A digital elevation model (DEM) is a numerical representation of the elevation of a surface at geographically referenced locations. Most commonly, it describes the land surface as a grid of height values, although elevation can also be represented by irregularly distributed points, connected triangles, and other structures. A DEM is not simply a shaded picture of terrain: it is a dataset from which measurements, profiles, drainage networks, and three-dimensional visualizations can be calculated.
Elevation models are foundational data in a geographic information system. They support the production of topographic maps, the delineation of watersheds, flood simulation, infrastructure design, and investigations of landscape change. Their interpretation depends on what surface was measured, how densely it was sampled, the reference system used for heights, and the errors introduced during measurement and processing. Two datasets with the same nominal grid spacing can therefore differ substantially in their suitability for a particular analysis.
Meaning, scope, and surface definitions
The term digital elevation model has both broad and narrow uses. In its broad sense, it includes digital representations of ground elevation and of surfaces containing vegetation, buildings, or other objects. In a narrower convention, DEM denotes a regularly gridded representation of bare-earth terrain. Because terminology varies among mapping agencies and scientific disciplines, the product specification is more informative than the label alone.
A digital terrain model (DTM) generally represents the ground after vegetation and artificial structures have been excluded. In some surveying traditions, however, DTM refers specifically to a structured terrain representation containing surveyed points, breaklines, and other information about landform geometry, rather than merely a bare-earth raster. A digital surface model (DSM) represents an upper surface that may include tree canopies, roofs, bridges, and exposed ground. Subtracting a compatible DTM from a DSM can estimate object height, but differences in survey date, horizontal alignment, or surface sampling can create misleading results.
“Bare earth” is itself an operational definition. A classification procedure may remove buildings while retaining embankments, road surfaces, and other human-modified ground. Bridges may be removed to expose the channel beneath them, whereas dams are normally retained as terrain features. In dense vegetation, the ground surface may be inferred from sparse observations rather than directly measured at every location. The resulting model is consequently a reconstruction of the intended surface, not a complete inventory of observations.
Most DEMs are 2.5-dimensional: each horizontal position has one elevation value. This is sufficient for many landscapes but cannot represent an overhang, cave, tunnel, or vertical wall with multiple heights at the same horizontal coordinate. Fully three-dimensional meshes or volumetric models are needed for such geometry. The same distinction matters in cities, where a conventional elevation grid cannot separately describe a bridge deck and the road underneath it.
Elevation models can also describe glacier surfaces, submerged terrain, or planetary topography. Underwater elevation is commonly discussed as bathymetry, often expressed as depth below a reference surface rather than height above it. A continuous coastal model must reconcile these conventions before joining land and seabed measurements. Elevation may additionally refer to the top of snow or ice rather than bedrock, making the identity and date of the modeled surface essential parts of its meaning.
Data structures, sampling, and spatial resolution
The most widespread DEM structure is a rectangular grid of raster cells. Each cell stores an elevation, together with a spatial reference that locates the grid on Earth. Regular grids are straightforward to store, display, and process because neighboring values can be found by row and column. They also support efficient numerical operations, such as calculating local slopes or accumulating flow across a landscape.
A grid value does not always have the same physical interpretation. Depending on the product, it may represent an interpolated elevation at a cell center, a sampled point on a regular lattice, or a statistic describing an area. The distinction between pixel-as-point and pixel-as-area conventions affects grid registration. Confusing these conventions can introduce a half-cell horizontal displacement even when coordinate systems and nominal resolutions agree. Missing observations are usually represented by a designated no-data value or a separate validity mask; they must not be interpreted as zero elevation.
An alternative is a triangulated irregular network (TIN), which divides a surface into non-overlapping triangles whose vertices have known horizontal coordinates and heights. Within a triangle, elevation is commonly estimated by linear interpolation. A TIN can place many vertices along ridges, channels, or abrupt slope changes while using fewer points on smooth terrain. Breaklines constrain triangle edges so that interpolation respects features such as the crest of an embankment or the edge of a channel. Unlike a raster, a TIN adapts its sampling density to surface complexity, although its connectivity requires additional storage and processing.
A point cloud contains discrete three-dimensional observations without necessarily defining a continuous surface. Millions or billions of laser measurements may form the input to a terrain model, but classification and interpolation are still required to convert those observations into a bare-earth grid or mesh. A contour dataset is another representation of elevation, although contours record selected height levels rather than providing an explicit height at every location.
