Remote Sensing Labs
Browse all docs

DEMs

A digital elevation model answers a question the optical sensors cannot: not what the surface looks like, but how high it is. A DEM is a raster whose every pixel holds a ground elevation, and that single extra dimension reshapes a surprising amount of analysis — where water flows, which ground floods, how a hillside faces the sun, what a valley hides from a viewpoint. Most of this site’s sources measure reflected or emitted energy; a DEM measures shape, and it usually enters a workflow as context and constraint rather than as the thing being classified. The sections below treat terrain as its own input surface: what a DEM represents, how raw elevation differs from the products derived from it, which quality issues matter, and how to choose among the models on offer — with reasoning that holds across datasets rather than any one product’s figures.

What a DEM represents

At its core a DEM is elevation on a grid: a regular raster, like the imagery the Digital Imagery concept describes, but with height as the pixel value instead of brightness. Two distinctions matter immediately. The first is what “the surface” means. A digital surface model records the top of whatever is there — tree canopy, rooftops, bridges — while a digital terrain model aims for the bare ground beneath them; “DEM” is often used loosely for either, so knowing which one you hold is essential before you reason about flow or exposure, because a canopy-top surface routes water very differently from bare earth. The second distinction is elevation versus its derivatives. The raw height grid is rarely the end product: from it you compute slope (steepness), aspect (the compass direction a slope faces), hillshade (a shaded-relief rendering for visualization), flow direction and accumulation (which way water moves and where it concentrates), and watershed boundaries. These are all derived measurements, calculated from neighborhood differences in elevation — the derivative is only as good as the DEM and its grid spacing underneath it, which is why the source model’s quality propagates into everything built on top. Units matter as much as quality: a DEM stored on a geographic grid has pixel spacing in degrees but heights in meters, and computing slope on it directly mixes the two and returns nonsense, so reproject to a projected system (or convert the horizontal spacing to ground distance) before deriving slope.

Quality concerns that shape use

A DEM’s fitness for a task turns on a handful of quality properties, and ignoring them is where terrain analysis quietly goes wrong. Spatial resolution — the ground distance per pixel — sets what landforms are even representable, exactly the tradeoff the Resolution concept frames: a coarse global model captures broad topography efficiently but smooths away small channels, terraces, and berms that decide local flooding, while a fine local model resolves them at the cost of coverage and size. Distinct from horizontal detail is vertical accuracy — how close the recorded heights are to true elevation — which varies with terrain, land cover, and how the model was produced, and which dominates any analysis of small height differences. Real DEMs also carry artifacts: voids (missing data, common in steep or radar-shadowed terrain), striping or banding from acquisition, and edge effects where tiles meet. And because elevation is measured against a reference, datum differences matter: heights given relative to one vertical reference do not equal heights relative to another. A model referenced to a geoid (Copernicus DEM uses EGM2008), ellipsoid heights of the kind GNSS and satellite altimetry report, and a national vertical datum can disagree by tens of meters in places, and mixing them injects a systematic offset. Finally, terrain is frequently resampled and reprojected to line up with imagery, and every resampling of a continuous surface smooths it slightly — so the order and method of those steps is itself an analysis choice, tied to the Coordinate Systems reasoning about aligning layers.

Where DEMs fit in workflows

Terrain earns its place in a workflow as the layer that adds physical context to spectral data. The clearest case is water: flow direction, accumulation, and elevation thresholds are central to hydrology and to the terrain screening behind Flood Mapping, where low, flat, water-accumulating ground is where inundation concentrates and a DEM tells you where that is before any image is examined. Terrain also provides contextual and exposure information for planning and for Infrastructure Monitoring, where slope and elevation frame what a road, dam, or settlement sits on and is exposed to. A third role is normalization and stratification: because slope and aspect change how much sunlight a surface receives, elevation-derived layers help correct or stratify optical analysis — comparing vegetation on north- and south-facing slopes as like with like rather than blaming the sensor for a shadow. DEMs are also a natural masking layer, letting you exclude or separate areas by height or steepness. In each case elevation is a distinct raster combined with spectral inputs — a Raster vs Vector companion to the imagery — and because bare ground changes slowly, a DEM is often a stable backdrop against which the imagery’s Change Detection plays out, though large earthworks or landslides can change terrain itself.

Accessing elevation data

DEMs come from several production methods, and the method largely decides what a model represents. Global models built from satellite radar or stereo optical imagery give near-worldwide coverage at moderate resolution — the widely used free examples are the Copernicus DEM (GLO-30 and GLO-90, from TanDEM-X radar), NASADEM (a reprocessing of the Shuttle Radar Topography Mission), and JAXA’s ALOS World 3D (AW3D30, from stereo optical). All of them are effectively surface models: radar and stereo optical see canopy and roofs rather than the ground beneath. National mapping agencies publish finer models over their own territories, increasingly from airborne lidar, which reaches the ground between leaves and so yields genuine bare-earth terrain; the USGS 3D Elevation Program is the United States example. Bare-earth versions of global models exist too, but they carry their own licenses and errors, so read the terms before building on one. Local surveys provide the finest detail over small areas. That tiering is the practical decision: global for broad context, national or local lidar when the question hinges on small height differences or on where water runs beneath a forest — exactly the flood-screening case where a global surface model misleads. Elevation collections are loaded as ready-made layers in hosted analysis environments such as Google Earth Engine, distributed directly by public agencies, and mirrored in cloud archives, often described by the same STAC metadata used for imagery and cataloged alongside the broader open geospatial data ecosystem. Before trusting any derivative, confirm whether you hold a surface or a terrain model, check resolution and vertical accuracy against what the task needs, watch the datum and any voids, and resample deliberately when aligning to imagery.

Sources