Shamba Pulse

Turning JKUAT Conduit weather data into late-blight alerts and safe spray windows, sent by SMS in English and Kiswahili to farmers in Juja and Kiambu.

Comment

Inspiration

Kenya's Meteorological Service expects above-average October–December 2026 short rains across most of the country, including Kiambu, with prolonged wet spells, driven by a strong El Niño. Long, humid, wet spells are exactly when late blight spreads through tomato and potato farms.

Smallholder farmers around Juja usually face two bad options: spray fungicide on a fixed schedule (costly, and often wasted when rain washes it off) or spray too late and lose the crop. Generic forecasts like "rain expected" don't answer the questions that matter: Is my crop at risk right now? When should I spray?

The JKUAT Conduit station measures exactly the variables that drive blight: humidity, temperature, rain and wind, every 15 minutes. We wanted to turn those readings into a decision a farmer can act on today.

What it does

Shamba Pulse turns Conduit station data into three decisions:

  1. Blight risk (HIGH / MODERATE / LOW / UNKNOWN) using the Hutton criteria, an established late-blight risk rule. A day is a Hutton day when

$$T_{\min} \ge 10\,^{\circ}\mathrm{C} \quad \text{and} \quad \text{hours with } RH \ge 90\% \;\ge\; 6$$

and two consecutive Hutton days trigger HIGH risk. Every verdict comes with plain-language reasons, e.g. "Humidity stayed above 90% for 11 hours, 2 days running."

  1. Spray windows: using the Open-Meteo forecast, it finds hours that are in daylight, have moderate wind (not calm, not gusty), have dry leaves (RH < 90%) and are followed by at least 6 rain-free hours, so the spray isn't washed off. It also explains why every other hour was ruled out.
  2. An SMS alert under 160 characters, in English and Kiswahili. The Kiswahili uses Swahili time (09:00 is saa 3 asubuhi); otherwise a farmer could read "09:00" as 3 p.m.

Two users, two products: farmers get the SMS; extension officers, agrovets and cooperatives use the dashboard to advise many farmers at once. A sending policy texts only when risk rises or Hutton fires, to avoid alert fatigue.

Does it work? A real-season backtest

We downloaded the station's full history (476 days, 46,183 readings) and replayed the real October–December 2025 short rains with no look-ahead (enforced by tests):

Result (1 Oct – 31 Dec 2025) Value
Days replayed 92
HIGH-risk days 10 (6 episodes; longest Hutton run: 5 days)
Days inside a Hutton run 21
SMS-eligible days → texts actually sent 46 → 15 (67% fewer)

The season's main HIGH period (29 Oct – 1 Nov 2025) coincided with the Kenya Meteorological Department's heavy-rainfall advisory for 23–30 October 2025, which listed Kiambu. This corroborates the weather, not disease on the ground; showing that alerts reduce blight needs a field trial.

Illustrative scenario (not a measured result): weekly scheduled spraying means about 13 sprays over those 92 days; spraying around the 6 high-risk episodes means about 6. Whether fewer, better-timed sprays protect the crop as well is exactly what a pilot must test.

How we built it

  • Data: the official Conduit API (form-encoded POST with a date range), fetched in cached, polite chunks. Timestamps are converted from UTC to Kenya time (a naive read would shift every night-time humidity window by 3 hours).
  • Engine: Python and pandas. It is rule-based and explainable, with the Hutton criteria, a humid-hours fallback, spray-window scanning and a send policy. All thresholds live in config.py with a comment explaining each.
  • Forecast: the Open-Meteo hourly 7-day forecast, mapped to the same columns as station data so the engine doesn't care about the source.
  • Cross-check: ERA5 reanalysis rainfall via Open-Meteo's archive.
  • App: Streamlit, Plotly and custom HTML/CSS components (risk ring, 7-day strip, daylight spray timeline, phone mockup, station map), deployed on Streamlit Community Cloud. If the station is unreachable, it falls back to committed real history and says so clearly.
  • SMS: an Africa's Talking integration. The demo runs in dry-run mode; no real SMS is sent.
  • Quality: 378 automated tests, including Hutton boundary cases, no-look-ahead checks, 160-character limits for every message, and Swahili time for all 24 hours.

Challenges we ran into

  • Our tests passed while a core feature was broken. In the backtest the HIGH pathway never fired: at alert time "today" is incomplete, so the consecutive-day counter always broke on it. Every test used complete days. We fixed the logic and added tests for real-world partial days.
  • The obvious rain field was wrong. The per-interval rain field (rg1) recorded 0.20 mm where the station's daily running total showed 6.20 mm, so it missed ~97% of the rain. We rebuilt rainfall from the running total, handling the daily reset.
  • Rain gauge 2 is faulty. Its "daily" total resets about 21 times a day and adds up to thousands of mm in a ~300 mm season. We excluded it and documented the evidence for the station maintainers.
  • Other sensor issues: the UV sensor reads constant, and the gust-direction field duplicates gust speed.
  • The API's end date is exclusive, and the station publishes with a lag. We fixed the query and show data freshness honestly instead of calling old data "live".
  • Alert fatigue: a naive version would have texted on most days of a wet season, so we designed a send policy.

Accomplishments that we're proud of

  • A real-season backtest on actual Conduit data, not a synthetic demo.
  • Finding and documenting real sensor faults. Rainfall rebuilt this way totalled 301.8 mm for the season, within 12% of ERA5 reanalysis (268.6 mm).
  • A system that says UNKNOWN instead of a falsely reassuring LOW when data is thin.
  • Kiswahili messages that respect Swahili time, not a literal translation.

What we learned

  • Inspect the data before trusting it: the most obvious field was the wrong one.
  • Explainability builds trust: a farmer needs to know why to spray, not just a number.
  • Tests must include messy real-world cases (partial days, gaps, lags), not just clean ones.
  • Honest limitations make a stronger case than overclaiming.

Limitations

  • A single station: it represents the JKUAT/Juja area, not every farm in Kiambu.
  • The Hutton criteria were developed in the UK; local calibration against observed blight outbreaks in Kiambu is our #1 next step.
  • The Kiswahili is pending native-speaker review.
  • No soil-moisture sensor. ERA5 is a reanalysis (a model that assimilates satellite and ground data), used only as a cross-check.

What's next for Shamba Pulse

  1. A pilot with a county extension office and farmer groups, measuring sprays, cost and disease incidence.
  2. Satellite rainfall (CHIRPS/GPM) in the live pipeline, plus more stations for Kiambu-wide coverage.
  3. An ML model trained on observed outbreaks, using the engine's existing slot for it.
  4. Live SMS/USSD via Africa's Talking, with native Kiswahili review. It could be funded by county extension services, cooperatives or input suppliers.

Data credits: JKUAT Conduit station (JHUB Africa / JKUAT), Open-Meteo, ERA5 (Copernicus/ECMWF), Kenya Meteorological Department.

Built With

Share this project:

Updates

Submission history