Spatial resolution describes the scale of sampling, not a guarantee of accuracy or detail. A 1-m grid created by resampling a 30-m grid still contains essentially the terrain information of the coarser source. Conversely, a densely measured surface may lose small channels or narrow ridges when aggregated to a coarse grid. Grid orientation, interpolation, and smoothing all influence the features retained.
Resolution expressed in angular units also changes its ground dimensions with location. One arc-second of latitude is approximately 31 m, while one arc-second of longitude is approximately m at latitude , with additional variation associated with Earth’s ellipsoidal shape. A one-arc-second grid therefore does not consist of equal-sized ground squares worldwide.
Coordinate systems and height references
A DEM requires both a horizontal location framework and a vertical reference. The horizontal coordinate reference system specifies how positions relate to Earth, commonly through geographic coordinates or a projected coordinate system. A map projection converts the curved reference surface into plane coordinates, introducing distortions whose importance depends on location and the operation being performed. Distances, cell areas, and terrain derivatives cannot always be calculated correctly by treating angular coordinates as ordinary Cartesian coordinates.
The vertical reference determines what “zero elevation” means. Satellite positioning commonly provides ellipsoidal height, , measured relative to a mathematical reference ellipsoid. Many national mapping systems instead use heights tied to a gravity-related surface and a specified vertical datum. For orthometric heights, the reference is associated with the geoid, an equipotential surface of Earth’s gravity field that approximates mean sea level in a physical sense. The geoid is not a simple ellipsoid and does not coincide with the instantaneous ocean surface.
A commonly used relationship is
where is orthometric height and is geoid undulation relative to the reference ellipsoid. This relationship is useful only when the quantities are consistent in their reference frame and conventions. Some countries use normal heights and a quasigeoid rather than orthometric heights and a geoid. Transformations between vertical systems may therefore require more than subtracting a single constant offset.
For example, a laser survey positioned using satellite navigation may initially contain ellipsoidal heights, whereas an existing engineering terrain model may use a national height datum. Merging them without transformation can create an apparent step across the survey boundary. The offset may vary spatially, so correcting one control point does not necessarily align the entire region. Units introduce another possible mismatch: elevations stored in feet and horizontal coordinates stored in meters produce incorrect slopes unless explicitly reconciled.
Time can be part of the reference problem. In areas of tectonic motion, land subsidence, or glacial isostatic adjustment, coordinates and elevations can change measurably between survey epochs. A dataset’s acquisition date and reference-frame epoch are distinct pieces of information: one describes when the surface was observed, while the other helps define the coordinates.
These distinctions are especially important in coastal applications. Tidal datums, such as mean lower low water, are not interchangeable with national land-height datums. Evaluating inundation or sea-level rise requires land elevations and water levels expressed in compatible vertical references; otherwise, a datum discrepancy can be mistaken for a physical difference between land and water.
Measurement technologies
Elevation models are produced through ground surveying, airborne observations, and satellite remote sensing. Each technique measures a different physical signal, and that signal determines which surface is detected. Surveying can provide highly controlled elevations along roads, construction sites, or reference networks, but collecting dense measurements over large regions is expensive. Remote sensing trades some local control for much broader coverage.
Photogrammetry reconstructs three-dimensional geometry from overlapping images taken from different viewpoints. Corresponding image features are located, and their positions are calculated through triangulation using camera geometry and orientation. Aerial stereo photography supported terrain mapping long before digital cameras became widespread. Modern automated image matching produces dense surface measurements from aircraft and satellites. Structure-from-motion methods estimate camera positions and scene geometry jointly from multiple overlapping images, often followed by dense reconstruction. Ground control or accurately positioned cameras establish geographic scale and location. Cloud, shadow, repetitive texture, moving vegetation, and featureless snow can hinder matching.
Lidar measures distance using laser light. Airborne topographic systems generally combine laser ranging with satellite positioning and an inertial measurement unit to determine the location of each return. Some pulses yield multiple returns from canopy layers and the ground, allowing terrain to be reconstructed beneath vegetation where sufficient light reaches the surface. Dense canopy still reduces ground sampling, and ground elevations between returns must be estimated. Water often produces unreliable returns for conventional near-infrared topographic lidar; specialized green-wavelength bathymetric systems can measure shallow submerged surfaces where water clarity and other conditions permit.
