Photogrammetry Explained: How Drone Images Become 3D Models

You fly a drone over a site and capture 2,000 photographs. Each photograph is a flat, 2D image. How do those flat images become a 3D model of the terrain?

The answer is photogrammetry — a 150-year-old science that has been transformed by modern computing into one of the most powerful tools in geospatial data collection.

Photogrammetry is based on the same principle your brain uses to perceive depth: stereo vision. When you look at an object with both eyes, each eye sees it from a slightly different angle. Your brain compares the two views and uses the difference (parallax) to calculate distance.

Photogrammetry does the same thing with photographs. By comparing the same feature as seen from multiple camera positions, the software can calculate the 3D position of that feature.

The more overlapping views of a feature, the more accurately its 3D position can be determined.

The drone flies a planned grid pattern, capturing images at regular intervals with significant overlap — typically 75–85% frontal overlap and 60–70% side overlap. This means each point on the ground is visible in 5–10 or more images from different angles.

Each image is tagged with GPS coordinates, altitude, and orientation data from the drone's sensors.

The first processing step is Structure from Motion (SfM) — an algorithm that identifies common feature points across overlapping images and uses them to reconstruct the relative positions and orientations of all the cameras.

SfM works by finding distinctive features in each image (corners, edges, texture patterns) and matching them across images. When the same feature is found in multiple images, the algorithm can calculate the camera positions that would produce those observations.

The result is a sparse point cloud — a set of 3D points representing the matched features — and a set of camera positions.

Once camera positions are known, Multi-View Stereo (MVS) algorithms use all the images to compute a dense point cloud — millions of 3D points covering the entire surveyed area.

MVS works by taking each pixel in each image and finding the corresponding pixel in every other image that shows the same point. The 3D position of that point is calculated from the camera positions and the pixel coordinates.

A dense point cloud from a typical drone survey might contain 50–500 million points.

The dense point cloud is used to reconstruct surfaces:

• Mesh — a triangulated surface connecting adjacent points, used for 3D visualization and printing
• DEM — a raster grid of elevation values interpolated from the point cloud
• Orthomosaic — the original imagery projected onto the DEM and stitched into a seamless, geometrically corrected map

The reconstructed model is anchored to real-world coordinates using GPS data from the drone and/or ground control points. This step transforms the model from an arbitrary coordinate system to a real-world coordinate reference system.

Photogrammetry depends on finding matching features across images. Surfaces with rich texture and contrast — grass, gravel, vegetation — produce many matching features and high-quality point clouds.

Surfaces with uniform appearance — smooth concrete, water, sand, snow — produce few matching features and poor-quality point clouds. These surfaces are a known limitation of photogrammetry.

More overlap means more views of each point, which means more accurate 3D reconstruction. The standard 75/60% overlap is a balance between accuracy and flight efficiency. For challenging surfaces or high-accuracy requirements, higher overlap (85/75%) improves results.

Consistent, diffuse lighting produces the best photogrammetry results. Harsh shadows create areas of high contrast that can confuse feature matching. Overcast days often produce better photogrammetry results than bright sunny days.

The photogrammetry software needs to know the precise characteristics of the camera — focal length, principal point, and lens distortion parameters. Most professional mapping drones have pre-calibrated cameras, but calibration can drift over time and should be verified periodically.

Photogrammetry cannot see through vegetation. In forested areas, the point cloud represents the top of the canopy, not the ground. For bare-earth terrain modeling in vegetated areas, LiDAR is required.

Photogrammetry assumes the scene is static during the survey. Moving objects — vehicles, people, water — create artifacts in the point cloud and orthomosaic. Active construction sites with moving equipment require careful flight planning to minimize these artifacts.

Highly reflective surfaces — windows, water, polished metal — produce inconsistent reflections across images, which confuses feature matching and produces poor-quality point clouds.

Photogrammetry is a mature, well-understood technology that produces high-quality 3D data from drone imagery. Understanding its principles and limitations helps you plan surveys that produce the best possible results for your application.

Contact Blackridge Geospatial to discuss how photogrammetry can support your project.