Methods and evidence

How every message is made, and what it cannot see.

For agronomists, scientists, funders and anyone doing diligence: what we compute, from which sources, under which rules, how each crop model is reviewed, and the known limits of the signal.

Last updated .

1 · What we compute

Sources, methods and cadence.

  1. Sentinel-2 / HLS / S-1Satellite passA new image of the field about every 5 days.
  2. cloud maskClean pixelsCloud and shadow removed; small plots flagged.
  3. NDVI, NDRE, NDWIField indicesComputed inside the field boundary only.
  4. baselineCompared with historyAgainst the field's own past weeks, plus weather.
  5. stateOne of four statesWatch, Act, No clear view or Escalate.
  6. WhatsAppOne weekly messageUnder 300 characters, with date and source.

One real message, step by step

The message of 20 May 2026 for the 3.2-acre field near Katol, traced from the satellite pass to the text. All values are this field's own readings.

  1. 1

    Satellite pass

    True-colour Sentinel-2 image of the field near Katol on 10 May 2026, outlined 10 May 2026

    Sentinel-2 photographs the field every few days. Clear passes on 10 and 20 May 2026.

  2. 2

    Clean pixels, inside the field

    Greenness (NDVI) of the same field on 10 May 2026, in 10 m squares NDVI 0.61

    129 squares of 10 m fall inside the boundary. Cloud and shadow are masked; the field's greenness is their average.

  3. 3

    Compared with its own year and its neighbours

    0.40.60.8OctDecFebAprJunAugOct 20 May

    This field Fields within about 500 m (median) 10 to 20 May

    0.61 to 0.38 in 10 days: down 38%. The fields around it fell 22%. A drop much faster than the neighbourhood is worth a visit.

  4. 4

    One state, one message

    Watch
    20 May: your field's greenness fell 38% in 10 days (0.61 to 0.38), faster than the fields around it (22%). This week: check the soil moisture and the irrigation.Sentinel-2 · built from this field's real Sentinel-2 readings of 10 and 20 May 2026

    Watch means something changed: check the field. It never carries a dose, a product or an irrigation volume.

Contains modified Copernicus Sentinel data, processed by VarshaKrishi. Field outline traced from the image; crop not confirmed on the ground.

QuantitySourceMethodCadence
Vegetation and water indices per field (NDVI, NDRE, EVI, NDWI)Sentinel-2 L2A surface reflectance (ESA Copernicus); NASA HLS to fill cloud gapsIndex formulas per the cited papers; per-field statistics over cloud-free pixels, with valid-pixel count and cloud fraction stored with every valueEvery pass (about 5 days), summarised weekly
Radar growth and standing-water stateSentinel-1 GRDBackscatter trend per field, used when optical passes are blocked by monsoon cloudEvery pass, summarised weekly
Field baseline and changeThe field's own historyChange against the field's own baseline once enough clean passes exist; until then the message says the baseline is buildingWeekly
Weather and reference evapotranspirationWeather provider forecast and observationsFAO-56 Penman-Monteith, with Hargreaves-Samani where only temperature is availableDaily
Irrigation timing (Decide only)Satellite, weather, crop coefficients, optional reference-kit soil moistureFAO-56 root-zone water balance with crop-specific coefficientsWeekly, with a decision window
Disease-risk window (Decide only, tomato)Hourly weather; leaf wetness where a kit existsTOM-CAST disease severity valuesDaily accumulation, weekly message

Every derived value is appended to a per-field, tamper-evident record chain together with the identifiers of its raw inputs, so it can be recomputed (see replay).

From reflectance to a message

Sentinel-2 records how much red and near-infrared light each 10 m square reflects. Healthy leaves absorb red and reflect near-infrared, so NDVI, (NIR minus red) divided by (NIR plus red), rises with green, dense canopy. We compute it over the cloud-free squares inside each field and compare it with that field's own history.

