadweep are left out, because the original covers no cell there. SOURCESource: ledger l17_levels.csv, cells the original also covers; 2011 census boundaries; read 21 November 2026. THE POINTThe number of states above India's standard rose from 7 to 16, led by the northern plain. |
| Table 5. Average PM2.5 in each state and union territory by period, with the rebuild and the original for 2010 to 2022 on the same cells. | ||||||
|---|---|---|---|---|---|---|
| State or Union Territory (2011 boundaries) | 1980s | 1990s | 2000s | 2010 to Sep 2026 | Rebuild, 2010 to 2022, cells the original covers | Original, 2010 to 2022 |
| NCT of Delhi | 49.3 | 54.8 | 88.8 | 92.2 | 93.9 | 89.9 |
| Haryana | 47.6 | 49.6 | 72.6 | 71.9 | 74.1 | 73.2 |
| Bihar | 42.3 | 47.0 | 71.6 | 71.4 | 75.3 | 76.0 |
| Uttar Pradesh | 42.5 | 46.3 | 71.0 | 69.6 | 74.0 | 80.7 |
| Punjab | 43.1 | 45.3 | 63.5 | 63.0 | 64.7 | 65.5 |
| West Bengal | 39.2 | 42.9 | 61.5 | 62.9 | 65.2 | 65.5 |
| Rajasthan | 56.8 | 54.4 | 65.4 | 59.5 | 63.0 | 63.5 |
| Jharkhand | 37.2 | 40.8 | 54.8 | 55.6 | 57.3 | 58.8 |
| Gujarat | 46.8 | 45.7 | 52.4 | 50.1 | 52.3 | 52.1 |
| Daman & Diu | 35.8 | 37.0 | 45.4 | 46.6 | not covered | not covered |
| Tripura | 30.8 | 31.8 | 42.7 | 45.2 | 46.5 | 50.3 |
| Madhya Pradesh | 35.4 | 36.1 | 45.5 | 45.0 | 46.0 | 48.3 |
| Odisha | 32.4 | 34.3 | 42.3 | 44.6 | 45.3 | 47.2 |
| Assam | 30.8 | 32.4 | 40.0 | 43.2 | 43.6 | 47.4 |
| Chhattisgarh | 32.4 | 34.0 | 41.3 | 42.3 | 43.1 | 45.4 |
| Dadara & Nagar Havelli | 33.3 | 33.6 | 40.9 | 42.1 | 43.2 | 47.4 |
| Maharashtra | 32.5 | 32.5 | 38.7 | 39.7 | 40.3 | 42.0 |
| Meghalaya | 25.2 | 26.4 | 34.1 | 36.6 | 36.7 | 42.5 |
| Andhra Pradesh | 28.8 | 28.7 | 34.7 | 36.1 | 36.5 | 37.6 |
| Uttarakhand | 26.7 | 27.3 | 31.8 | 31.8 | 32.8 | 31.5 |
| Goa | 25.3 | 25.7 | 29.7 | 31.6 | 32.1 | 32.6 |
| Himachal Pradesh | 29.1 | 29.5 | 32.1 | 31.6 | 33.2 | 28.7 |
| Karnataka | 25.6 | 25.6 | 29.6 | 30.6 | 30.9 | 31.6 |
| Jammu & Kashmir | 31.1 | 31.1 | 32.2 | 29.7 | 31.5 | 25.9 |
| Mizoram | 23.5 | 23.5 | 28.2 | 29.4 | 29.5 | 34.1 |
| Tamil Nadu | 23.7 | 23.5 | 27.1 | 29.0 | 29.0 | 30.0 |
| Puducherry | 23.3 | 23.1 | 26.1 | 28.3 | 28.1 | 27.6 |
| Kerala | 21.5 | 22.2 | 25.9 | 28.1 | 28.1 | 28.9 |
| Lakshadweep | 20.2 | 20.8 | 23.9 | 26.5 | not covered | not covered |
| Nagaland | 19.0 | 19.5 | 23.1 | 24.9 | 24.8 | 29.2 |
| Manipur | 19.4 | 19.8 | 23.0 | 24.5 | 24.5 | 31.4 |
| Sikkim | 21.6 | 22.3 | 24.7 | 23.6 | 24.5 | 22.0 |
| Arunanchal Pradesh | 17.6 | 17.9 | 19.7 | 20.2 | 20.4 | 18.0 |
| Andaman & Nicobar Island | 17.0 | 17.8 | 19.0 | 19.5 | 19.3 | 19.6 |
READINGEach row is one state or union territory, ordered from the highest last-period level to the lowest. CAUTIONAndhra Pradesh includes Telangana and Jammu and Kashmir includes Ladakh. In Jammu and Kashmir the original leaves 1,153 of 2,171 cells blank, so its comparison rests on under half of its land. SOURCESource: ledger l17_levels.csv; 2011 census boundaries; the column for 2010 to September 2026 is model AI to 2022 and model BI from 2023; read 21 November 2026. THE POINTThe highest levels lie in a band of states across the north, from Punjab to West Bengal. |
On the cells both cover, most states lie within 10 per cent of the original in 2010 to 2022. Uttar Pradesh reads 74.0 on the rebuild against 80.7 on the original. The widest gaps lie in the hills and the north-east, where the original's monitors are few.
