Home/Solutions/Detection and mapping from imagery (GeoAI)
Where is every object, across a whole region?
Object detection in satellite imagery finds buildings, tanks, trees or panels across an area too large to check by eye, and turns them into an inventory on a map. We prepare the training data, train deep-learning models on your objects, run them over the imagery and review the results, so the inventory you receive has been checked.
Some questions are simple to ask and slow to answer: how many storage tanks stand in a state, how many trees grow in an orchard, which roofs carry solar panels. A person can answer them for one town; a trained model can answer them for a country, the same way every time.
GeoAI is the name for this kind of work: machine learning applied to geospatial data. Detection is its most visible use, next to land-cover classification and change detection, and it works on satellite, aerial and drone imagery alike.
Inputs
The objects set the resolution: tanks and panels need very-high-resolution satellite or aerial imagery, and single trees need around 15 cm. We collect examples of each object, draw them as labels and keep some apart to test the model later.
Analysis
A deep-learning model, for object detection or segmentation, learns the objects from the labels and then scans the whole area tile by tile. Each detection comes with its outline or point, its class and a confidence score.
Result
Detections are cleaned, merged across tiles and reviewed in a map viewer, where doubtful ones are confirmed or removed. The result is one layer with every object, ready to count, filter and join to your records.
One point or outline per object, with its class and confidence.
Detections checked in a map viewer before delivery.
The model and its settings, to run on new imagery.
The labelled examples the model learned from.
How the model performed on test areas it never saw.
Limits
A model finds only what the imagery shows: objects under trees, roofs or shadow are missed, and accuracy drops on imagery unlike the training data, so results are reviewed, not taken on trust.
Infrastructure & UtilitiesAssets such as storage tanks detected across a region and gathered into one inventory. 
EnergySolar panels and wind turbines detected from imagery across a region. 
Agriculture & Land ManagementTrees counted one by one in orchards, from very-high-resolution imagery. 
Real Estate & DevelopmentBuilding footprints extracted from imagery where maps are missing or out of date. A deep-learning model is trained on examples of the object drawn on imagery. It then scans new imagery, marks each object it recognises and gives each a confidence score. The detections are reviewed and become a layer you can count and map.
It depends on how distinct the object is and how varied the landscape. A clear object on similar imagery needs fewer; a small or variable one needs more. We start with a pilot area, measure the results and add examples where the model fails.
Yes. Each tree found becomes a point on the map and each gap is located on its row, with the count per row. We did it for a hazelnut orchard in Romania from 15 cm satellite imagery. The imagery cannot tell a live tree from a dead one, so those positions are flagged for a check.
Yes. A segmentation model outlines each roof on very-high-resolution imagery, and the outlines become a building layer for planning, flood exposure or market studies.
Often, but not blindly. New imagery, seasons or building styles change what the model sees. We test it on a sample of the new area first and add local examples if the results drop.
Tell us the area, the decision and the deadline. We reply within 24 hours, from the person who will deliver the work.
Based in Bucharest, working with clients on five continents.