True-colour Sentinel-2 image of farm parcels west of Katol, Nagpur district, on 12 Oct 2026 The same area as an NDVI greenness map: dark green is dense healthy canopy, pale yellow and brown are bare or dry fields 4 km x 3 km · 10 m pixels
Real image: Farmland west of Katol, Nagpur district, Maharashtra, Sentinel-2, 12 Oct 2026. Contains modified Copernicus Sentinel data, processed by VarshaKrishi.
2 · The four states

Every message is one of four states.

StateMeaningWho can receive itExample reason
ActA decision is due within a window (irrigate within 2 days; inspect the north block)Decide fields, for crops with a reviewed modelRoot-zone water below the safe level; no rain forecast for 5 days
WatchSomething changed; check the field. No prescriptionWatch and DecideGreenness fell 12% since last week while rain stayed low
No clear viewThe satellite could not see the plot well enough this weekWatch and DecideCloud on 6 of the last 7 passes; radar shows no standing water
EscalateBeyond the model; a person should lookDecide fieldsTwo weeks of decline across the whole plot; agronomist asked to review

  • Plan rules

    A field without a reviewed crop model, or on the Watch plan, never receives Act or Escalate. Watch messages contain no dose, no product name, no pesticide and no irrigation volume.

  • A message every week

    Every active field gets one message a week, even when the satellite could not see it. Silence never means "healthy".

  • Plot size

    Below 0.5 ha the signal is approximate, so we report direction only and show how many clean pixels we used. Below 0.2 ha a field is on Watch only.

  • Monsoon

    A field without a clean optical pass for 15 days gets a radar-based message (standing water, waterlogging and growth trend), never an optical stress call.

3 · Crop models

Decide opens crop by crop, after review.

A crop moves to Decide only when an agronomist has reviewed and signed its model. The review document is then linked here.

CropRegionPlanStatus
citrus (Nagpur orange)Vidarbha, MaharashtraDecide, 2026-27 seasonModel built on FAO-56 and technically reviewed; awaiting agronomist sign-off
Rice, Wheat, Cotton, Sugarcane, Soybean, Maize, Pulses, VegetablesAll of IndiaWatchCrop calendars for stage estimates; no prescriptions
Any other cropAll of IndiaWatchTrend and weather
The limits, in real data

What 10 m pixels and monsoon cloud look like.

Two real examples from Katol, in the Nagpur orange belt, that explain the first two entries of the register below.

Sentinel-2 true-colour close-up of 600 m x 450 m of farmland near Katol, showing individual 10 m pixelsTrue colour
NDVI close-up of the same areaNDVI greenness
L1, small plots. 600 m x 450 m near Katol on 12 Oct 2026: 60 x 45 pixels of 10 m. A half-hectare plot covers only about 50 of these squares, fewer once its edges are excluded. Contains modified Copernicus Sentinel data, processed by VarshaKrishi.
Cloud cover on each Sentinel-2 pass over Katol, May to October 2026 39 passes. From late June to early September almost every pass was more than 70 percent cloud, so optical images could not see the fields; clear passes returned from late September. 70% 100% 0% May Jun Jul Aug Sep Oct
Cloud cover of each Sentinel-2 pass over tile 44QKJ (Katol), 15 May to 12 Oct 2026, from the scene metadata. Source: Copernicus Sentinel-2 scene metadata via Earth Search.

L2, monsoon cloud. From late June to early September 2026 almost every Sentinel-2 pass over Katol was too cloudy to use. In those weeks the message says so and relies on Sentinel-1 radar.

Every pass over this field for a year, free to download (open data)

4 · Limitations register

What the signal cannot do.

Numbered so a paper or a partner can cite an entry. Entries are added, never removed.

