How Dhaan reads a rice field
Every number on the map as a picture and one line. "How exactly" opens the working.
Where rice is
33.9million ha, 2024
A published 30 m satellite map, counted pixel by pixel inside each district. Land once; a field grown twice is two hectare-seasons.
How exactly
A published 30 m map (Landsat and Sentinel-2 phenology, 84–88% accurate), counted pixel by pixel inside each district with the pixel's own area. Land is counted once; a field grown twice is two hectare-seasons.
The season today
22states in harvest
Each district takes its state's crop-calendar stage today. The radar clock replaces it district by district next.
How exactly
Each district takes the stage of its state's crop calendar today: sown, growing, flowering, harvest, fallow. Radar (Sentinel-1 sees a paddy flood and sees it cut) replaces the calendar district by district next; until then every page says by state calendar.
- fallow 10
- harvest 22
- sowing 1
- growing 1
Straw
207Mt a year
Official rice crop × 1.5 straw per tonne of grain, apportioned to districts by hectare-seasons.
How exactly
1,378.25 lakh t of rice (DES 2023-24) × 1.5 straw per tonne of grain, apportioned to districts by hectare-seasons. Under-counts Punjab's high yields until state production replaces it.
Methane, the chain
IPCC Tier 1; the range is the water regime we cannot yet see from orbit. Radar days-flooded narrows it.
How exactly
IPCC 2019 Tier 1. The range is the water regime we cannot yet see: multi-season fields are irrigated (single drainage 0.71 to flooded 1); single-season fields may be rainfed (0.54) or flooded. Days flooded from radar narrows it; AWD fields take 0.55. Most: Purba Bardhaman 38 kt, Paschim Medinipur 30 kt, Bathinda 27 kt, Birbhum 26 kt, Ludhiana 23 kt. Punjab 238–421 kt.
Dhuan Lipa
counted, burned, carried, breathed
Fires on rice land
NASA FIRMS, open files. A detection counts when a rice pixel lies within 200 m. Series use Suomi NPP alone, so years compare.
How exactly
NASA FIRMS, open files, no key. A detection counts when a 2024 rice pixel lies within ~200 m (a VIIRS pixel is 375 m). Series use Suomi NPP alone so years compare (NOAA-20 from 2020, NOAA-21 from 2023; S-NPP ends 1 Nov 2026, the series then switches with the break marked). Official bulletins use the same passes and miss the same evening fires.
Fields burned · being re-measured
No figure yet
Our first rule called a field burned when its Sentinel-2 burn index fell; it also took fields only cut or ploughed. A stricter rule, asking for dark ash, misses fires under smog. Neither is good enough to print.
How exactly
First rule: on rice pixels at 20 m, NBR = (B8A − B12)/(B8A + B12) falls by more than 0.10 between the last clear look before a window and the first in it, and ends under 0.10. A cut field loses its green too, so it passes. Stricter rule: the after-look must also be ash-like (NBR below −0.05, near-infrared reflectance under 0.18); under the haze of the peak weeks nothing reads that dark. What would settle it: field truth, such as the Punjab Remote Sensing Centre's burnt-area maps; a measure that compares each field with unburned ground in the same hazy scene; or radar, which sees through smog. Cloud is always unseen, never unburned.
Where the smoke goes
48hours ahead
Our own ECMWF run; a parcel from each district's fires is followed hour by hour. Transport, never a share of a city's air.
How exactly
Our own ECMWF IFS 0.25° run, sampled hourly on a half-degree grid: 10 m (the particles) and 850 hPa, their mean as a crude boundary-layer wind. A parcel leaves each district's fire centre and is followed hour by hour; each hour over a district or a bulletin city is a fire-hour. Transport, not chemistry: never a share of a city's PM2.5.
What cities breathe
715bulletin days since 2022
CPCB's 4 pm bulletin, one AQI per city, parsed daily. The number newsrooms quote.
How exactly
CPCB's daily 4 pm bulletin, one AQI per city (24 h average, prominent pollutant), parsed from the PDF every day. It is the number newsrooms quote and it is open. Station-level hourly PM2.5 (CAAQMS) needs a key and is next.
The model: where it burns next
small, graded the house way
What it is
15features, 200 trees
Gradient-boosted trees on the atlas's own numbers, not a language model. Predicts a heavy day within 3 and 7 days.
How exactly
Gradient-boosted trees, not a language model. 2,77,840 district-days from 2019–2023; features the atlas already computes: fires in the last 1, 3, 7 days, last year's same week, season so far, rice hectares, three days of rain, temperature and humidity, the calendar's harvest stage, days to Diwali. Predicts a heavy day (≥ 20 detections at the 13:30 pass) within 3 and 7 days. Trained 29 Sept 2026.
Against the baselines, 3 days
Brier on a season it never saw; it must beat climatology and persistence or it does not ship. ships.
How exactly
Brier score on the season it never saw (2024, 55,568 district-days): lower is better. It must beat climatology (share of heavy days on that day of season, last three seasons) and persistence (heavy in the last three days) or it does not ship. ships. AUC 0.9879.
Calibration, 3 days
On the dotted line a 70% is a 70%. Calibration error 0.0036.
How exactly
What it said against what happened, in ten bins of its own confidence; on the dotted line a 70% is a 70%. Expected calibration error 0.0036. 7 days: Brier 0.0135 vs 0.0189 climatology, ships.
By state, 3-day Brier
A shorter bar than the number beside it means the model helps; where it does not, the map shows no forecast.
How exactly
Model against climatology; a shorter bar than the number beside it means the model helps. Where it does not (few heavy days), the map shows no forecast: not enough signal.
Known failures
- inherits the 13:30 overpass blind spot: predicts detections, not fires
- no 2025 season in the open archive: trained to 2023, tested on 2024
- first week of a season, before the district's own count settles
- smog weeks, when the satellite sees less
- districts with under 20 heavy days across five seasons: little signal, no forecast shown
Every day of the season the burning tab shows yesterday's forecast beside what happened.
How exactly
Every day of the season the burning tab shows yesterday's forecast beside what happened and the score so far. A model that hides its misses is worth nothing to a newsroom.
What it can't tell you
- Whether a field is rice this year. The map is 2024. Our own annual map from radar is planned after this kharif.
- When a district's harvest actually started. Calendar windows until the radar clock lands.
- How much methane, exactly. A range by design, until days flooded are measured.
- How many fields burned today. Only what the satellites saw at their passes; burned area is being re-measured.
- What share of Delhi's air is smoke. Not without a chemical transport model; not this season.
- Anything about one farmer. Nothing below block level; embers are 375 m pixels, not fields.