Source: facts.json states, rebuild_common and original_2010_2022, and blank_by_state; research draft scholar_draft_levels.md, section 3, section 5 claim 4; read 21 November 2026. Analysis in-house.
In every full year from 2023, the rebuild reads below CPCB's Delhi monitors. On the same station-days, model BI gives 94.7, 95.9 and 91.9 for 2023, 2024 and 2025. The monitors give 101.4, 104.0 and 98.7, so BI reads 6.6 to 7.7 per cent below them.
Model AI reads closer to Delhi's monitors, at 98.8, 98.4 and 93.7. The series still uses BI from 2023, because AI was trained only to 2022 and drifts upward elsewhere in later years.
Source: ledger l17_cpcb_check.csv, Delhi, 2023 to 2025, the monitors whose nearest cell lies in the 12-cell Delhi boundary; read 21 November 2026. Analysis in-house.
Table 6. CPCB's Delhi monitors beside model BI and model AI, year by year from 2023, on the monitors' own station-days.
| Year | Monitors | Station-days | CPCB monitors | Model BI (the rebuild) | Model AI | BI against monitors, per cent | AI against monitors, per cent |
|---|---|---|---|---|---|---|---|
| 2023 | 35 | 11290 | 101.4 | 94.7 | 98.8 | -6.6 | -2.6 |
| 2024 | 35 | 11725 | 104.0 | 95.9 | 98.4 | -7.7 | -5.4 |
| 2025 | 36 | 11550 | 98.7 | 91.9 | 93.7 | -6.9 | -5.1 |
| 2026, Jan to Sep | 40 | 10019 | 68.0 | 67.6 | 72.1 | -0.6 | +6.0 |
| 2023 to Sep 2026, pooled | 44584 | 93.9 | 88.2 | 91.4 | -6.0 | -2.7 |
READINGEach row is one year or one pooled span, with the monitors, the two models and the gap of each model to the monitors. CAUTIONThe 2026 row covers January to September only. The monitors' mean is taken where and when they reported, while the grid level covers every cell and every day, so the two are different averages. SOURCESource: ledger l17_cpcb_check.csv, Delhi; read 21 November 2026. THE POINTWhere Delhi's PM2.5 can be checked against monitors, the rebuild is the lower reading. |
The nine months from January to September give a fair comparison across years. In those months of 2023, 2024 and 2025, the rebuild again reads below Delhi's monitors. It gives 69.3 against 74.0, 75.4 against 80.3 and 65.1 against 70.4.
In January to September 2026, the rebuild reads 67.6 and the monitors 68.0. That gap is far smaller than in the same months of the three years before. So 2026 is the unusual year, and the reason is not known. One fact on record is that Delhi had 40 monitors in these months of 2026, against 35 in 2023 to 2025.
At the monitors, January to September 2026 read lower than the same months of 2023, 2024 and 2025. This is a reading of nine months at a changing set of monitors. It is not a trend, and it is never set against a full year.
