Digital Terrain Models

These get used interchangeably in marketing copy, and the confusion costs real money: a volume computed on the wrong one is a wrong volume.

• DSM — digital surface model. The top of everything: tree canopy, rooftops, stockpile crests, parked trucks. What photogrammetry gives you natively, because the camera sees surfaces, not the dirt under them.
• DTM — digital terrain model. Bare earth, with vegetation and structures stripped out, often enriched with breaklines at the grade changes. What engineering design, drainage, and earthwork actually need.
• DEM — digital elevation model. The umbrella term — a gridded elevation raster of either flavor. USGS national data is DEM in this sense. When someone hands you "a DEM," your first question is whether it’s surface or terrain.

The work of going DSM → DTM is classification — deciding which points are ground. Software does a first pass; the difference between a cheap surface and a defensible one is a human checking it where it matters.

• TIN — triangulated irregular network. Your points become triangle vertices; the surface is the set of flat triangle faces. Honors every measured point exactly, densifies where you measured densely, and accepts breaklines. This is what Civil 3D surfaces are.
• Grid / raster. Elevations resampled onto uniform cells (e.g. 1 ft pixels). Compact, fast, perfect for rasters like slope maps and flood models — but every cell is an interpolation, and sharp grade breaks get rounded off.
• The rule of thumb. Measure and design on TINs; publish and analyze on grids. Converting TIN → grid is cheap and safe at an appropriate cell size; grid → TIN cannot recover the edges the grid already smoothed away.

Breaklines are the fix for the left side of that picture: linear features (top and toe of slope, curb lines, ditch flowlines, wall faces) that force triangle edges to follow the real grade break instead of cutting across it. A few hundred shots plus good breaklines beats ten thousand shots without them.

• 1. Classify ground. Run the automatic ground filter (Terra, Pix4D, TBC, or LiDAR-specific tools), then audit it on cross-sections — under trees, against buildings, on stockpile toes, anywhere it matters to the deliverable.
• 2. Thin with intent. A 400-million-point cloud is not a surface; it’s raw material. Decimate to a spacing the project needs (e.g. 0.5–1 ft on structures, 2–5 ft on open ground) so the TIN stays workable.
• 3. Add breaklines. Extract or field-survey the grade breaks. This is where surveyor judgment outperforms any filter.
• 4. Build and groom the TIN. Clip to the data boundary, fix flat spots and crossing edges, exclude low-confidence areas honestly rather than letting interpolation invent them.
• 5. Check against independent shots. Compare the finished surface to check shots it has never seen — RTK topo singles on hard surfaces and open ground. Report the residuals with the deliverable.

Here's the pipeline above run for real — extracting a surface model from a DJI L2 point cloud and landing it in Civil 3D as a usable design surface.

• Contours. Still the universal language of terrain. Generated from the TIN, smoothed lightly for drafting — and honest about interval (see the formula above).
• Volumes. Surface-to-surface comparison: existing vs design, this month vs last month on a stockpile. The number is only as good as BOTH surfaces and the boundary definition — state all three with the result.
• Machine control. Dozers and graders consume the design DTM directly and cut to it without stakes. This raises the stakes on surface quality: a bust in the model becomes a bust in the dirt at production speed, which is why machine-control surfaces get checked against control before upload.

Capture-side details — flight design, GSD, and the ground control that anchors all of this — live in the Drone Surveying guide at /free/drone-course, and the control workflow itself in the Control Surveys reference.