Geospatial raster data are geographically referenced datasets that represent information through an ordered grid of cells, usually arranged in rows and columns. Each cell contains one or more values associated with a location or area on Earth. They combine the array structure of raster data with information describing where the grid is located and how its values should be interpreted. Common examples include satellite images, elevation grids, land-cover classifications, and gridded weather observations. Raster datasets are a major component of a geographic information system (GIS), alongside vector geographic data, which represent features using points, lines, and polygons.
Structure and geographic reference
A basic raster consists of a rectangular array with a specified number of rows and columns. Cells are commonly called pixels, particularly in imagery. A single-band raster stores one value per cell; a multiband raster stores several corresponding arrays. In a color image, three bands may represent red, green, and blue intensities. In scientific datasets, bands or additional dimensions may represent wavelengths, dates, depths, or different variables.
The rectangular array does not necessarily mean that the actual geographic coverage is rectangular. Cells outside a study boundary, beyond a sensor’s valid coverage, or within gaps can be marked as invalid.
Positioning the grid
Georeferencing establishes the relationship between array positions and geographic locations. A coordinate reference system (CRS) supplies the framework in which coordinates are interpreted, including the relevant geodetic datum and, for projected coordinates, a map projection.
For a regularly spaced grid, an affine transformation commonly relates column and row coordinates to map coordinates:
where and are column and row coordinates, and locate the grid origin, and the remaining coefficients describe cell spacing, rotation, and shear.
For an unrotated, north-up raster, the rotation terms and are zero. If rows increase downward while map northing increases upward, the vertical coefficient is negative.
A critical convention is whether integer array coordinates refer to cell corners or cell centers. If the origin is the upper-left corner, the center of the first cell is obtained by substituting and . Confusing these conventions can shift a dataset by half a cell.
Not all geospatial rasters use a single affine transformation. Raw satellite imagery, aerial photographs, and some scientific grids may require control points, sensor models, or coordinate arrays to establish geographic position.
Cell size, extent, and alignment
A raster’s spatial extent is its geographic footprint. Its cell size describes the spacing or dimensions of cells in coordinate units. Two rasters with the same CRS and cell size may still be misaligned if their origins differ.
A grid expressed in latitude and longitude has angular cell dimensions. Equal angular cells do not have equal ground areas: their east–west dimensions generally decrease toward the poles. Projected grids also require attention to spatial distortion; a constant cell size in map units does not automatically imply an identical true ground area everywhere.
What cell values represent
Raster values require interpretation beyond their numeric storage type. A value may describe:
- A category, such as forest, cropland, or urban land.
- A continuous quantity, such as elevation or temperature.
- A count or total, such as the estimated population within a cell.
- A proportion or density, such as canopy cover or population per unit area.
- An image measurement, such as recorded digital intensity or calibrated reflectance.
- A probability or quality indicator, such as the probability of a land-cover class or a cloud flag.
A digital elevation model, for example, represents elevation, but its interpretation also depends on whether it describes the ground surface, includes buildings and vegetation, or uses a particular vertical reference.
The distinction between point samples and area-based values is important. A cell value may approximate a quantity at the cell center, an average across the cell, or a total for its footprint. These meanings affect how values can be combined or transferred to another grid.
Continuous phenomena and continuous numeric storage are not the same thing. Elevation can be stored as integers, while categorical class codes may be stored in floating-point arrays.
Missing values and masks
Cells without valid observations are commonly identified using a special NoData value, a validity mask, or a floating-point NaN. Missing data are not equivalent to zero. Zero rainfall, for example, is a valid observation, whereas missing rainfall indicates that no usable value is available.
Additional masks may distinguish clouds, shadows, sensor saturation, water, or low-quality retrievals. Because these conditions have different meanings, collapsing them into a single missing-data category can discard useful information.
Resolution and accuracy
Spatial resolution is often described using cell size, but cell spacing alone does not determine how much independent spatial detail a raster contains. A sensor’s response, preprocessing, interpolation, and aggregation also influence effective resolution.
A raster resampled from 1-kilometre cells to 10-metre cells has more cells but does not thereby acquire genuine 10-metre information. Similarly, coordinate precision and fine numeric increments do not establish positional or measurement accuracy.
