Satellite cropland NDVI by county and district in Kenya, Tanzania, Uganda and Rwanda

Cropland vegetation indices per county, region or district in Kenya, Tanzania, Uganda and Rwanda: a 10 m weekly product and a 300 m dekadal product with 8 years of history, the input to the demand signal.

Catalogue entry: Satellite demand indices. Cropland vegetation indices per administrative unit against that unit's own seasonal normal — the input to the fertilizer demand signal. Kenya plus TZ/UG/RW.

Questions this data answers

  • Where will fertilizer demand start in the next 2 weeks: which counties and districts are greening ahead of their own normal?
  • Which of Uganda's 135 districts are furthest below their seasonal normal right now, for pricing risk or planning relief?
  • Where in Tanzania is the season on time and where is it late, region by region, before yield reports exist?
  • Which counties should a sales team visit this week, ranked by how far cropland is above normal?
  • Is a county's greenness a rainfall story or a planting story, from the 300 m and 10 m series together?
  • How did Nakuru's cropland develop through the 2026 long rains, week by week, against the 8-year history?

What the satellite crop condition data holds

Cropland vegetation per administrative unit across 4 countries: 47 Kenyan counties, 31 Tanzanian regions, 135 Ugandan districts and 30 Rwandan districts. 2 products sit under the source tag. A 10 m product supplies 4 indices per unit and week: NDVI for greenness, NDRE for chlorophyll, SAVI for greenness with the soil signal damped down, NDMI for canopy moisture. A 300 m product supplies NDVI from 2018 onward and, from July 2026, a soil water index alongside it. The 10 m values are averaged over cropland only, so a value reflects what the fields themselves are doing, not the towns, forests and lakes sitting around them.

Against its own normal

The demand signal compares each unit only with itself: this week’s cropland vegetation against what the same weeks looked like across the 8 years before. A county that simply runs greener than its neighbours only counts as above normal once it beats its own record. The 10 m series is the live reading, the 300 m series is the climatology it’s read against; the platform calibrates the 2 per unit and, wherever they disagree past a set tolerance, serves the lagged reading and says so directly in the response. Thresholds of ±0.02 and ±0.05 sort the anomaly from well below to well above normal, and demand_outlook reads rising, steady or softening off the anomaly and its trend together. This dataset carries only the raw values; the anomaly, the outlook and the ranking all come from the demand signal and demand board analyses built on top of it.

Nakuru in 2026

The chart tracks the county’s 10 m NDVI across the year: 0.47 in early January, a low of 0.35 in early March just before the long rains, 0.68 at the start of June with the canopy fully closed, and back down to 0.46 by September. The 300 m product underneath follows the same season about 40 days later.

Across counties at the snapshot

Nandi read 0.66, Kericho 0.62, Kakamega 0.58, Marsabit 0.09. Against each county’s own normal, though, the picture shifts: Nandi sits at +0.03, Kakamega is exactly on normal, Marsabit at −0.18. The raw value says how green a county is; the anomaly says whether that’s unusual for it. The demand board ranks all 243 units this way every week, blended with how intensively each one farms, so a breadbasket county sitting merely on normal still outranks a pastoral district enjoying a rare green spell.

Sample data

Nakuru county, 2026: the 10 m series against the 300 m series

26 weekly composites. The long rains take cropland from 0.35 in early March to 0.68 at the start of June; the dry season brings it back to 0.46 by September.
The same season in the 300 m product, monthly means. It runs about 40 days behind and reads a little higher because its pixels are not cropland-masked; the per-county offset between the 2 is what the calibration measures.

Sample rows, two Kenyan counties, latest composite

StampCountyIndexValue
2026-10-18BaringoNDMI-0.0742
2026-10-18BaringoNDRE0.2061
2026-10-18BaringoNDVI0.3422
2026-10-18BaringoSAVI0.2033
2026-10-18BungomaNDMI0.0865
2026-10-18BungomaNDRE0.3789
2026-10-18BungomaNDVI0.5787
2026-10-18BungomaSAVI0.379

4 indices for each of the 243 units every week.

Dataset idmarket.demand_indices
Cadence10-daily / weekly
EndpointGET /market/demand-indices
Key scopemarket:read
Declared fields1
History2018-01 to today
CoverageKenya 47 counties, Tanzania 31 regions, Uganda 135 districts, Rwanda 30 districts
StatusServed over the API

Global reference data — the same rows for every caller.

Source: Contains modified Copernicus Sentinel data; administrative boundaries from UN OCHA COD-AB / geoBoundaries (CC BY-IGO); cropland mask from ESA WorldCover (CC BY 4.0). Rain outlook: © 2024-6 Google LLC, whose machine learning models were used to create the experimental data made available under the following licence terms https://storage.googleapis.com/weathernext-public/terms-of-use.pdf. This data is intended for experimental modelling only and is not intended, validated, or approved for real world use..

Fields

FieldUnitNote
value—index value; see the index tag for which

Filters

countrycountyindexlevelsource

The same call, three ways

curl -G https://api.nuasense.com/api/nuasense/v1/market/demand-indices \
  -H "Authorization: Bearer nua_live_…" \
  -d start=-30d -d country=ke -d county=nakuru -d index=ndvi

Through the MCP connector the same read is the platform_dataset tool with dataset=market.demand_indices. Through a scheduled export it is one job with the same filters, delivered as a file. Every dataset takes the same arguments: start and end, every with agg, fields, limit up to 5,000, order, and the tags it declares.