An offline irrigation controller for smallholder farms. One box at the water manifold keeps a daily water account for each soil zone of the farm and runs the pump itself: it opens the right zone valve when a zone's account runs low, stops exactly when the refill is committed, and never needs internet, a server, or a person driving out to turn the water off.
Saqi v1 argued, in simulation, that a farm can share scarce water with a mesh of per-tree devices and no central computer, and that this survives network failures that break a cloud controller. That research stands.
Then I interviewed the farmer I know best: my father. Two things came out of it.
First, per-tree sensing is redundant on a real smallholder farm. Trees in the same soil, on the same slope, watered on the same line, sit at practically the same moisture. What actually differs is the handful of soil zones the farm splits into. Commercial precision agriculture reaches the same conclusion; it calls them management zones.
Second, information is not the bottleneck. Products already exist here (in Tunisia, Seabex tells you from satellite data when and how much to irrigate) and most farmers around us irrigate from their own well, so saving water is not what they care about. What costs them is the pump's diesel and the trips: somebody still has to drive out, open the valve, wait, and close it. Every existing tool stops at the phone notification. Nobody automates the stopping.
So v2 keeps v1's allocation brain and re-aims it: zones instead of trees, one local controller instead of a mesh, and the objective translated from litres into what a well irrigator feels, pump hours and trips. The mesh idea from v1 does not die; it moves up a level, to several farms sharing one well or canal with no central authority, which is future work.
- The farm is split once into zones by soil type, using satellite imagery and the farmer's own knowledge. That map is static and lives on the device. No live satellite in the loop: satellite soil moisture reads the top few centimetres on a multi-day revisit, while olive roots drink from a metre down.
- Each zone has a water account: a running depletion estimate in mm, drained daily by evapotranspiration (computed FAO-56 style from a cheap local weather reading), topped up by rain and irrigation, and corrected by one soil probe per zone.
- When a zone's account crosses its refill line, the controller sequences the pump: driest zone first, within the pump's daily capacity, stopping each refill on committed water (what has been applied, not what a lagging sensor reads) so nothing drains past the roots.
The zoning and the water accounting are both in this repo, with one simplification each. The
zone map is a given input (the ZONES table at the top of saqi2.py, five zones with their
soil capacities); on a real farm that map is drawn once, from satellite imagery and the
farmer's knowledge, then lives on the device as exactly such a table. The ET accounting runs
every simulated day (crop coefficient times reference ET per zone), but the simulation draws
ET0 from typical dry-season values; the device computes it from its own weather reading.
In short: the repo proves the decision logic, the pilot proves the physics.
A 120-day dry season on a 5-zone, 1.1-hectare farm, averaged over 30 random seasons, comparing the three things a farmer can actually buy: a timer box (blind automation), an advice app (perfect information, human execution), and Saqi v2 (informed automation).
The timer box stresses the fast-drying sandy zones for ~18 days each while over-watering the clay. The advice app eliminates stress, but at the cost of 79 trips to the farm in one season, and manual stopping still drains water past the roots. The controller matches the app's zero stress with zero trips, the fewest pump hours, and nothing wasted.
Knowing when to water is not the hard part. Acting on it is.
controller.py is the reference implementation of the full loop the firmware ports:
water accounts, rain hold, weekly probe anchoring, and the fail-safes. It replays a
season where a probe freezes, a storm hits, the power dies mid-season, and a valve jams:
The box detects the dead probe (and keeps that zone alive on the model alone), closes everything on reboot so a power cut can never leave a valve open, credits the storm instead of watering through it, raises exactly one alarm for the jammed valve and parks the zone until a human fixes it, then recovers. One subtlety worth reading in the code: a frozen probe looks exactly like a broken valve, so the valve check refuses testimony from a probe that has stopped moving.
pip install -r requirements.txt
python3 saqi2.py # the strategy comparison -> saqi2_season.png, saqi2_timer_vs_auto.png
python3 controller.py # the fault season -> saqi2_faults.pngOne ESP32 controller at the manifold, a latching solenoid valve per zone, one capacitive soil
probe per zone, a rain gauge, a small solar panel and battery. Rough parts cost for a 5-zone
farm is under $100. A two-zone pilot on my family's farm in Sidi Bou Zid is the next step:
the design is in docs/DESIGN.md, the wiring, calibration protocol, and acceptance checklist
in docs/BUILD.md.
The soil and weather model is deliberately simple, and the manual-watering baseline encodes assumptions (visit behaviour, 15% overfill) that should be measured, not assumed. The constants that matter (each zone's capacity, drying rate, and how fast applied water settles) have to be calibrated on a real farm; that calibration is the first job of the pilot.
I used AI assistants to help debug the code.
MIT. See LICENSE.