Source: ledger l17c_cpcb_windows.csv, Delhi, January to September; facts.json jan_sep; model BI; read 21 November 2026. Analysis in-house.
| Table 7. PM2.5 in January to September of each year, 2018 to 2026, at CPCB's monitors and in the rebuild on the same station-days, for Delhi, NCR and all India. | ||||||||
|---|---|---|---|---|---|---|---|---|
| Year | Area | Monitors | CPCB monitors, Jan to Sep | Rebuild | Rebuild against monitors, per cent | Model AI | Model BI | Original |
| 2018 | Delhi | 29 | 82.0 | 77.2 | -5.9 | 77.2 | 75.1 | 74.9 |
| 2019 | Delhi | 31 | 83.8 | 78.3 | -6.5 | 78.3 | 77.1 | 78.0 |
| 2020 | Delhi | 32 | 65.4 | 66.1 | +1.0 | 66.1 | 65.6 | 64.9 |
| 2021 | Delhi | 34 | 81.5 | 80.4 | -1.3 | 80.4 | 79.3 | 78.9 |
| 2022 | Delhi | 34 | 79.4 | 76.2 | -4.1 | 76.2 | 75.2 | 74.8 |
| 2023 | Delhi | 35 | 74.0 | 69.3 | -6.3 | 73.9 | 69.3 | not covered |
| 2024 | Delhi | 35 | 80.3 | 75.4 | -6.1 | 82.2 | 75.4 | not covered |
| 2025 | Delhi | 35 | 70.4 | 65.1 | -7.5 | 69.8 | 65.1 | not covered |
| 2026 | Delhi | 40 | 68.0 | 67.6 | -0.6 | 72.1 | 67.6 | not covered |
| 2018 | NCR | 49 | 84.4 | 77.2 | -8.5 | 77.2 | 74.6 | 74.9 |
| 2019 | NCR | 71 | 79.0 | 74.0 | -6.4 | 74.0 | 72.0 | 74.1 |
| 2020 | NCR | 78 | 61.0 | 62.8 | +2.9 | 62.8 | 61.7 | 62.2 |
| 2021 | NCR | 80 | 75.4 | 76.8 | +1.8 | 76.8 | 74.8 | 75.9 |
| 2022 | NCR | 83 | 73.9 | 73.2 | -0.9 | 73.2 | 71.7 | 72.7 |
| 2023 | NCR | 84 | 65.3 | 65.1 | -0.4 | 70.8 | 65.1 | not covered |
| 2024 | NCR | 85 | 70.0 | 70.4 | +0.5 | 79.2 | 70.4 | not covered |
| 2025 | NCR | 77 | 62.0 | 62.7 | +1.3 | 69.0 | 62.7 | not covered |
| 2026 | NCR | 104 | 66.7 | 63.6 | -4.8 | 68.7 | 63.6 | not covered |
| 2018 | All India | 121 | 66.2 | 59.3 | -10.5 | 59.3 | 56.0 | 59.1 |
| 2019 | All India | 191 | 60.0 | 56.1 | -6.5 | 56.1 | 53.5 | 57.5 |
| 2020 | All India | 217 | 45.4 | 48.0 | +5.7 | 48.0 | 46.0 | 48.3 |
| 2021 | All India | 285 | 52.1 | 54.0 | +3.5 | 54.0 | 50.9 | 54.5 |
| 2022 | All India | 346 | 52.0 | 53.0 | +2.0 | 53.0 | 50.3 | 54.1 |
| 2023 | All India | 480 | 45.8 | 45.3 | -1.0 | 50.4 | 45.3 | not covered |
| 2024 | All India | 507 | 44.1 | 45.2 | +2.6 | 51.6 | 45.2 | not covered |
| 2025 | All India | 504 | 39.0 | 41.8 | +7.3 | 47.9 | 41.8 | not covered |
| 2026 | All India | 539 | 40.9 | 43.4 | +6.0 | 50.1 | 43.4 | not covered |
READINGEach row is the same nine months of one year in one area. The rebuild is model AI to 2022 and model BI from 2023. CAUTIONThe set of monitors changes from year to year in every area. Nine months leave out October to December, the worst months, so no row is a year. SOURCESource: ledger l17c_cpcb_windows.csv; facts.json jan_sep; read 21 November 2026. THE POINTModel BI read 6 to 7.5 per cent below Delhi's monitors in these months of 2023 to 2025, and close to them in 2026. |
A fixed panel of 30 Delhi stations shows no fall in the monitors over the full years. That panel read 95.4 in 2020, 98.1 in 2022 and 97.9 in 2025.
