From farm data to the hours on your phone
Four kinds of data go into our physics-informed neural network. One WhatsApp message comes out, with the hours a day for every block.
Illustration of how the model works. The hours shown are an example.
What each input tells the network
Satellites
PlanetScope images at 3 m, almost every day, show how big your trees are and when a block weakens. NASA Landsat thermal images show which blocks run hot. Bigger trees in hotter blocks drink more.
Weather
Our FAO engine turns the 7-day forecast into how much water the trees will lose each day, heat waves included.
Water meter
A small reader on your meter counts what each block really got. The network compares it with its forecast and learns from the gap.
Soil and salt
Soil maps give a first guess of your soil, and a test of your well water tells it the salt. Probes on reference farms read soil water at 20, 40 and 60 cm.
Satellites alone can't set the hours. They can't see the soil or the salt under the trees. That is why the network combines all four.
Why physics inside the network
A normal AI model needs huge amounts of data, and it can still guess things that are physically impossible. Ours learns from two teachers at once: the measured data, and the equations of how water moves through soil.
So its forecasts have to agree with how water really behaves, and it can learn a soil from far fewer probes. This is a state-of-the-art approach from soil physics and machine learning research, and the 2027 pilot tests it on real farms.
Raissi, Perdikaris & Karniadakis, Journal of Computational Physics. Bandai & Ghezzehei, Water Resources Research. FAO-56.
A hub-and-spoke model
Probes sit on a few reference farms, the hubs. They teach the network how each soil in the region holds water and salt. Every other farm is a spoke: it only needs the meter reader, and the network brings what it learned on the hubs to its soil.
Your week with Arvessa
Monday: the plan arrives
One WhatsApp message with the hours a day for each block.
You run the pump
You decide. The plan only advises; it never touches your pump.
The meter counts
The reader records what each block really got.
The network learns
It compares its forecast with what happened, so next Monday’s plan fits your blocks a little better.
Then it starts again. Every week the plan learns a little more about your blocks.
Look inside: the soil twin
This is what the network sees under one block: water at 20, 40 and 60 cm, hour by hour, with a 7-day forecast. Change the soil, the weather and the trees.
Example simulation of a block on a reference farm with a probe.
How we test it in 2027
First the network only watches: you water as usual, and it forecasts. Once its forecasts hold, it advises on a few blocks. On each pilot farm, four blocks compare four ways to set the hours.
Join the 2027 pilot
Do you run a drip-irrigated citrus farm or a packhouse in Souss-Massa? We're looking for three farms for a free pilot. Researchers and funders are welcome to write too.
Fill in the application Or write to pilot@arvessa.org