Object detection
Locate supported objects or regions in an image, such as tanks, vehicles, rooftops or field features.
AI object detection & classification
Find meaningful features in complex imagery.
Satvitech applies computer-vision workflows to suitable satellite and aerial images to detect, classify, and organise visible objects or features according to a defined analysis objective.
Example visualisationWhat the service can include
From a visual target to a structured answer
Not every project needs every output. Formats and categories are agreed during scoping.
Classification
Assign defined categories to detected objects or image areas when the data supports reliable separation.
Image annotation
Prepare images with bounding boxes, segmentation masks, labels or other visual annotations where appropriate.
Structured results
Organise outputs in tables or machine-readable files that slot into your reporting or analysis.
Geospatial attributes
Include positions or spatial attributes when the source imagery is properly georeferenced and the method supports them.
Tailored categories
Assess whether a customer-specific class or visual target can be detected reliably enough for your intended use.
Potential application areas
Possible targets may include…
Whether a target is detectable depends on its size relative to the image resolution, the sensor, image quality and your permission to use the imagery. These examples show the kind of features that may be suitable:
- Supported infrastructure features, such as buildings, tanks or rooftop installations
- Selected vehicles or vessels in imagery with sufficient resolution
- Solar panel arrays on rooftops or on the ground
- Storage areas, container stacks and stockpiles
- Agricultural features such as field structures or farm buildings
- Visible modifications to industrial sites, compared across dates
How a project works
Five steps from class definition to delivery
-
Step 1: Define classes and use
Agree the target classes and the decision the results will support.
-
Step 2: Review imagery
Check source imagery against the expected object size and resolution.
-
Step 3: Choose the approach
Select, adapt or develop an appropriate computer-vision method.
-
Step 4: Evaluate the output
Review false positives and missed detections against reference examples.
-
Step 5: Deliver
Provide the agreed annotations, data and documentation of limitations.
What affects detection quality?
Results depend on the image as much as the model
A detection workflow can only find what the imagery shows clearly. These factors decide what is realistic for each project — we assess them before committing to an approach.
Pixel size vs. object size
A small object needs a fine enough resolution to cover it with enough pixels to be recognised.
Sensor characteristics
Optical, multispectral and radar sensors show objects differently and need different methods.
Acquisition conditions
Image date, lighting, shadows, viewing angle, haze, cloud cover and overall image quality.
Scene complexity
Crowding, occlusion, similar-looking backgrounds and varied object appearance.
Training & evaluation data
Representative examples and clear class definitions matter as much as the model.
Success criteria
How success is measured and how much manual review is appropriate for the use.
Detection is not identity verification
Detecting “a vehicle” or “a vessel” describes what is visible in the image. It does not establish which specific object it is, who owns it or what it is used for.
Confidence scores need context
Where results include a model confidence score, it is not automatically a calibrated probability that the detection is correct. We explain how scores were reviewed and what thresholds mean for your use.
Potential outputs
What a delivery can look like
Below is an illustrative output package for the example scene on this page. Every count and box corresponds to the objects drawn in that example image.
Example visualisationCategory counts
Illustrative data4Storage tanks
6Vehicles
3Container stacks
2Buildings
1Solar array
Detection table
Illustrative data · pixel coordinates| ID | Class | x, y (px) | Size (px) |
|---|---|---|---|
| T-01 | Storage tank | 735, 175 | 100×100 |
| T-02 | Storage tank | 855, 175 | 100×100 |
| T-03 | Storage tank | 735, 295 | 100×100 |
| T-04 | Storage tank | 855, 295 | 100×100 |
| V-01 | Vehicle | 379, 366 | 22×95 |
| V-02 | Vehicle | 459, 366 | 22×95 |
| V-03 | Vehicle | 724, 496 | 95×22 |
Showing 7 of 16 rows. With properly georeferenced imagery, positions can be given as map coordinates instead of pixels.
Structured export (excerpt)
Illustrative format · agreed per project{
"image": "example-site.tif",
"note": "Illustrative output. Pixel bounding boxes refer to the example image.",
"detections": [
{ "id": "T-01", "class": "Storage tank", "bbox_px": [735, 175, 100, 100], "review": "accepted" },
{ "id": "T-02", "class": "Storage tank", "bbox_px": [855, 175, 100, 100], "review": "accepted" },
{ "id": "S-01", "class": "Solar array", "bbox_px": [322, 180, 316, 132], "review": "accepted" }
]
}FAQ
Object detection questions
What types of objects can be detected?
Visible objects or features that are large enough relative to the image resolution and distinct enough from their surroundings — for example buildings, tanks, vehicles, solar arrays or stockpiles. Each target class is assessed for feasibility before work starts.
Does every satellite image contain enough detail?
No. Many freely available satellite images have pixels several metres across or larger, which is suitable for broad features but not small objects. The required resolution depends on the size of the target and how reliably it must be separated from similar features.
Can customer-supplied images be used?
Often, yes — provided you have the right to use them for the analysis, the format is workable, and the imagery has enough quality and detail for the agreed classes.
Can results include geographic coordinates?
Yes, when the source imagery is properly georeferenced and its metadata supports it. Otherwise results are given in image (pixel) coordinates, as in the example on this page.
How is accuracy reviewed?
Outputs are checked against reference examples or a reviewed sample. We record false positives and missed detections, explain the limitations, and agree beforehand what level of performance is good enough for your use.
Related services
Combine with other capabilities
Object detection is often paired with change detection to see how the objects on a site vary between dates.
Have a specific object or feature in mind?
Describe what you need to detect, how the results will be used, and any imagery you already have. We will assess whether it can be detected reliably enough.