Optical imagery is often blocked by cloud during kharif. Sentinel-1 SAR gives wetness and crop-structure signals when Sentinel-2 is unavailable.
Research-backed product choices
Designed from field evidence, not app assumptions.
KisanVaani uses satellite fusion, crop-stage water logic, agromet adoption evidence, Indic voice models and human expert validation because those are the hard parts in rural deployment.
Design principles
What the research changes in the product
Dry-spell risk is not generic. FAO-56 style ETc = ETo x Kc logic lets alerts change by sowing, vegetative and flowering stage.
Field-photo diagnosis is noisy. Low confidence, flowering-stage risk and clustered reports must go to RSK experts with context attached.
Evidence stack
Five technical choices judges can defend
Fuse radar wetness with optical NDVI/NDWI. If clouds block optical images, SAR still keeps the dry-spell sentinel alive.
Advisories are more useful when they are crop-specific, timely and trusted. The app localizes advice to village, crop stage and RSK validation.
Bhashini and AI4Bharat capabilities are used for ASR, translation and TTS, while approved agricultural advice remains template controlled.
The crop engine gives safe, balanced and avoid options so a small farmer is not pushed into one fragile recommendation.
The alert fires only when forecast deficit agrees with rainfall anomaly, satellite stress or sensor evidence, reducing false alarms.
RSK corrections become district labels, making the product improve from local agronomy instead of generic model guesses.
Research to product
How evidence becomes a clickable feature
| Remote sensing | Multimodal crop-monitoring literature supports combining radar and optical observations. Product feature: Sentinel-style wetness board and crop scoring input. |
|---|---|
| Water requirement | FAO-56 provides the operational crop evapotranspiration basis. Product feature: dry-spell advice by crop stage rather than generic rain alerts. |
| Advisory adoption | India AAS studies show crop advisories can affect outcomes when farmers can access and trust them. Product feature: village-specific voice advice plus RSK approval. |
| Language access | Indic language model work enables ASR, translation and TTS across Indian languages. Product feature: Telugu/Hindi voice flow with dialect phrase handling. |
| AI safety | Field diagnosis uncertainty is handled by confidence gates. Product feature: instant advice only for high-confidence routine cases; uncertain cases become RSK tickets. |
Honest limits
What is feasible now vs. production
One district, three villages, deterministic agronomy engine, synthetic satellite features, RSK demo queue and CI-deployed static app.
Load real Soil Health Card, village boundaries, IMD/CHIRPS, Sentinel features, RSK roster, consent and IVR/SMS provider credentials server-side.
No pesticide or fertilizer dose is sent from a low-confidence model. The RSK expert must approve high-risk diagnosis and broadcast cases.