What is detection?
Object detection answers what is present and where it is. It produces a location for each recognised object, usually a bounding box, together with a class and score. A box locates an object but does not trace its exact boundary.
Inputs and outputs
| Input | Output |
|---|---|
| Image tiles with consistent bands and resolution; training labels contain a box and class for each target. | Object boxes, classes and scores. Convert image coordinates to map coordinates before GIS export. |
Where it helps in GIS
Locate ships, inventory visible vehicles or find buildings across large imagery. Small targets need enough pixels to distinguish them.
Model families and examples
- YOLO: a widely used detector family with training and prediction workflows.
- Faster R-CNN: a region-proposal-based detector.
- RetinaNet: a single-stage detector to include in comparisons.
These examples explain the task. The sidebar shows only models currently published on GISSchools.
How to start training
Define classes and consistent box-labeling rules. Include background tiles. Split by geographic area before tiling, then fine-tune a suitable checkpoint. Use validation data to choose confidence thresholds and inspect both false detections and missed objects.
How to judge the result
Report precision, recall and average precision with stated overlap thresholds. Check small and crowded objects separately. A checkpoint trained on ordinary photographs is not automatically suitable for satellite imagery.
Learn more
The category image is an illustration, not an evaluated model prediction.