Crop-Guardian

মাঠের সৎ ডাক্তার — a Bengali, mobile-first UI for advising potato farmers on late blight, and paddy farmers on rice blast: daily risk from the weather, leaf photo diagnosis that admits uncertainty (potato, and paddy as an experiment, see Paddy diagnosis), an area risk map, and a season ledger. Every number on screen comes from a real source: the weather forecast, the farmer's own records on the phone, or cases an agriculture officer has confirmed. There is no sample or demo data.

Built with Next.js (App Router), TypeScript and Tailwind CSS v4.

Why potato late blight

This is a deliberate focus: one crop and one disease done honestly, rather than many done thinly.

• Potato is a main winter crop of northern Bangladesh, and late blight (নাবি ধসা) is its most destructive disease. A 16-district survey found it cuts yields by up to 30% (Potato Research, 2022).
• It is driven by the weather, so it can be forecast. Fog and cold, wet spells in December–January spread it within days. A warning the afternoon before lets a farmer spray in time, and a quiet forecast saves the calendar sprays most farmers pay for anyway (the season ledger).
• It has look-alikes (early blight, other leaf spots), which is where a photo check that admits doubt helps most.

Pests are not covered. The challenge asks for pests and diseases. We chose to cover one disease end to end (forecast, SMS alert, photo check, officer review, ledger) first. The pieces are built to grow:

• More crops and diseases. Crops are a registry (src/lib/crops.ts), and paddy blast already runs on it as a draft rule (experimental).
• More photo classes and crops. Each crop's photo model, classes and thresholds come from its own exported config (src/lib/diagnosis/crops.ts). Paddy diagnosis runs end to end on a model trained by notebooks/paddy_training.ipynb, labelled experimental; pests such as potato tuber moth or aphids would be new classes the same way.
• More places. The area list already covers the 17 potato districts of Rangpur and Rajshahi divisions and Munshiganj, down to unions.

See Limitations and next steps for what is not yet validated.

Run

npm install
npm run dev      # http://localhost:3000
npm run build && npm start
npm test         # unit tests (node:test)
npm run typecheck
npm run lint

On a locked-down network npm install can fail because onnxruntime-node (used only by the parity test) downloads its GPU files from NuGet. Set ONNXRUNTIME_NODE_INSTALL_CUDA=skip first.

Screens

Route	Screen
/	Home: today's live risk at the farmer's union for each field in season (a chip per field), 5-day forecast, today's decision; a 4-card intro on first open
/camera	Leaf photos (top, then underside) and on-device diagnosis
/result/uncertain	"Not sure": the candidates the photos left
/questions	Two questions that narrow the candidates
/result	Final diagnosis, including "healthy"
/unknown	Can't identify: send to agriculture officer
/map	The unions of the farmer's upazila (Shibganj, Bogura, until one is chosen): today's weather risk per crop (আলু · ধসা / ধান · ব্লাস্ট) and officer-confirmed cases
/ledger	A field's season (?field=potato|boro|aman) from the real weather and the sprays the farmer logged
/planting	"আমার জমি": the farmer's place (district → upazila → union, ?d=/?u= are the steps), and their fields: crop (আলু, or ধান with বোরো/আমন) and planting date (?season= opens one)
/guide	How to use the app, in plain Bengali, each part read aloud
/potato	Potato growing advice for Bangladesh: planting, varieties, late blight, wilt, harvest, storage, stored-tuber diseases, with sources
/ask	«কৃষি সহায়ক»: questions answered in Bengali from the app's own advice, with sources
/docs	How it works, in English, for judges, partners and agronomists: the evidence and how AI is used
/sign-in	Mobile number, then the SMS code (?next= is where to go after)
/account	The signed-in number and role, SMS risk alerts on/off, sign out (from the person button on Home)
/officer	Agriculture officers only: where farmers' cases will be listed

Late-blight risk

Home's risk is computed from the weather, with no ML. src/lib/weather.ts fetches a 5-day hourly forecast (temperature, relative humidity) at the centroid of the farmer's union (or Shibganj upazila, Bogura, before one is chosen) from Open-Meteo (no API key, cached for 1 hour, 5 s timeout). src/lib/risk.ts applies the Hutton criteria to it:

• A day qualifies when its min temperature is ≥ 10°C and RH ≥ 90% for ≥ 6 hours.
• বিপদ (danger): the day and the day before or after it both qualify (a Hutton period).
• সতর্ক (watch): the day qualifies alone, or is ≥ 10°C with 4–5 humid hours.
• নিরাপদ (safe): otherwise.
• Today's overall risk is today's level, raised to watch if a danger day falls within the next 2 days.

The same result is served as JSON at GET /api/risk (optional lat, lon, by default the farmer's union; optional crop=potato|paddy, potato by default). If the forecast can't be fetched, Home says so ("আবহাওয়ার তথ্য আনা যায়নি") and shows no risk at all, never an old or made-up one.

This is the beginning of the README. Read the rest at the original. Read the full README ↗