Georeferencing is the process of establishing a relationship between data and locations in a coordinate reference system (CRS). It allows a scanned map, aerial photograph, satellite image, survey drawing, or other spatial record to be positioned relative to the Earth. In a geographic information system (GIS), this relationship makes it possible to overlay datasets, measure distances and areas, and associate observations with geographic features. Georeferencing may involve recording an existing coordinate relationship, estimating one from known locations, or deriving it from measurements of a sensor’s position and orientation.
Coordinate foundations
An image initially has an internal coordinate system, usually expressed as pixel columns and rows. These coordinates locate information within the image but do not, by themselves, identify positions on Earth. Georeferencing supplies a mapping from these internal coordinates to geographic or projected coordinates.
Geographic coordinates commonly use latitude and longitude. Projected coordinates use a map projection to represent positions on a plane, often in metres. A complete spatial reference also depends on a geodetic datum, which defines how coordinates relate to the Earth. Coordinate units, axis order, and, where relevant, coordinate epoch and vertical reference must also be understood.
Georeferencing is closely related to several other operations, but they are not interchangeable:
- Assigning a CRS identifies the reference system in which existing coordinate values are to be interpreted. It does not change those values or establish a missing pixel-to-ground relationship.
- Reprojection transforms already referenced data into another CRS. It assumes that the original reference information is sufficiently known.
- Image registration aligns one image with another. Registration becomes geographic referencing when the resulting alignment is tied to an Earth-referenced coordinate system.
- Orthorectification corrects image displacement caused by terrain relief and viewing geometry, producing an image with map-like geometry.
A dataset can therefore have a CRS label yet remain incorrectly positioned, or have a valid mathematical relationship between pixels and coordinates but lack enough CRS information for unambiguous interpretation.
Methods of establishing position
Control-point georeferencing
A common method uses ground control points: locations identifiable both in the source dataset and in a trusted geographic reference. Examples include road intersections, survey markers, and corners of stable structures. Their reference coordinates may come from surveying, a global navigation satellite system, an accurately positioned image, or an existing map.
Each point pairs a source position with a target coordinate. A transformation is fitted to these pairs and then used to calculate positions throughout the source dataset.
Control points should be distributed across the area of interest rather than clustered in one location. Their configuration matters as well as their number: collinear points, for example, cannot determine a general two-dimensional affine transformation. Stable, sharply identifiable features are preferable to features such as shorelines or vegetation boundaries that may change or be difficult to locate precisely.
Direct and sensor-based georeferencing
Images can also be positioned using measurements of the imaging system itself. Aircraft and satellite systems may combine navigation observations, inertial measurements, camera calibration, and acquisition timing to establish the relationship between image coordinates and the ground.
This approach is commonly called direct georeferencing. Its accuracy depends on sensor calibration, synchronization, position and orientation measurements, and the geometric model used. Ground control may still be used to correct systematic offsets or validate the result.
Some datasets already contain sufficient referencing information. A digital map exported with a known CRS and coordinate transformation may require no manual control-point selection.
Transformation models
The appropriate transformation depends on the geometry of the source and the distortions that must be represented.
Affine transformation
A two-dimensional affine transformation is widely used for scanned maps and other approximately planar sources. It accommodates translation, rotation, unequal scaling, and shear while preserving straight lines and parallelism.
For source coordinates and target coordinates ,
The six coefficients can be determined from three non-collinear control-point pairs. With additional points, they are generally estimated by least squares, allowing discrepancies among observations to be assessed.
For a north-up raster without rotation, the relationship simplifies to a geographic origin and pixel spacing. The distinction between pixel corners and pixel centres is important: confusing the two can introduce a half-pixel displacement.
Projective and polynomial transformations
A projective transformation can represent perspective effects when the photographed subject is effectively planar. It preserves straight lines but not necessarily parallelism, and generally requires at least four control-point pairs in a suitable configuration.
