Public data
Uses Sentinel-2, which is free and refreshed every five days.
UnB research project
Satellites and AI to know what each plot produced.
MapaSafra is a research project of the University of Brasília (UnB), developed as an undergraduate thesis. It trains artificial intelligence models on Sentinel-2 satellite time series to identify, inside a property's perimeter, which plots were productive in each month.
You draw your production units on the map and test the model on them, with no field visit and no reliance on self-reported data.
OUR MISSION
A second source of information about the field, public and verifiable
We want to show that public satellite imagery, combined with machine learning, can answer at scale and at low cost a question that today depends on field visits: what was produced, where and when.
Every result is presented with its limits. The model was trained and evaluated on data from Bahia; for other regions, we measure accuracy before claiming anything.
Monthly answer
The result comes per plot and per month, not just per harvest.
Stated limits
Metrics and methodology are documented, including what the model does not do yet.
Academic origin
Born at the university, with a method described so it can be evaluated and challenged.
THE PROBLEM
Knowing what each plot produced is still expensive
In most cases, the outline of the plots does not exist and what was planted in each area depends on what someone declared.
Field visits are expensive, sample-based and repeated with every new question. By the time the information arrives, the harvest is already over.
Satellites already observe all of this every five days. What was missing was turning the image into a useful answer: which plot produced, and in which month.
No outline of the plots
Reliance on declarations
Information that arrives late
Expensive, sample-based field visits
Cost that grows with area
An answer per harvest, not per month
HOW THE MODEL WORKS
From satellite image to productive plot
MapaSafra processes Sentinel-2 time series indexed on H3 hexagons, a grid that splits the terrain into cells of uniform size.
A contrastive learning model compares each cell's trajectory over time and indicates whether production occurred. It was trained on real agricultural fields in Bahia.
Public imagery
Sentinel-2 images cover the whole property at no acquisition cost.
H3 cells
The perimeter becomes a grid of hexagons, and each cell is analyzed separately.
Month-by-month reading
Results come back grouped by plot, month by month.
RESEARCH PROJECT
Academic work in development
MapaSafra is an undergraduate thesis project at the University of Brasília, with academic supervision and a focus on research applied to agriculture.
The work has three fronts: detecting productive plots, classifying the crop produced, and making the models available on this platform.
Detecting productive plots is the most advanced stage. Crop classification and full platform integration come next.
OUR VISION
Field information within reach of those who need to decide
We want anyone with a perimeter drawn on the map to get a monthly reading of what was produced, without someone having to drive out there.
In the future, we imagine expanding coverage to new regions and crops, always with measured validation before each expansion.
Less guesswork. More evidence. One answer per plot, every month.
IN DEVELOPMENT
Built with people who know the field
MapaSafra is evolving and depends on those who work the land to get better.
Every property tested, every question and every mismatch between the map and reality helps measure where the model is right and where it still errs.
That is why we invite producers, technicians, cooperatives and researchers to test the model on areas they already know.
Producers
Tell us whether the map matches what happens in the field.
Technicians and consultants
Point out where the reading helps in day-to-day field work.
Cooperatives and institutions
Bring known areas to validate the model in new regions.
Researchers
Review the method and suggest improvements.
See what your production unit produced
Draw the perimeter on the map and get a monthly reading of the productive plots.
Try the model