Interferometric synthetic aperture radar derives elevation from phase differences between radar observations made from separated antenna positions. The geometry of this baseline makes the phase difference sensitive to surface height. Processing must account for imaging geometry, phase ambiguities, and other signal contributions. Depending on wavelength, vegetation, and moisture, the effective radar scattering surface may lie within the canopy rather than at the ground or canopy top. Radar-derived elevation is therefore not automatically equivalent to optical surface elevation or classified lidar terrain.
Radar can acquire data through clouds and without sunlight, but steep terrain causes characteristic problems. Foreshortening compresses slopes facing the sensor; layover can reverse the apparent order of elevated and lower features; radar shadow leaves some slopes unobserved. Optical stereo has its own visibility limitations, especially on steep, shaded, or snow-covered slopes. Combining acquisition directions or measurement technologies can reduce gaps, but it also creates a need to reconcile different surface responses.
Satellite and airborne observations may be supplemented by satellite altimetry, sonar, or field measurements. Altimeters provide precise observations along tracks rather than uniform wall-to-wall coverage. Their value may lie in controlling, validating, or updating a regional surface rather than directly producing a dense grid.
From observations to an elevation product
DEM production is a sequence of geometric, statistical, and interpretive operations. Raw measurements must first be positioned, calibrated, and screened. In an airborne laser survey, this includes reconstructing the aircraft trajectory, applying sensor offsets and orientation corrections, and checking agreement between overlapping flight strips. For imagery, camera calibration, image orientation, and correspondence matching precede surface reconstruction. A processing error at this stage can affect an entire block rather than a few isolated cells.
Producing a bare-earth model usually requires separating ground from nonground observations. Classification algorithms examine characteristics such as local height differences, slope, neighborhood geometry, and return attributes. A method that performs well on a flat agricultural plain may remove genuine terrain features in steep mountains or retain low vegetation in a marsh. Manual inspection and targeted correction remain important where automatic classifications confuse cliffs, retaining walls, dense shrubs, or buildings with the underlying ground.
Spatial interpolation converts scattered elevations into a continuous surface or regular grid. Nearest-neighbor interpolation preserves sampled values but can create discontinuities. Linear interpolation over a TIN produces planar facets. Inverse-distance weighting gives greater influence to nearby observations, while kriging estimates values using a model of spatial covariance. Splines produce smooth surfaces but may overshoot near abrupt changes. No method reconstructs unobserved terrain without assumptions; the appropriate method depends on point density, surface geometry, and intended use.
Breaklines and supplementary observations can constrain interpolation where ordinary proximity-based methods would be misleading. A channel bank, ridge crest, or seawall may need explicit representation because interpolation across it would smooth away a physically important boundary. Water bodies are often assigned controlled surface elevations through hydro-flattening, which removes spurious slopes or irregularities from lakes and represents rivers according to product-specific rules. Such editing improves consistency but means that the final grid contains modeled or imposed values as well as direct measurements.
Mosaicking combines overlapping surveys or tiles. This requires consistent horizontal and vertical references, comparable surface definitions, and treatment of seams. Differences in acquisition date can create real discontinuities where excavation, construction, flooding, or vegetation change occurred. Blending those boundaries may improve visual appearance while concealing the actual measurement history.
A finished product should preserve metadata describing acquisition dates, source sensors, resolution, height reference, processing methods, classification, no-data areas, and quality assessment. Formats such as GeoTIFF can store elevation grids together with georeferencing information, but a file extension alone does not establish that all necessary metadata are present. Distribution pyramids and compressed tiles improve access; they should not be confused with independent measurements at additional resolutions.
Accuracy, uncertainty, and validation
DEM quality has several dimensions: horizontal accuracy, vertical accuracy, effective resolution, completeness, and fidelity to the intended surface. Accuracy concerns agreement with an appropriate reference, whereas precision concerns the repeatability or dispersion of measurements. A surface can be internally smooth and highly repeatable while carrying a substantial systematic height offset.
Vertical validation commonly compares model elevations with independent checkpoints of higher accuracy. If the elevation residual at checkpoint is
the mean residual estimates bias, while the root mean square error is
RMSE combines systematic and random components and is sensitive to large residuals. Median error, percentile-based measures, and robust spread statistics can provide additional information, particularly where errors are skewed or contain outliers. Confidence statements based on a normal distribution are appropriate only when their assumptions are justified.
