Projects Computer Vision Live product
Identifying cultivable land from satellite imagery through image processing rather than manual survey.
Free to open · no account · nothing to upload
Deciding which land is worth cultivating is normally a survey exercise: slow, expensive, and stale by the time it is finished.
Satellite imagery is processed to classify terrain suitability, with results surfaced through an R Shiny interface so the output is explorable rather than a static report.
Making the result interrogable. A classification nobody can question is a classification nobody will act on.
(01)
The paper fused four things that are normally kept apart: a sharpened satellite image, a colour mask that finds the green, an open weather and soil archive, and a language model asked what would grow in those conditions. Its crop suggestions were then checked against a government crop calendar.
The product runs that same chain on real Sentinel-2 imagery, in the browser, while you watch each stage happen. Nothing is pre-rendered: the sharpening and the mask are computed on the scene in front of you.
(02)
The paper fixed its green bounds by hand — hue 40 to 80, saturation and value from 50 up. Those numbers are right for a wheat plain in March and wrong for a date oasis, where the palms are darker and duller than anything a fixed floor will catch.
So the product has no sliders at all. It picks a sharpening weight, seeds on the scene’s own greenest pixels, reads the hue, saturation and value bounds off that seed, then strikes the mask — printing each decision as it makes it, so a disagreement is with a number rather than a black box.
On the pivots at Wadi Al-Dawasir it settles on hue 31 to 55, saturation from 100 and value from 73, and measures 480 acres of green — a fifth of the scene. The paper’s fixed bounds would have taken half the desert with it.
(03)
The paper scaled its areas against a printed map. Here every scene carries its own metres per pixel from the satellite, so the struck pixels convert straight into acres and hectares that a farm manager can act on rather than a percentage of a picture.
Ten scenes come loaded — five in the Kingdom, the rest over the districts the paper tested — and each is about three kilometres across, which is a neighbourhood of farms rather than a single field. Or type any town on Earth and it fetches the clearest pass of the last six months and runs the whole chain on it, live.
(04)
The paper’s own evaluation is on the page rather than in a footnote. Across twenty Indian districts the model named the crop calendar’s dominant crop fifteen times — 80% counting only its first-ranked crop — and the five it missed are shown as plainly as the fifteen it got.
The per-crop scores say where to trust it: chickpea, cotton, groundnut and mustard land perfectly, wheat and rice mostly, maize and soybean not at all on a single district each. That is the honest shape of a small evaluation, and it is why the rulebook answers first and the model second.
(05)
The paper’s decision rules answer instantly and show every branch they took, on the day’s real temperature, humidity, rain, sunshine, evapotranspiration and soil moisture at that exact place. A language model answers separately, with seasons and cautions, and only when asked.
Both are then held against what is actually grown there. A desert pivot at twenty-six degrees with no rain in a month is not a mystery to solve; it is irrigation, and an adviser that cannot say so is not worth running.
It started as a paper
Scientific Reports, 14 (1), 16241 2024
Riyadh, Saudi Arabia
The demo is free, and everything it needs is already inside it. When you want it on your own ground — your data, your people, your rules — that is a conversation and a number in writing, not a licence you click through.
Examinations
Photograph a handwritten script. It reads English, Arabic or equations, scores it against your keywords, explains every mark, and reports on the whole class.
Logistics
Eight drones, four pairs, no dispatcher. Order to a roof over a real Riyadh neighbourhood and watch the swarm bid for it, fly it, map the streets and cover for a drone that falls.