Source: ledger l13_cpcb_delhi.csv, set fixed panel 2020 to 2025, as given in the research draft scholar_draft_levels.md, section 3, section 6 claim 3; read 21 November 2026. Analysis in-house.
In Delhi in 2018 to 2022, both models read below the monitors. On the same station-days the monitors read 102.9, model AI 99.2 and the original 96.9. Across NCR the monitors read 96.1, AI 94.5 and the original 93.3. Across all India they read 65.9, AI 65.8 and the original 66.9.
From 2023, model AI reads well above model BI on the monitors' own days. Across 2023 to 2025, AI gives 93.9 in NCR against BI's 87.6, and 59.6 across all India against BI's 52.7. The monitors give 87.9 and 51.0. This gap between the two models is the step that every spliced chart shows at 2023.
Source: facts.json cpcb_2018_2022 and cpcb_pools; ledger l17_cpcb_check.csv; NCR is the outline of the 20 NCR districts; read 21 November 2026. Analysis in-house.
READINGIn Delhi the black line of the monitors sits above both models in every full year. From 2023 the blue line of model AI runs above the red line of model BI in every area. CAUTIONOn the top row the 2026 points are January to September alone, and they are not joined to 2025. The set of monitors grew in every area, so the all-India monitor line mixes new places with any change in the air. SOURCESource: ledgers l17_cpcb_check.csv and l17c_cpcb_windows.csv; CPCB station-days through CREA's mirror; read 21 November 2026. THE POINTThe step at 2023 comes from the change of model, and Delhi's monitors read above the rebuild throughout. |
The model smooths the extremes. In Delhi in 2018 to 2022, 6.6 per cent of station-days at the monitors were above 250, against 4.6 per cent for model AI. Above 120 the figures were 30.5 and 27.8 per cent. Above 60 the model counted more, with 64.0 per cent against the monitors' 61.3.
The same pattern held on model BI in 2023 to 2025. The monitors recorded 6.7 per cent of station-days above 250, and the rebuild 3.5 per cent. So this report's days above 250 are understated, and its days above 60 run slightly high.
Source: facts.json cpcb_pools and cpcb_shares, Delhi; ledger l17c_cpcb_windows.csv; read 21 November 2026. Analysis in-house.
| Table 8. Year by year from 2018, CPCB's monitors beside the rebuild, model AI, model BI and the original, on the monitors' own station-days, for Delhi, NCR and all India. | |||||||
|---|---|---|---|---|---|---|---|
| Year | Area | Monitors | CPCB monitors | Rebuild | Model AI | Model BI | Original |
| 2018 | Delhi | 30 | 113.0 | 107.8 | 107.8 | 105.8 | 105.1 |
| 2019 | Delhi | 31 | 107.4 | 101.3 | 101.3 | 100.5 | 99.2 |
| 2020 | Delhi | 34 | 94.7 | 94.3 | 94.3 | 93.7 | 91.1 |
| 2021 | Delhi | 34 | 104.0 | 102.2 | 102.2 | 101.3 | 100.1 |
| 2022 | Delhi | 34 | 97.3 | 92.2 | 92.2 | 91.3 | 90.7 |
| 2023 | Delhi | 35 | 101.4 | 94.7 | 98.8 | 94.7 | not covered |
| 2024 | Delhi | 35 | 104.0 | 95.9 | 98.4 | 95.9 | not covered |
| 2025 | Delhi | 36 | 98.7 | 91.9 | 93.7 | 91.9 | not covered |
| 2026, Jan to Sep | Delhi | 40 | 68.0 | 67.6 | 72.1 | 67.6 | not covered |
| 2018 | NCR | 53 | 114.8 | 107.7 | 107.7 | 105.1 | 105.2 |
| 2019 | NCR | 73 | 101.9 | 96.0 | 96.0 | 94.1 | 95.2 |
| 2020 | NCR | 80 | 88.3 | 89.2 | 89.2 | 87.8 | 87.0 |
