What is the difference between DEM, DTM, and DSM?

A DSM (digital surface model) records the first reflective surface - canopy, rooftops, structures. A DTM (digital terrain model) records bare earth with those features removed. DEM (digital elevation model) is the general term; in common usage it means a bare-earth grid, close to a DTM, but check the metadata because usage varies.

Should I use a DTM or a DSM for terrain analysis?

Use a DTM for slope, hydrology, watershed delineation, and geomorphology, where you want bare ground. Use a DSM for line-of-sight, viewshed, solar, flight-obstruction, and canopy/building analysis, where the objects on the surface are the point.

How do I make a canopy height model from a DSM and DTM?

Subtract the DTM from the DSM cell by cell (CHM = DSM - DTM), after confirming both share the same CRS, cell size, extent, and vertical units. The result is height above ground for canopy or structures. Clamp small negatives to zero and mask NoData.

Is a higher-resolution DEM always better?

No. Finer cells reveal real detail but also amplify noise and micro-relief that can swamp the landform signal, and they cost storage and processing time. Match cell size to the feature scale you are mapping; a 10-30 m DEM is often better for regional landforms than a 1 m grid.

DEM vs DTM vs DSM: Differences Explained

Short answer

These three terms describe what surface the elevation values represent. A DSM (digital surface model) is the first reflective surface: tree canopy, rooftops, bridges, anything the sensor hits first. A DTM (digital terrain model) is the bare earth with vegetation and structures stripped out. DEM (digital elevation model) is the umbrella term, and in everyday use it usually means a bare-earth grid (effectively a DTM) — but the convention is inconsistent across providers, so you must read the metadata. The practical rule: bare earth (DTM/DEM) for slope, hydrology, and geomorphology; first surface (DSM) for viewsheds, line-of-sight, solar, and canopy or building height.

Defining the three precisely

DSM — every cell holds the elevation of the topmost object. Over a forest it is the canopy top; over a city it is the rooftops. Derived from the first return of a LiDAR pulse, or from photogrammetric image matching (which sees the surface, not under it).
DTM — every cell holds the bare-ground elevation. Produced by classifying LiDAR points into ground vs non-ground (e.g. progressive morphological or cloth-simulation filtering) and interpolating a surface from ground returns only. A DTM may also carry breaklines (ridges, stream centrelines) for a more faithful terrain.
DEM — generic "elevation grid." USGS 3DEP products, SRTM, and Copernicus DEM are all called DEMs; SRTM at 30 m is effectively a DSM over forest and built-up areas because C-band radar did not penetrate canopy, whereas a LiDAR-derived 3DEP DEM is bare earth. Same word, different surface — hence: always check the metadata.

The relationship is simple arithmetic: over an object, DSM ≥ DTM, and the difference is the object's height.

Choosing the right model for the task

Analysis	Use	Why
Slope, aspect, curvature	DTM	Derivatives must describe the ground, not roof pitch or canopy
Hydrology, flow accumulation, watersheds	DTM	Water runs on bare earth; buildings/canopy create false barriers
Geomorphology, landform mapping	DTM (often smoothed)	Bare-earth relief is the signal
Viewshed, line-of-sight, telecom	DSM	Trees and buildings block sightlines
Solar radiation, shadow modelling	DSM	Canopy and structures cast the shadows
Flight obstruction, drone planning	DSM	Obstacles are the whole point
Canopy height, biomass, building height	DSM − DTM (CHM/nDSM)	Height above ground is the measurement

Running hydrology on a DSM is a classic error: buildings and tree clumps become artificial dams that divert modelled flow.

Worked example: canopy height model

To get a canopy height model (CHM), also called a normalized DSM, subtract bare earth from the surface. The non-negotiable prerequisite is that both rasters are co-registered: same CRS, same cell size, same extent/origin, same vertical units.

