How we use AI
AI reads the images. Our team decides what they mean.
A satellite looks at every piece of land every few days. No team can read that by hand. Our models do and they only bring you what matters. Below: how a model gets made, what goes in and comes out, where the value is and how your data is handled.
How a model gets made
From raw imagery to a flag someone can act on.
Every model follows the same five steps. The last one feeds the first: what the field team finds goes back into training.
- 01 · Collect
Imagery and records
Satellite archives back to 2017, drone flights per block and the plot, farmer and certificate records you already hold.
- 02 · Prepare
Clean and label
Clouds masked, drone photos stitched into one orthomosaic, everything on one grid. Field plots become labelled training examples.
- 03 · Train
Fit and validate
A model per task (count, compare, project), trained on your land and validated on plots it has never seen. We keep the error bars.
- 04 · Run
Score every pass
Each new satellite pass or drone flight is scored within hours. Only blocks that changed become flags.
- 05 · Review
People check, models learn
The field team confirms or rejects each flag on the phone. Those answers retrain the model next season.
What goes in
Open data plus your own
- Sentinel-2 and Planet imagery, rainfall since 1981, climate projections to the 2040s.
- Elevation, soil and land-cover maps.
- Your drone flights, plot boundaries, field records and photos.
- Nothing you did not agree to share.
What comes out
Numbers with a trail
- Per block: counts, change, risk, with a confidence score and the images behind it.
- Maps and layers in open formats (GeoTIFF, GeoJSON, CSV) you can open in any GIS.
- A dashboard, a weekly brief, alerts to the field team.
- An evidence pack a certifier, lender or verifier can audit line by line.
Three kinds of models
Count. Compare. Project.
Count
Trees, crops, buildings, water: counted from drone and satellite imagery, each with a confidence score.
Compare
What changed since the last pass: clearing, canopy loss, flooding, fire. Flagged with date, area and the before-and-after.
Project
Where crops, trees, fire and floods will be in the 2040s climate. Weights written down, so it can be rerun.
Count
Every tree, from the air.
A drone flies the block at about 80 metres. The model finds each crown, measures it and flags the trees that are missing or stunted. Our field team checks a sample of plots and that sample sets the error bar on the number.
- Counts, survival and canopy per block.
- A confidence score on every tree.
- Validated against field plots, error bars published.
On the ground
The field app closes the loop.
A flag from the model becomes a task on the phone: walk the route, take two photos, confirm what you see, save offline. When the phone finds signal, the dashboard and the evidence pack update themselves.
- Works for weeks without internet.
- GPS, heading and the before-and-after image on every photo.
- English and Kiswahili.
Compare
What changed since the last pass.
Each new satellite image is compared with the last one, block by block. Clearing, canopy loss, standing water and burn scars come out as a flag with a date, an area and the before-and-after.
- A flag per block, not a wall of pixels.
- Fire history back to 2017.
- Every flag becomes a field visit.
Project
Run the climate forward to 2040.
Rainfall, temperature, altitude, soil and people, under today's climate and the projections for the 2040s. The result is a map of where a crop, a tree species or a flood will be, with every weight written down.
- Suitability for crops and tree species.
- Flood, fire and erosion risk.
- Rerunnable by your own team.
Where the value is
Faster, fuller and provable.
Hours, not months
A new satellite pass is scored the day it arrives. You know which block changed this week, not at the end of the season.
Every block, not a sample
Field teams can visit a few plots. The model reads all of them, so the visits go where they matter.
Evidence, not opinion
Each number carries its confidence, its images and its field check. That is what a certifier or lender can accept.
Data security and ownership
Your land, your data.
Six commitments we put in every contract.
You own it
Boundaries, field records, drone imagery and the results stay your property. Export everything in open formats at any time.
Access is per person
Dashboards and field apps sit behind individual logins. You decide who sees which block.
Encrypted, backed up
Encrypted in transit and at rest, hosted on infrastructure you choose or on ours, with daily backups.
Farmers consent first
Personal data is collected with a consent script in the field, kept to what the task needs and never sold.
No cross-client training
Your imagery does not improve a competitor's model. Not without written agreement.
Delete on request
When a project ends, you choose what we keep, hand over or erase.
Honest about limits
What AI cannot do, and what we do about it.
Clouds happen
In the wet season some passes are unusable. We combine optical and radar sources and say when a block was last seen clearly.
A pixel is not a cause
The model sees canopy loss, not why. Fire, clearing or disease is a question for the person on the ground.
Every number has an error bar
We publish it. A tree count of 1,214 ± 3 % is useful. A count without the ± is a guess.
An assistant that speaks your language
Chat and voice in English and Kiswahili, grounded in your own maps and the weather forecast. "Which block today?" gets an answer, not a spreadsheet.
Never the last word
Every flag goes to the field. Every model is checked against ground plots and we publish the error bars. That is what turns a map into evidence.
Want to see it on your own land?
Send us the blocks or the question. We will show you what the models see.