Skip to main content

Case Study · GeoAI Telangana

A production deployment serving the entire Indian state of Telangana.

Goal​

Predict ten soil parameters from Sentinel-2 satellite data and deliver soil advisories to government, agricultural officers, and farmers — through one technical foundation.

Stakeholders​

  • Government of Telangana
  • Mandal Agricultural Officers (MAOs)
  • PJTSAU (agricultural university)
  • GIZ (international cooperation)
  • Indian AgriTech start-ups (BharatRohan, Carbonmint, Transity)
  • Farmers (notified via WhatsApp)

Architecture​

Three audiences, one digital farm twin:

  • State. District NDVI overview for programme oversight and policy targeting.
  • Mandal. Ten soil and crop parameters per mandal for MAO advisory.
  • Field. Time-series at the field level, available through the farmer's existing FMIS or the platform's farmer-facing front-end.

APIs used​

  • Spatio-Temporal API — Sentinel-2 ingest as Zarr DataCubes, model outputs as additional Zarr layers in the same collection, STAC items per time slice.
    • Uploads via POST /api/v2/collections/{id}/zarr/init → POST /layers → PUT /layers/{name}/data/binary.
    • Frontend reads layers via TileJSON: GET /api/v2/collections/{id}/tiles/WebMercatorQuad/tilejson.json?layers=NDVI.
    • State / mandal / field views all use the aggregations endpoints (/aggregate/time-series, /aggregate/summary) with per-region bboxes.
  • Activity API — advisory delivery, attached to fields and regions as typed Records.
  • Farm API — region hierarchy: state → district → mandal → field, modelled as nested Regions.

Data flow​

  1. New Sentinel-2 scene becomes available.
  2. The GeoAI service is commissioned for all of Telangana.
  3. AI outputs are written back as Zarr layers in the Soil monitoring Telangana collection. The platform automatically indexes them as STAC items and exposes them via TileJSON + WebMercator tiles.
  4. Subscribed farmers are notified via WhatsApp.
  5. Soil report and decision support are available in the dashboard — driven by the MapLibre style the API generates for the collection.

Outcome​

  • Ten trained models in production
  • Four currently cross R² > 0.55
  • Trained on 43,000+ cleaned soil samples plus 600 GIZ AgriFabriX samples
  • Rolled out state-wide across all 33 districts in March 2026

See the GeoAI workshop report for detailed performance metrics.