Service
Geospatial AI and machine learning
Models that read imagery at a scale no team could manage by hand, and predictive models that tell you how far to trust them. The oil palm project counted 718,751 palms this way.
What you get
- Object detection and counting: trees, buildings or other features in drone, aerial or satellite imagery.
- Classification of land cover, crops or other classes you care about.
- Predictive maps, such as flood susceptibility, built from many spatial factors.
- Explanations of each prediction with SHAP, so you can see what drives a result.
- An honest error rate, from tests on areas the model did not train on rather than on random samples.
Who it's for
Plantation and farm operators, lenders and insurers, planners and researchers: anyone with more imagery or more ground than people can review by hand.
How it works
- FrameThe question and the decision it supports
- LabelA sample annotated by hand, or existing data
- TrainCompare models on the same test set
- ValidateTest on held-out areas
- DeployRun at scale and hand over the code
Questions clients ask
How much labelled data do you need?
Often less than people expect. The oil palm detector learned from 4,041 hand-labelled crowns on 435 image tiles, and improved in rounds as new failure patterns showed up.
Why test on held-out areas?
Neighbouring pixels look alike, so random splits flatter a model. GeoAID scored about 0.798 on a random split and about 0.642 under spatial cross-validation. The second number is the one to plan with.
Do I get the model?
Yes. You get the trained model, the code and documentation, so your team can run it again.
Other services
Drone mapping and photogrammetry
Survey flights planned and flown in-house, then turned into orthomosaics, surface and terrain models, and counts.
How it worksSatellite remote sensing and earth observation
Land cover, vegetation, water and change over time, read from optical and radar satellite imagery, anywhere.
How it worksEUDR geolocation and deforestation checks
Plot points and polygons checked and screened for deforestation since 31 December 2020, ready for your EU due diligence.
How it worksLocation assessment and site selection
Where to build, open, expand or invest: candidate sites scored on access, population, competition, risk and cost.
How it worksGIS mapping and spatial analysis
Site selection, access and proximity analysis, map production and clean, well-structured spatial data.
How it worksFlood, climate and environmental risk
Where flooding is more likely, which people, roads and facilities sit in the way, and how land cover is changing.
How it worksWeb GIS development and dashboards
Interactive web maps, Streamlit apps and Power BI dashboards that work on any device.
How it worksSpatial databases and data engineering
PostGIS databases and repeatable Python and Earth Engine pipelines that keep spatial data current.
How it worksTraining and capacity building
Hands-on Python, GIS and remote sensing training for teams, from small workshops to seminars that have reached 1,500 people.
How it works