Checkpoint selection is consequential. Measurements concentrated on accessible roads or flat open ground may poorly characterize forested slopes, wetlands, or dense urban areas. A meaningful assessment distinguishes relevant land-cover and terrain classes and avoids reusing production control as if it were independent validation. Comparing two DEMs does not by itself reveal the absolute accuracy of either one, because both may contain error or share a common source.
Horizontal misalignment creates apparent vertical error on sloping terrain. For a small displacement, an approximation is
Thus, even a modest horizontal offset can generate large elevation differences on a steep mountainside. This is one reason co-registration is essential before interpreting differences between repeated surveys as erosion, glacier thinning, or excavation.
Errors also have spatial structure. Flight-strip offsets, radar artifacts, interpolation patterns, and mismatched source tiles can create correlated errors extending across many cells. Uncertainty in an average elevation or calculated volume therefore cannot generally be estimated by assuming that every cell provides an independent observation. Similarly, slope and curvature calculations can amplify short-wavelength noise even when elevation errors appear small.
Product specifications must be read at the scale of the intended analysis. A global model with errors of several meters may support regional watershed delineation but be unsuitable for identifying a low coastal embankment or designing a small drainage structure. Nominal cell spacing, visual sharpness, and the number of decimal places stored in a file are not substitutes for validation.
Terrain derivatives and hydrological structure
A DEM is often most useful as the basis for derived measurements. Local slope describes the rate of elevation change, while aspect gives the direction of steepest descent. If and in compatible horizontal and vertical units, slope angle can be expressed as
In a raster, derivatives are estimated from neighboring cells, so the result depends on the neighborhood size and numerical method. A coarse grid generally suppresses short steep slopes, whereas a very fine but noisy surface can exaggerate them. Aspect is undefined on a perfectly flat surface and unstable where the slope is very small.
Curvature describes changes in slope and can help distinguish convergent hollows, divergent ridges, and convex or concave profiles. Surface roughness, local relief, and topographic position measure different aspects of terrain form. Hillshading simulates illumination from a chosen light source to reveal landforms, but it is a visualization rather than a direct measurement of rock type or terrain condition. Changing the illumination direction can make some features conspicuous and others nearly invisible.
In hydrology, terrain grids support the identification of drainage basins, stream paths, and contributing areas. The D8 method routes each cell toward one of its eight neighbors, usually the one with the steepest downward gradient. Multiple-flow-direction methods divide flow among several downslope neighbors, while other approaches estimate a continuous direction from local facets. These alternatives can produce different channel networks and wetness patterns, particularly on broad slopes and in low-relief terrain.
Real and artificial depressions complicate routing. Some are measurement artifacts, but others are genuine features such as lakes, closed basins, or karst sinkholes. Depression filling raises low cells to create an outlet; breaching cuts a path through an obstruction. Both modify the original surface. A hydrologically conditioned DEM is therefore a purpose-specific representation, not necessarily a more accurate model of physical ground elevation.
Roads and bridges illustrate the distinction. A road embankment may block flow in a terrain grid even though a culvert carries water underneath it. Removing the bridge deck or inserting an appropriate drainage connection can restore realistic connectivity. Conversely, cutting through an actual levee creates a false pathway. Conditioning requires information about structures and drainage, not merely a rule that every cell must drain.
Flow accumulation estimates contributing area under the chosen routing assumptions; it does not directly measure river discharge. Simulating runoff also requires precipitation, infiltration, storage, and other processes. Likewise, a DEM-derived channel is a geometric inference, not proof of a perennial stream.
Applications in environmental science and engineering
In flood modeling, elevation defines the geometry through which water moves and the land elevations against which simulated water levels are compared. Channel depth, floodplain barriers, road crossings, and embankments can strongly influence the result. Airborne topographic data often do not measure submerged channel beds, so bathymetric surveys or other channel information may be required. A simple “bathtub” map marks land below a chosen water level, but it does not reproduce flow resistance, hydraulic connectivity, timing, or the dynamics of a storm surge.
