Drone imagery analysis and detection pipelines

From flights to analysable data

The first decision is ground sample distance: how much ground one pixel covers, set by flying height, camera sensor and lens. The smallest object you need to detect must cover enough pixels for a model to recognise it, so the target object drives the flight plan, not the other way round. (TODO: set the minimum object size in pixels from a pilot test on your own imagery.)

Overlapping still images can be stitched into an orthomosaic, a single georeferenced image of the whole area, using photogrammetry software such as the open-source OpenDroneMap or commercial packages. The output is typically a GeoTIFF, often with an elevation model alongside. Video suits moving objects and quick inspections, but is harder to map precisely.

Detection and counting

Detection models such as the YOLO family are trained on annotated examples from imagery like yours. An orthomosaic is far larger than a model's input, so it is cut into overlapping tiles, each tile is processed, and detections on tile borders are merged so objects are not counted twice. Each detection's pixel position is converted to map coordinates using the image's georeference, which turns a picture into a point layer you can count, filter and map.

In video, a tracker links detections across frames so that one object passing through the view is counted once.

Change detection over time

Comparing surveys from different dates needs the images to line up. That means good georeferencing, from ground control points or RTK positioning, plus handling for differences in light and season that are not real change. Change can then be measured by comparing classified maps, detected objects or elevation models between dates, and reported as areas and locations rather than impressions.

Being honest about accuracy

• Accuracy is measured on held-out imagery from your own area, reported as precision and recall, not taken from a model's published benchmark.
• A model trained on one region, season or camera usually loses accuracy on another, and the pipeline should be re-checked when those change.
• Annotation is often the largest cost. A small, well-labelled set from your own surveys is worth more than a large generic one.

Outputs that fit existing workflows

Results are delivered in standard GIS formats (GeoJSON, shapefile, GeoTIFF and CSV) that open directly in QGIS or ArcGIS. The trained model weights and inference code are handed over, and the pipeline can run offline on your own machines where data cannot leave the organisation.