Higher-order polynomial transformations can model more complex deformation, such as distortion in a scanned paper map. A complete second-order transformation has six coefficients for each output coordinate and requires at least six suitably arranged control points. Additional points provide redundancy.
Increasing model complexity does not necessarily improve geographic accuracy. A flexible transformation may fit control points closely while introducing unrealistic distortion between them or unstable behavior beyond their extent.
Local warping and physical models
Local methods, including thin-plate splines and piecewise transformations, allow different parts of a dataset to deform differently. They are useful for irregularly distorted documents or local alignment, but may alter shapes, distances, and areas.
Aerial and satellite imagery may instead require physical camera models or sensor-specific mathematical models. Orthorectification commonly incorporates a digital elevation model to account for terrain-related displacement. A simple two-dimensional warp is not generally an adequate substitute where relief and viewing geometry cause substantial distortion.
Raster transformation and resampling
Georeferencing does not always require rewriting image pixels. Software may retain the original raster data and store the information needed to position them. Alternatively, it may transform the source onto a new, regularly spaced output grid, a process often called warping or rectification.
Because output pixel centres seldom coincide exactly with source pixel centres, generating a new raster requires resampling:
- Nearest-neighbour resampling selects a nearby source pixel. It preserves original class codes and is commonly used for categorical data.
- Bilinear resampling combines four neighboring values and is commonly used for continuous data.
- Cubic methods use a larger neighborhood to produce smoother results, but some can create values outside the local input range.
The output pixel size determines sampling density, not the accuracy or information content of the source. Resampling a coarse or poorly positioned image to smaller pixels does not create finer observations or improve its underlying positional accuracy. Repeated resampling can also modify values and soften detail.
Accuracy and uncertainty
Georeferencing quality depends on source quality, reference accuracy, control-point identification, transformation choice, and spatial coverage. Relevant sources of measurement uncertainty include:
- Errors in the reference map or survey.
- Ambiguous identification of corresponding features.
- Paper shrinkage, folds, and scanner distortion.
- Terrain displacement and unmodeled sensor geometry.
- Incorrect datum, projection, units, or axis order.
- Changes in the landscape between acquisition dates.
A common diagnostic is the root-mean-square error of control-point residuals. For points, a combined two-dimensional form is
where are reference coordinates and are transformed source positions. Software may instead report separate coordinate errors, pixel-based residuals, or statistics with different normalization conventions.
Small fitting residuals do not by themselves establish high positional accuracy. A flexible model can pass exactly through inaccurate control points, and errors shared by all reference observations may remain undetected. Independent checkpoints, excluded from model fitting, provide a stronger assessment of external accuracy.
Errors may also vary spatially. Areas surrounded by reliable control points are generally better constrained than areas far outside them. A single overall error statistic can conceal local misalignment.
Storage and interoperability
Georeferencing information may be embedded in a file or stored separately. GeoTIFF can store coordinate-system information together with raster positioning information. A world file stores six parameters describing an affine pixel-to-coordinate relationship, but does not by itself identify the CRS.
Other datasets may retain control points or sensor-model coefficients instead of a single affine relationship. GIS software can use these descriptions to position or transform the data.
Reproducible processing also depends on metadata recording the source CRS, reference dataset, transformation method, control-point provenance, output resolution, resampling method, and accuracy assessment.
Development and applications
Georeferencing draws on surveying, cartography, and photogrammetry. Computer-based mapping extended these practices to scanned documents and digital imagery, while satellite navigation and digital sensors enabled increasingly automated positioning.
Applications include integrating historical maps with modern geography, comparing imagery from different dates, aligning environmental observations, and preparing datasets for spatial analysis. In archaeology and cultural heritage work, it can connect excavation plans and archival records with present-day locations.
For historical maps, alignment does not remove uncertainty in the original document. Some maps prioritize schematic relationships, navigation, or symbolic importance rather than uniform geometric accuracy. Forcing such a map into close agreement with a modern reference can obscure its original structure. Its georeferenced position must therefore be distinguished from the accuracy and meaning of the source representation.