| 2021 | NCR | 81 | 96.2 | 97.3 | 97.3 | 95.6 | 96.3 |
| 2022 | NCR | 83 | 88.2 | 88.3 | 88.3 | 86.7 | 88.0 |
| 2023 | NCR | 84 | 87.3 | 87.0 | 94.2 | 87.0 | not covered |
| 2024 | NCR | 85 | 86.9 | 87.2 | 94.0 | 87.2 | not covered |
| 2025 | NCR | 86 | 89.7 | 88.9 | 93.4 | 88.9 | not covered |
| 2026, Jan to Sep | NCR | 104 | 66.7 | 63.6 | 68.7 | 63.6 | not covered |
| 2018 | All India | 128 | 83.4 | 76.7 | 76.7 | 73.4 | 76.7 |
| 2019 | All India | 196 | 73.1 | 68.5 | 68.5 | 65.7 | 70.1 |
| 2020 | All India | 241 | 61.5 | 63.4 | 63.4 | 60.8 | 63.5 |
| 2021 | All India | 306 | 63.6 | 65.4 | 65.4 | 62.0 | 66.6 |
| 2022 | All India | 376 | 60.6 | 62.4 | 62.4 | 59.0 | 64.3 |
| 2023 | All India | 500 | 55.1 | 55.2 | 61.8 | 55.2 | not covered |
| 2024 | All India | 517 | 50.8 | 52.5 | 59.4 | 52.5 | not covered |
| 2025 | All India | 521 | 47.5 | 50.6 | 57.8 | 50.6 | not covered |
| 2026, Jan to Sep | All India | 539 | 40.9 | 43.4 | 50.1 | 43.4 | not covered |
READINGEach row is one year in one area, with the count of monitors first. The rebuild column is model AI to 2022 and model BI from 2023. CAUTIONThe 2026 rows cover January to September only. All India had 128 monitors in 2018 and 539 by 2026, so its monitor mean is not a fixed set of places. SOURCESource: ledger l17_cpcb_check.csv; read 21 November 2026. THE POINTDelhi's monitors read above every model in every full year. |
Table 9. The share of station-days above 60, 120 and 250 micrograms, at the monitors and in the rebuild, pooled over 2018 to 2022, 2023 to 2025 and January to September 2026.
| Years | Area | Station-days | Above 60: monitors / rebuild | Above 120: monitors / rebuild | Above 250: monitors / rebuild |
|---|---|---|---|---|---|
| 2018 to 2022 (model AI) | Delhi | 53,878 | 61.3 / 64.0 | 30.5 / 27.8 | 6.6 / 4.6 |
| 2023 to 2025 (model BI) | Delhi | 34,565 | 59.4 / 61.0 | 29.5 / 26.4 | 6.7 / 3.5 |
| 2026, Jan to Sep (model BI) | Delhi | 10,019 | 42.9 / 44.5 | 11.6 / 9.7 | 1.3 / 0.7 |
| 2018 to 2022 (model AI) | NCR | 117,126 | 58.4 / 62.4 | 27.1 / 25.2 | 5.4 / 3.7 |
| 2023 to 2025 (model BI) | NCR | 80,140 | 53.5 / 58.4 | 22.9 / 23.1 | 4.3 / 2.2 |
| 2026, Jan to Sep (model BI) | NCR | 24,332 | 43.7 / 40.6 | 10.8 / 7.7 | 0.9 / 0.4 |
| 2018 to 2022 (model AI) | All India | 346,840 | 38.7 / 41.5 | 13.3 / 11.6 | 2.2 / 1.3 |
| 2023 to 2025 (model BI) | All India | 449,099 | 27.6 / 30.1 | 6.6 / 5.2 | 0.9 / 0.4 |
| 2026, Jan to Sep (model BI) | All India | 121,745 | 18.1 / 17.7 | 2.6 / 1.6 | 0.2 / 0.1 |
READINGEach cell gives the monitors' share first and the rebuild's share second, both in per cent of the same station-days. CAUTIONThese are shares of station-days at the monitors, which are mostly urban. They are not days a year at a cell, so they do not convert directly to this report's counts. SOURCESource: ledger l17c_cpcb_windows.csv; facts.json cpcb_pools and cpcb_shares; read 21 November 2026. THE POINTThe rebuild counts too few of the worst days, so its very poor and severe counts are low. |
The all-India monitor mean and the all-India grid mean measure different things. The monitors sit mostly in towns, and the grid covers every cell of India's land. So the two are never set against each other as a check.
