Photogrammetry Workflow

Photogrammetry is the software side of drone mapping: turning a folder of overlapping photographs into a usable 3D model or map. A well-flown mapping mission supplies the raw material; this guide covers what happens to those images afterward, from the first alignment pass through a finished, checkable deliverable.

Step 1: image alignment

The first processing step identifies distinct visual features across the image set and matches the same features between overlapping photos. From those matches, the software calculates the approximate position and orientation each photo was captured from, a process sometimes called structure from motion. This step produces a sparse point cloud, a rough skeleton of the scene built from the strongest matched features, along with a calculated camera position for every photo in the dataset. Poor overlap or low visual texture in the original flight shows up here first, as gaps in coverage or as photos the software cannot confidently place.

Step 2: dense point cloud generation

Once camera positions are known, the software revisits every image and calculates many more matched points than the sparse alignment pass used, producing a dense point cloud: potentially millions of individual 3D points representing the surface of the scene in far more detail than the sparse pass. This step is one of the more computationally intensive parts of the pipeline and is where processing time scales most directly with image count and desired detail level.

Step 3: mesh generation

A mesh connects the dense point cloud into a continuous triangulated surface, turning a set of disconnected points into something closer to a solid 3D model with defined faces and edges. A mesh is generally the deliverable of choice when the goal is a visual or structural 3D model, such as for a building exterior or a stockpile volume calculation, rather than a flat map.

Step 4: orthomosaic and DEM export

For flat, top-down mapping deliverables, the software instead projects the model into an orthomosaic: a single, geometrically corrected, seamless image assembled from all the source photos with lens distortion and perspective differences removed, so every part of the image sits at a consistent, accurate scale. A digital elevation model, or DEM, is a related output that represents surface height as a grid of elevation values rather than an image, used for terrain analysis, drainage studies and cut-and-fill volume calculations.

Photogrammetry output types

Output	What it is	Common use
Sparse point cloud	Rough feature skeleton from alignment	Intermediate step, not a deliverable
Dense point cloud	Millions of matched 3D points	Detailed surface analysis, further processing
Mesh	Triangulated continuous 3D surface	Visual models, structural inspection
Orthomosaic	Flat, geometrically corrected top-down image	Site maps, land surveys, agriculture
DEM	Grid of elevation values	Terrain analysis, volumetric calculations

Checking accuracy the right way

A model built entirely from onboard GPS positions with no ground control is useful for a visual reference but carries positional uncertainty on the order of several meters, which is unacceptable for anything measurement-driven. Ground control points placed and surveyed before the flight, then identified manually in the software after capture, let the model be corrected against known real-world coordinates rather than GPS alone, as covered in mission planning. To actually verify the result rather than assume it, hold back a subset of surveyed points as checkpoints, excluded from the correction step, and compare the model's calculated position for each checkpoint against its known surveyed position. The resulting discrepancy is a real, defensible accuracy figure you can put in a client deliverable, rather than an assumed number.

Common failure points and what causes them

Gaps or holes in a finished model trace back almost every time to one of three causes: insufficient overlap during the flight, an area the flight lines simply did not cover, or a surface with too little visual texture to match reliably, such as still water, a uniform painted roof, or fresh snow. None of these are fixable in processing software after the fact; the fix is reflying the affected area with tighter overlap or a different flight pattern. Blurry or doubled features in a mesh usually point to motion blur in the source images, which traces back to shutter speed and ground speed settings covered in drone camera settings, even though that guide is written primarily for video.

Hardware and processing time

Processing time depends heavily on image count, resolution and target output quality, and it scales faster than linearly with image count since matching complexity grows with every added image relative to every other image. A small dataset of a few hundred images can process on a capable laptop in under an hour at draft quality; a large dataset of several thousand images at full mesh detail can take many hours, sometimes run overnight, and benefits substantially from a dedicated workstation with a strong GPU rather than a laptop. Offload captured images to reliable storage immediately after each flight using a fast card reader and a portable SSD, rather than processing directly off a memory card, both for speed and to protect the only copy of a full day's capture.

Managing a photo dataset before processing even begins

Good file organization before processing starts saves significant time compared to sorting it out afterward. Create a dedicated folder per flight or per site visit, named consistently with a date and site identifier, and copy images directly from the card using the workflow covered in the field kit guide rather than processing straight off the memory card, which is both slower and risks the only copy of the dataset if the card fails partway through a long processing run. Keep the original, unedited images in a clearly labeled folder separate from any processed outputs, since photogrammetry software occasionally needs to reprocess a dataset from scratch if a setting was wrong the first time, and having the clean originals readily available avoids a return trip to re-fly a site that was already captured correctly the first time.

Culling obviously bad images, such as ones affected by motion blur, an obstructed lens, or a momentary exposure spike, before starting the main processing run both speeds up processing and can improve the final result, since a small number of genuinely bad images mixed into an otherwise good dataset can occasionally confuse the alignment step. Be conservative about culling, though; removing too many images, especially ones that only look marginally soft, reduces the overlap the software depends on and can introduce the same gaps a poorly flown mission would produce.

Choosing processing settings for the actual deliverable

Most photogrammetry software offers processing quality presets, commonly labeled something like draft, medium and high, and picking the right one for the job matters as much as any capture decision. A draft-quality pass is genuinely useful early in a project to quickly check coverage and catch any gaps before committing hours to a full-quality run, and re-flying a missed area is far cheaper to fix at draft-quality-discovery time than after a full high-quality process has already completed. Reserve full processing quality for the final deliverable pass, since it is where the meaningful detail difference actually shows up in the finished mesh or orthomosaic, and running every intermediate check at full quality wastes significant processing time for no real benefit until the dataset itself is confirmed complete.

Export settings deserve the same deliberate attention as processing quality. An orthomosaic exported at full native resolution can be an extremely large file, sometimes multiple gigabytes for a large site, which is unnecessary if the client's use case, such as a general site reference image, does not require that level of detail. Match the export resolution and file format to how the client will actually use the deliverable: a GIS-ready GeoTIFF for a client who will load it into their own mapping software, a compressed JPEG or PDF for a client who mainly needs to view and reference the map, or both if the use case is not yet clear.

Where this fits into commercial work

Photogrammetry output feeds directly into client deliverables for roof inspection reports, construction progress tracking, and land survey work, and every one of those jobs assumes the operator is flying under Part 107 with appropriate insurance in place, covered in drone insurance explained.

Storage for a photogrammetry set

A single mapping mission produces hundreds of full-resolution stills, and processing reads them all repeatedly. Working from a SanDisk 1 TB portable SSD (USB-C) ($166) rather than a card reader cuts the slowest step of the pipeline, and gives you somewhere to keep the raw set once the model is built.