UAV photogrammetry is a continuous chain of steps that begins with flight planning and ends with map-scaled, georeferenced products. From near-vertical images with sufficient overlap, the three-dimensional structure of the scene is reconstructed, and the position of the camera at each exposure is recovered together with the coordinates of ground features. The quality of the result depends far more on the design of the image block, the ground control and the reference data than on the processing software alone. This page describes the technical steps of the service, from ground sample distance and overlap through dense point cloud, elevation models, orthomosaic and contours. It also explains how accuracy is reported as root mean square error and states the known limits of the method.
Flight Planning, Ground Sample Distance and Overlap
Every project starts from a technical question: how small is the smallest feature that must be visible in the product, and what spatial accuracy does the client need. These two criteria set the ground sample distance, and the GSD together with the focal length, the sensor pixel size and the flying height then determines the mission altitude. Flying lower produces a finer GSD but reduces the footprint of each frame and increases the number of images required.
Forward overlap between consecutive images along a strip and side overlap between adjacent strips are the precondition for automatic image matching and correct reconstruction. The required overlap is not a fixed figure; over rough, densely built or well-vegetated terrain more overlap is needed to compensate for feature displacement and occlusion. Flight lines are planned around sun angle, wind direction and speed, safe clearance from obstacles, and available flight endurance.
Large areas are divided into several flight blocks so that lighting and altitude remain consistent within each block. Oblique acquisition is added when facades and building sides are part of the intended product rather than a by-product.
- Derive the GSD from the smallest target feature and the required accuracy, not from the highest altitude the aircraft can reach.
- Increase forward and side overlap over urban and well-vegetated areas.
- Split the area into blocks of uniform light and elevation, and record sun and wind conditions per block.
- Check full coverage of the boundary and a surrounding margin, including a buffer outside the contract area.
Ground Control, RTK and PPK
A reconstructed model is only a relative structure until it is tied to real coordinates. Ground control points are selected across the project area, marked on the ground and measured with precise satellite receivers. Their distribution must be spread and balanced, covering the boundary and the centre of the area; concentrating all points in one part introduces systematic error in the regions far from control.
With RTK, positioning is performed relative to a base station operating during the flight, while with PPK the carrier-phase data recorded on board the aircraft are processed after the flight against a base station. Both methods tag the exposure centre of each image with precise coordinates and can therefore reduce the number of ground control points needed; keeping a few independent check points is nevertheless advisable for validating the product.
These check points, usually called check points or independent check points, are excluded from the adjustment and used only to measure final accuracy objectively. Documenting the coordinate reference system, the datum and the measurement method of the control is part of every project deliverable.
Aerial Triangulation and Bundle Adjustment
In aerial triangulation, corresponding points between overlapping images are detected automatically and form a set of tie points. From these points, the position and orientation of each image are estimated in a common frame. Bundle adjustment, usually solved by linearised least squares, determines the camera parameters that minimise the reprojection error of image points in object space.
At this stage the interior orientation of the camera, including focal length and distortion coefficients, is also estimated by self-calibration so that lens distortion is absorbed by the model. Robust estimation such as RANSAC is used to detect and remove blunders, because a single wrongly matched tie point can bias the whole block.
Introducing the ground control points then transforms the relative network into absolute coordinates. Inspecting the residuals at this stage shows whether a systematic error remains in the control data or in the imaging geometry.
Structure from Motion and Dense Matching
Structure from motion uses the multi-view geometry of the same image set to produce a sparse point cloud from the tie points, which represents the overall scene structure. This sparse cloud is enough to display the geometry but not dense enough for surface modelling and accurate measurement. In the next step, dense matching algorithms produce a coloured dense point cloud from per-image depth maps, representing the surface detail of the ground.
A dense cloud requires cleaning: removing outlier points, clipping it to the contract boundary and, where needed, classifying points automatically into ground and non-ground. Point cloud quality directly affects the quality of the elevation models and the orthomosaic, so the density and the cleaning method are documented in the technical report.
- A sparse point cloud from triangulation, used to check geometry and coverage.
- A coloured dense point cloud, used for surface modelling and measurement.
- Automatic classification of ground and non-ground points, a prerequisite for the terrain model.
- Clipping to the project boundary and delivery in standard point cloud formats.
Elevation Models: Surface, Terrain and Height
From the dense point cloud a digital surface model is generated, representing the visible surface of the ground including buildings, trees and other objects. The digital terrain model is built from the same cloud after non-ground objects have been removed, showing only the natural configuration of the ground. The difference between the two forms the canopy height model, used to estimate tree crown height and vegetation volume.
Terrain modelling depends on the filtering and classification of points, and its quality decreases where vegetation is dense, because the ground surface below the canopy is not directly visible. In such places the terrain elevation is an estimate and must be reported as a limitation of the product.
The models are normally delivered either as rasters at a selectable spatial resolution or as triangulated irregular networks. The chosen resolution should match the GSD of the images and the scale of the client’s application; a resolution finer than the underlying data does not increase real accuracy.
Orthomosaic, Contours and Breaklines
An orthomosaic is produced by geometrically rectifying each image with an elevation model and then merging them into a single continuous image. Displacement caused by feature height and camera tilt is corrected so that the output image has map-scale properties. Seamlines are selected carefully and colour is balanced between images; the result is normally delivered as GeoTIFF and published as COG.
Contours are extracted from the digital terrain model at a vertical interval suited to the map scale, with controlled smoothing to avoid artificial noise. Breaklines record natural and constructed edges such as road centrelines, stream banks and terrace edges so that the surface between them is interpolated correctly.
Vector outputs, including contours, breaklines, parcel boundaries and topographic layers, are prepared in interchange formats such as Shapefile and DXF with their attribute tables, so they can be used in the client’s GIS.
Accuracy, Independent Validation and Method Limits
Product accuracy is expressed as root mean square error in GSD units, giving a measure that is comparable across projects and independent of scale. Horizontal and vertical error are computed separately and reported as a ratio to the GSD. This validation is performed on independent check points that took no part in the adjustment; using the control points themselves to measure accuracy produces an optimistic result.
The photogrammetric method has known limits that must be established before the project. Shadows and occlusion hide parts of the ground from the camera, and there the elevation is interpolated. Dense vegetation makes both terrain reconstruction and image matching difficult.
Water, glass and polished metal reconstruct poorly because of specular reflection and the lack of reliable texture, and moving objects appear as smearing in the model. These effects are recorded in the product report together with the flight metadata and the coordinate reference system.
- Separate computation of horizontal and vertical error in GSD units.
- Validation with independent check points outside the adjustment.
- Explicit statement of limits for water, glass, dense vegetation and moving objects.
- Recording of flight metadata, spatial reference and processing method with each product.
Summary
UAV photogrammetry produces dependable results when every step is designed as one chain: from GSD and overlap through ground control and aerial triangulation to point cloud, elevation model and orthomosaic. A weakness in any link is passed on to the following ones, and correcting it at the end of the project is expensive.
The deliverables of this service are a set of georeferenced products with documented accuracy: point cloud, surface, terrain and canopy models, orthomosaic, contours, breaklines and topographic vector layers. Every product is delivered with metadata and a method report so the client can judge whether the data suits the intended use.
To start an assignment it is enough to define the project boundary, the smallest target feature, the required spatial accuracy and the form of the final product; from these, flight altitude, image coverage, ground control design and the delivery list can be prepared.