Align the grids. If the DSM and DTM differ, warp one to match the other exactly:
```
gdalwarp -t_srs EPSG:25832 -tr 1 1 -te 600000 5600000 601000 5601000 \
         -r bilinear dsm.tif dsm_aligned.tif
```
(-te sets a common extent; -tr a common 1 m cell.)

Subtract with gdal_calc.py, clamping spurious negatives to zero:

gdal_calc.py -A dsm_aligned.tif -B dtm.tif --outfile=chm.tif \
  --calc="(A-B)*(A>B)" --NoDataValue=-9999 --type=Float32

Sanity-check the result: canopy heights should be physically plausible (a few metres to a few tens of metres), with bare fields near zero. Large negatives or implausible spikes mean misalignment or a vertical-datum mismatch.

In QGIS the same is the Raster Calculator ("dsm@1" - "dtm@1"); in ArcGIS Pro it is the Minus tool or Raster Calculator.

Resolution: finer is not automatically better

Cell size sets the smallest feature you can represent and how derivatives behave:

Terrain derivatives are neighbourhood operations. Slope at a cell is computed from a 3×3 window (Horn's method in GDAL gdaldem slope and ArcGIS Slope). On a 1 m DEM that window spans 3 m; on a 30 m DEM it spans 90 m. The same algorithm therefore reports different slope because it samples different scales.
A 1 m LiDAR DEM exposes tractor furrows, footpaths, and noise that are irrelevant to regional landform analysis and can dominate the slope histogram. For mapping drainage and landforms over a watershed, a 10-30 m DEM often gives a cleaner signal and runs far faster.
Finer grids cost storage and compute non-linearly: halving cell size quadruples cell count.

The discipline is to match cell size to the feature scale you care about, and to resample deliberately (with documented method) rather than defaulting to the finest available product.

Units, NoData, and vertical datums

Vertical units must match horizontal expectations: if elevation is in metres but the DEM is in a geographic CRS (degrees), gdaldem slope will produce nonsense unless you pass -s (scale) to convert, or better, reproject to a projected CRS first.
Slope units — degrees vs percent rise — are not interchangeable. Decide and label which one the output uses.
Vertical datum — orthometric (NAVD88, EGM2008 geoid) vs ellipsoidal heights can differ by tens of metres. A CHM built from a DSM and DTM on different vertical datums is meaningless.
NoData — voids, water masks, and edge fill must be set as NoData, not 0 or -9999 treated as real elevation, or they corrupt slope and hydrology at the boundary.

Common pitfalls and why they happen

Assuming "DEM" means bare earth. SRTM and many global DEMs are surface-like over vegetation/buildings. Running hydrology on them inserts canopy "dams." The word does not tell you the surface — the metadata does.
Computing a CHM from misaligned rasters. Different origins or cell sizes make the subtraction sample the wrong cells, producing salt-and-pepper height. Warp to a common grid first.
Slope on a geographic-CRS DEM. Degrees of latitude/longitude are not metres, so a degree-based slope is wrong by a latitude-dependent factor. Reproject to UTM or State Plane first.
Chasing maximum resolution. A 1 m grid amplifies micro-relief and noise; for landform-scale work it can be worse than a smoothed 10 m product.

QA and validation

Confirm horizontal CRS, vertical datum, cell size, and NoData before any derivative.
Spot-check elevations against known benchmarks or GNSS points.
For a CHM, verify open ground sits near zero and canopy heights are physically plausible.
Inspect edges, voids, striping (radar/photogrammetry artifacts), and pits before using slope or hydrology outputs.
Record source product, resolution, vertical datum, slope units, and processing method with the layer.

Bathyl perspective

We choose the elevation surface from the question, not from whatever file is on hand: bare-earth (DTM/DEM) for slope, hydrology, and geomorphology; first-surface (DSM) for visibility, solar, and canopy or structure height. Derivatives ship with their resolution, vertical datum, and units stated, because a slope or watershed layer is only as trustworthy as the elevation model and processing assumptions behind it.

DEM vs DTM vs DSM