On the level, the rebuild and the original agree within 5 per cent in every period, for Delhi, NCR and India. For Delhi the rebuild reads 2.9 per cent below the original in the 1980s and 4.8 per cent above it in 2010 to 2022. For India it reads 0.5 per cent above in the 1980s and 2.6 per cent below in 2010 to 2022.
Every comparison uses 2010 to 2022, because the original ends in 2022. It also uses only the cells the original covers, because the original leaves 1,204 of India's 28,885 cells blank.
Source: facts.json levels, rebuild_minus_original_pct, blank_cells_original and india_cells; ledger l17_levels.csv, cells the original also covers; read 21 November 2026. Analysis in-house.
READINGThe two rows draw the same India, with the northern plain darkening from the 2000s. The all-India average is 35.9 against 35.7 in the 1980s and 47.4 against 48.6 in 2010 to 2022. CAUTIONBoth rows end in 2022 and use model AI alone. The hills and the north-east differ most between the two. SOURCESource: the per-cell store l17_cell_levels.npz; facts.json levels; Wang et al. 2024 data on Zenodo; read 21 November 2026. THE POINTThe office's rebuild reproduces the published picture of India's PM2.5. |
Table 10. Average PM2.5 on the rebuild and the original, and the share of cells above 40 on each, by period, for Delhi, NCR and all India.
| Period | Area | Rebuild | Original | Rebuild against original, per cent | Cells above 40, rebuild | Cells above 40, original |
|---|---|---|---|---|---|---|
| 1980 to 1989 | Delhi | 49.3 | 50.8 | -2.9 | 100 | 100 |
| 1990 to 1999 | Delhi | 54.8 | 55.1 | -0.4 | 100 | 100 |
| 2000 to 2009 | Delhi | 88.6 | 87.0 | +1.8 | 100 | 100 |
| 2010 to 2022 | Delhi | 93.4 | 89.2 | +4.8 | 100 | 100 |
| 1980 to 1989 | NCR | 47.1 | 48.2 | -2.3 | 96 | 100 |
| 1990 to 1999 | NCR | 49.8 | 50.8 | -1.9 | 100 | 100 |
| 2000 to 2009 | NCR | 75.1 | 77.2 | -2.7 | 100 | 100 |
| 2010 to 2022 | NCR | 77.9 | 78.2 | -0.4 | 100 | 100 |
| 1980 to 1989 | All India | 35.9 | 35.7 | +0.5 | 29 | 33 |
| 1990 to 1999 | All India | 36.6 | 36.9 | -0.9 | 32 | 36 |
| 2000 to 2009 | All India | 46.1 | 47.3 | -2.5 | 57 | 62 |
| 2010 to 2022 | All India | 47.4 | 48.6 | -2.6 | 63 | 67 |
READINGEach row is one area in one period, with the rebuild, the original and the gap between them in per cent. CAUTIONThe last period here is 2010 to 2022 on model AI, so it does not include model BI. A close match shows that the method was reproduced. It does not show that either map is right about any single year. SOURCESource: ledger l17_levels.csv, cells the original also covers; read 21 November 2026. THE POINTOn levels, the report's findings do not depend on the office's rebuild. |
The two differ more on counts of days. Across India, the rebuild counts fewer days above 60 in every period, for example 92.8 against 105.3 in 2010 to 2022. In Delhi it counts more days above 120 from 2000, with 92.8 against 82.7 in 2010 to 2022.
Against Delhi's monitors in 2018 to 2022, the rebuild is the closer of the two. It reads 3.6 per cent below them, and the original 5.8 per cent below.
Source: facts.json days and cpcb_2018_2022; ledger l17b_cell_days.csv; read 21 November 2026. Analysis in-house.
This record ends on 30 September 2026, because NASA's MERRA-2 aerosol inputs stop on 1 October 2026. Its last period therefore lacks October to December 2026, the three worst months of Delhi's year. Each limit below is stated so that no figure in this report is read for more than it shows.