In geomorphology, elevation models quantify landforms and landscape change. Repeated surveys can reveal riverbank retreat, landslide displacement, coastal erosion, mining, or sediment deposition. A DEM of difference subtracts one surface from another after alignment and datum reconciliation. For equal-area raster cells, a volume change can be estimated as
where is cell area and is elevation change. Interpretation requires accounting for measurement uncertainty, changing vegetation, gaps, and the distinction between surface displacement and actual material gain or loss.
For glaciers and ice sheets, repeated surface elevations support estimates of volume change and geodetic mass balance. Converting volume to mass requires assumptions or observations about density. Snow accumulation, firn compaction, and differences between radar penetration depths can alter measured surface elevation without representing equivalent changes in solid ice thickness. Acquisition season is consequently important.
Engineering applications include route selection, grading, earthwork estimation, reservoir geometry, and drainage design. A terrain model allows alternative alignments for a railway or road to be assessed before detailed field investigation. However, the required tolerances often exceed those of regional elevation products. Local surveys, structural details, and explicit breaklines may be necessary where narrow features control design.
A viewshed identifies locations theoretically visible from an observation point given an elevation surface and specified viewing conditions. Bare-earth terrain supports landscape-scale visibility analysis, while buildings and vegetation require a surface model or additional three-dimensional information. Applications include communications planning, visual-impact assessment, and the study of archaeological landscapes. Long-distance calculations may also need Earth curvature and atmospheric refraction.
Elevation influences ecosystems and atmospheric processes through temperature gradients, solar exposure, drainage, and barriers to airflow. Models of habitat distribution, soil erosion, and local climate therefore use terrain variables alongside other observations. For example, soil erosion assessments derive slope-related factors from DEMs, while weather models represent topography at their own computational scale. In each case, changing DEM resolution can change the explanatory variables as well as the apparent detail of the map.
Historical development and major elevation datasets
Digital terrain representation developed alongside electronic computing, analytical surveying, and automated engineering design during the mid-twentieth century. Charles L. Miller and Robert A. Laflamme’s 1958 work on the digital terrain model is an early landmark, particularly in the context of highway engineering. Subsequent national mapping programs digitized contours, photogrammetric measurements, and surveyed heights, making elevation available for automated analysis rather than solely for visual interpretation on a topographic map.
The expansion of geographic information systems made regular elevation grids increasingly practical. Early national products often reflected the scale and contour interval of their source maps, so neighboring datasets could differ in accuracy and production method. Later airborne laser scanning greatly increased local detail and enabled extensive bare-earth mapping beneath partial vegetation cover. In the United States, the US Geological Survey’s 3D Elevation Program, or 3DEP, supports systematic elevation acquisition and distribution, with lidar central to mapping in much of the country and interferometric radar used for Alaska.
A major advance in near-global coverage came from the Shuttle Radar Topography Mission (SRTM), flown aboard Space Shuttle Endeavour in February 2000. Its interferometric radar observations covered land between approximately north and south. SRTM became a widely used source for terrain analysis in regions lacking accessible national elevation data. One-arc-second data, often described approximately as 30-m postings, were progressively made available beyond the United States from 2014 onward. Voids, vegetation effects, and radar geometry remained important considerations despite the mission’s broad coverage.
The ASTER Global Digital Elevation Model, first released in 2009, used optical stereo imagery from the Advanced Spaceborne Thermal Emission and Reflection Radiometer aboard NASA’s Terra satellite. It extended coverage to high latitudes and offered one-arc-second postings, but cloud contamination and stereo-processing artifacts required careful assessment. The Japanese AW3D30 product provides another global source, derived from optical stereo observations acquired by the PRISM instrument on the ALOS satellite.
The German TanDEM-X mission, launched in 2010 to fly in formation with TerraSAR-X, used bistatic radar interferometry to produce a highly consistent global elevation model. The Copernicus DEM is derived from the WorldDEM product based on TanDEM-X observations, with editing and supplementary data. It is a digital surface model rather than a universally bare-earth terrain model. Its globally available GLO-30 and GLO-90 products have approximately 30-m and 90-m postings, respectively.
These products differ in acquisition period, sensor physics, editing, vertical reference, and licensing. A global grid may combine observations from many dates rather than depict a single moment. Selecting among products therefore involves matching their documented properties to the landscape and analytical task, rather than assuming that the newest release or smallest pixel always provides the most appropriate terrain representation.