  1. L1. Small plots are approximate.Below 0.5 ha few clean 10 m pixels fall inside the boundary; we report direction only and show the valid-pixel count. Below 0.2 ha a field is on Watch only.
  2. L2. Cloud hides fields in the monsoon.Optical indices are impossible for weeks at a time. Sentinel-1 radar shows standing water, waterlogging and a coarse growth trend only, and the message says so.
  3. L3. Confidence values are expert priors until calibrated.A confidence value describes how well the signals agree, not a probability of truth. Calibration needs field measurements, which the reference plots provide.
  4. L4. Crop stage on Watch is calendar-based.Stage is inferred from the sowing date and a crop calendar, not observed.
  5. L5. Satellites do not see pests or disease directly.Optical indices show stress, not its cause. Disease-risk windows come from weather models, and the message asks the farmer to inspect rather than naming a disease.
  6. L6. Translations.Marathi, Hindi and English copy is written with AI assistance and checked against published agricultural guidance; review by native speakers and an agronomist is still pending. Bengali, Gujarati, Punjabi, Tamil and Telugu are machine translations, labelled as such in the product.
  7. L7. Reference-kit readings depend on installation.Soil-moisture probes read a small volume of soil; depth, contact and calibration dominate accuracy, so kit metadata ships with every export.
  8. L8. Weather is a third-party model.Evapotranspiration uses forecast and gridded observations, not an on-site station unless a reference kit includes one.
5 · Citations

The standards behind every number.

  1. FAO-56, crop evapotranspiration.Allen, R.G., Pereira, L.S., Raes, D., Smith, M. (1998). FAO Irrigation and Drainage Paper 56. FAO. Reference evapotranspiration, crop coefficients and the root-zone water balance.
  2. TOM-CAST and FAST, disease severity values.Madden, L., Pennypacker, S.P., MacNab, A.A. (1978), Phytopathology 68; Pitblado, R.E. (1992), Ridgetown College. Tomato disease-risk windows from temperature and leaf wetness.
  3. Sentinel-2 MSI Level-2A.European Space Agency, Copernicus Data Space Ecosystem. dataspace.copernicus.eu.
  4. NASA HLS.Claverie, M. et al. (2018). Remote Sensing of Environment, 219, 145-161. doi:10.1016/j.rse.2018.09.015.
  5. Sentinel-1 GRD.European Space Agency, Copernicus. C-band radar backscatter.
  6. NDVI.Rouse, J.W. et al. (1974). NASA SP-351.
  7. NDRE.Gitelson, A., Merzlyak, M.N. (1994). Journal of Plant Physiology, 143(3), 286-292.
  8. EVI.Huete, A. et al. (2002). Remote Sensing of Environment, 83(1), 195-213. doi.
  9. NDWI (canopy water, NIR and SWIR).Gao, B.-C. (1996). Remote Sensing of Environment, 58(3), 257-266. doi.
  10. Citrus crop coefficients.FAO-56 Tables 12 and 17 for citrus by canopy cover, with the drip adjustment of FAO-56 Eq. 75; regional agronomist review to follow before Decide opens.
6 · Results

Results are published with their protocol.

The first season's outcome study will be published here, with its design and its data. We publish counts and measured outcomes only, never projections.

  • Design

    Sponsored Decide fields compared with Watch-only fields in the same region, with measures registered before the season.

  • Measures

    Reply rate, action rate, decision windows closed with a logged action, irrigation events and farmer-reported outcomes.

  • Review

    Written up with the reviewing agronomist and the institution hosting the reference plots, with data available to them.

  • Publication

    Published here first, with the limitations register updated alongside.

7 · Replay

Check any record yourself.

  1. Request an exportfor a field or a group, through your dashboard or by email. Each derived record carries its raw-input identifiers, processing version and record id.
  2. Fetch the raw inputsfrom the public sources above (Sentinel-2, HLS and Sentinel-1 scenes by id; weather by provider and time).
  3. Recomputethe indices and the FAO-56 balance with the cited formulas, then the record hash over the ordered records.
  4. Compare.A matching hash shows the record was not altered after it was written. Report any mismatch to us and we investigate it openly.