Table 11. Eight questions this record cannot answer, and the reason in each case.
| The question | Why the record cannot answer it |
|---|---|
| What happened after 30 September 2026? | The model cannot run without MERRA-2, whose aerosol inputs stop on 1 October 2026. |
| Did PM2.5 fall from 2022 to 2023? | The step at 2023 is the change from model AI to model BI. On the grid in 2018 to 2022, AI reads 11.2 to 12.4 per cent above BI over all India. |
| What was the level in any single year before 2018? | Both models are back-casts trained on monitors from 2018. The original's own tests give R squared of 0.66 for years it had not seen. |
| How many people breathed these levels? | The averages are over land, cell by cell, and carry no population weight. |
| What caused the rise? | The model reads reanalysis and aerosol inputs, and it does not attribute PM2.5 to a source. |
| Did a cell meet a monitor's standard? | India's 24-hour standard and CPCB's bands are written for a monitoring site, and are applied here to cell averages. |
| Was a season above a limit? | India's limits apply to a year or a day, so a season's average is printed as a level only. |
| Does the all-India grid match the all-India monitors? | The monitors sit mostly in towns, and the grid covers all of India's land. |
READINGEach row names one question and the fact that keeps the record from answering it. CAUTIONThese are limits of the record, and none of them reverses a finding about the level over the decades. SOURCESource: facts.json cpcb_pools and ai_minus_bi_grid; Wang et al. 2024, Table 3, page 3570; read 21 November 2026. THE POINTThe decade averages stand, and single years before 2018 are read only as part of a decade. |
Every value before 2018 is a back-cast, which means a model trained on recent years and run on the inputs of older years. NASA's MERRA-2 took in no satellite aerosol data before 2000. So part of the step from the 1990s to the 2000s may come from the inputs, and the levels before 2000 may be understated.
Source: Wang et al. 2024, section 4.3 and Table 3, page 3570; PROGRESS.md of the rebuild, entry of 21 November 2026 on MERRA-2; read 21 November 2026.
The rebuild follows the recipe in the LongPMInd paper. Model AI is LightGBM, a method that builds many small decision trees in turn. It learned from CPCB station-days across India in 2018 to 2022, as the paper's model did. Model BI uses the same recipe and learned from 2018 to September 2026.
Shri Ajay Maken chose the series on 7 November 2026 and again on 21 November 2026. It runs on model AI from 1 January 1980 to 31 December 2022, and on model BI from 1 January 2023 to 30 September 2026. AI is the closer copy of the published grid. On the monitors in 2023 to September 2026, which it had not seen, AI gave an R squared of 0.430 and read 17.8 per cent high, so it is not carried past 2022.
Source: PROGRESS.md of the rebuild, decisions and the entry of 7 November 2026, as given in the research draft scholar_draft_levels.md, section 4, Annexure A; Wang et al. 2024, sections 2.1 and 2.2; read 21 November 2026.
Table 12. The definitions behind every figure in this report.
| Term | How it is computed |
|---|---|
| Cells | India's 28,885 cells on the paper's own grid, about 10 km apart. Delhi is 12 cells inside the office's Delhi boundary, and NCR is 599 cells. |
| Cells the original covers | The 27,681 cells the original fills. Of the 1,204 blank cells, 1,153 lie in Jammu and Kashmir. |
| Regions | 58 in all: all India, Delhi, NCR, NCR outside Delhi, 20 NCR districts on FAO GAUL 2015 boundaries, and 34 states and union territories on 2011 census boundaries. |
| Chandigarh | Chandigarh is too small to hold a 10 km cell of its own, so it carries no figure. |
| Level | The mean of the area's monthly means over every month of the period, each month weighted once. |
| Times WHO, times India | The level divided by WHO's annual guideline of 5, or by India's annual standard of 40. |
| Cells above 40 | The share of the area's cells whose own period mean is above 40. |
| Days above a limit | At each cell, the days whose daily mean was above 60, 120 or 250, scaled to 365.25 days, then averaged over the area's cells. |
| Seasons | Winter, December to February; pre-monsoon, March to May; monsoon, June to September; post-monsoon, October and November; smog season, November to February. |
| Periods | 1980 to 1989, 1990 to 1999, 2000 to 2009, and 2010 to 30 September 2026. Comparisons with the original use 2010 to 2022, and 2010 to 2025 is the full-year check. |
| CPCB check | Each station sits on its nearest cell. A station-day needs 16 valid hours, and the monitors and the models are averaged over the same station-days. |
READINGEach row names one term and says how the office computed it. CAUTIONThese are choices the office made, and each can be changed and run again. SOURCESource: ledgers l17_levels.csv, l17b_cell_days.csv, l17_seasons.csv and l17_cpcb_check.csv, and the scripts that write them; read 21 November 2026. THE POINTEvery figure in the report can be traced to one of these definitions and one ledger row. |
Days above WHO's 24-hour guideline of 15 are not counted, because Delhi's average was above 15 on nearly every day of every period. The year 2025 has 364 days in the record, because MERRA-2 holds no hours for 5 September 2025.
Source: ledger l17_days_above.csv, column days_gt_15; research draft scholar_draft_levels.md, section 4, Annexure A; read 21 November 2026.
The ledgers named in the source lines hold every number the tables print, and facts.json holds every pooled figure the text quotes. The table below names each file and says what it holds. The office keeps all of these files.
| What it holds | File |
|---|---|
| Levels, days, seasons, annual means and the CPCB check | l17_level_metrics.py |
| Days above the limits, counted at each cell | l17b_cell_days.py |
| Figures and numbers for this report | r7_levels_assets.py |
| This report's builder and its checks | r7_build_levels.py |
| Average PM2.5 by period and area | l17_levels.csv |
| Days the area's average exceeded each limit | l17_days_above.csv |
| Days above each limit, counted at each cell | l17b_cell_days.csv |
| Season levels by period and area | l17_seasons.csv |
| Annual means by area | l17_annual.csv |
| Monthly means by area | l17_monthly_regions.csv |
| CPCB monitors against the rebuild and the original, 2018 to 2026 | l17_cpcb_check.csv |
| The same check by January to September, with shares above 60, 120 and 250 | l17c_cpcb_windows.csv |
| The script behind that check | l17c_cpcb_windows.py |
| Figures and numbers behind this report | facts.json |
| Training and held-out tests, all-India models | l7_india_vs_ncr.py |
| The original LongPMInd grids, monthly | longpmind |
Each source below is named with the page or table this report relies on. The office's own ledgers are listed in Annexure B.
Table 13. The sources this report relies on, with the place each was read.
| Source | What the report uses | Where in the source |
|---|---|---|
| Wang, S. and others (2024), LongPMInd, Earth System Science Data 16, pages 3565 to 3577 | The method, the training years and the held-out tests | Section 2.1, page 3567; Table 3 and section 4.3, page 3570 |
| CPCB hourly PM2.5, through CREA's mirror of the CPCB feed | Station-days from 2018 to 30 September 2026 | Ledgers l17_cpcb_check.csv and l17c_cpcb_windows.csv |
| India's National Ambient Air Quality Standards, CPCB notification of 18 November 2009 | PM2.5 annual standard of 40 and 24-hour standard of 60 | Page 3, table row 4 |
| WHO Global Air Quality Guidelines, 2021 | PM2.5 annual guideline of 5 and 24-hour guideline of 15 | Table 0.1, page xvii |
| CPCB, National Air Quality Index, 2014 | PM2.5 bands: very poor 121 to 250, severe above 250 | Table 3.11, page 35 |
| Press Information Bureau, National Clean Air Programme launch, 10 January 2019 | The target of a 20 to 30 per cent fall by 2024 against 2017 | Local text copy in reports/sources |
| Reanalysis inputs and NASA MERRA-2, through Google Earth Engine | The model's reanalysis and aerosol inputs | As listed in the paper's Table 1, page 3568 |
READINGEach row names one source and the part of it that this report uses. CAUTIONThe paper's page numbers were read from the office's local copy on 6 and 7 November 2026. SOURCESource: the documents named in each row; read 21 November 2026. THE POINTEvery outside fact in the report can be traced to a named page or table. |