Other relevant dimensions include:
- Spectral resolution: the wavelength ranges and widths represented by image bands.
- Temporal resolution: the spacing of observations or the period summarized by a value.
- Radiometric resolution: the quantization levels available for recording a signal.
Accuracy concerns the agreement of positions or values with an appropriate reference. Measurement uncertainty may reflect sensor noise, geolocation errors, retrieval assumptions, sampling limitations, and model error. A nominally fine-resolution product can still have substantial uncertainty.
Origins and development
The raster approach developed through several overlapping traditions: digital image processing, gridded numerical modeling, scanned cartography, and computerized spatial analysis. Early raster GIS methods demonstrated how geographic layers could be represented as arrays and combined through arithmetic and neighborhood operations.
Earth-observation satellite programs, including Landsat from the 1970s onward, greatly expanded the availability of repeated digital imagery. Increasing computing capacity subsequently enabled larger elevation models, continental land-cover products, and global environmental datasets.
More recently, network-accessible archives and cloud computing have encouraged processing data near their storage locations. Tiled storage, multiresolution representations, and multidimensional data structures make it possible to work with portions of datasets too large to load into memory at once.
Processing and spatial analysis
Raster structure supports efficient array-based computation. Many operations can be expressed as calculations over corresponding cells or moving neighborhoods.
Local, neighborhood, and zonal operations
Local operations calculate an output from values at the same location, such as subtracting temperatures from two dates or combining spectral bands. The systematic combination of raster layers is often called map algebra.
Neighborhood operations use surrounding cells. Examples include smoothing filters, edge detection, terrain slope, and local variability. Results depend on neighborhood size, shape, weighting, and treatment of missing cells.
Zonal operations summarize values within regions, such as calculating mean elevation within each drainage basin. Zones may come from another raster or from vector polygons. Decisions about cells intersected by polygon boundaries can influence the results.
Broader spatial analysis includes distance calculations, least-cost paths, viewsheds, and terrain-based flow routing. In hydrology, elevation rasters support the estimation of drainage directions and flow accumulation, although results depend on terrain quality and the treatment of depressions and flat areas.
Resampling and reprojection
Raster resampling estimates values on a new grid. It is required when changing cell size, shifting grid alignment, or transforming a raster into another CRS.
Common methods include:
| Method | Basic approach | Main considerations |
|---|---|---|
| Nearest neighbor | Selects the nearest input cell value | Preserves existing codes; commonly used for categorical rasters |
| Bilinear interpolation | Combines four neighboring values by distance-based weights | Smooths continuous data and changes original values |
| Cubic interpolation | Uses a larger neighborhood and cubic functions | Can preserve visual detail but may overshoot the input range |
| Aggregation | Summarizes contributing cells | The appropriate statistic depends on what values represent |
Reprojection transforms geographic positions and usually requires resampling. It changes the grid representation, not necessarily the underlying information content.
For downsampling, averages may be appropriate for area-mean quantities, while totals require methods that preserve the relevant sums. Categorical data may use a majority rule or retain class proportions. Simple averaging of class identifiers is generally meaningless. Repeated resampling can progressively alter values, boundaries, and spatial patterns.
Conversion between raster and vector
Rasterization converts vector features into grid values. The result depends on cell size and rules for deciding whether a feature occupies a cell. Narrow features may disappear if they are smaller than the grid spacing.
Vectorization extracts points, lines, or polygons from raster information. It can produce cell-shaped boundaries, fragmented polygons, or very large datasets. Neither conversion is inherently lossless.
Storage, formats, and distribution
Raster formats differ in their support for georeferencing, compression, metadata, multidimensional arrays, and partial access.
GeoTIFF stores geographic referencing information within the TIFF image framework and is widely used for imagery and gridded surfaces. Cloud Optimized GeoTIFF organizes tiles and reduced-resolution overviews to support efficient access to selected portions of a file over a network.
NetCDF is widely used for scientific datasets with dimensions such as time, height, latitude, and longitude. Other storage systems support similarly structured multidimensional arrays.
Large rasters commonly use:
- Tiling, which divides arrays into manageable blocks.
- Compression, which reduces storage requirements.
- Overviews, which store reduced-resolution representations for display or coarse-scale access.
- Chunking, which organizes multidimensional data into blocks suited to particular access patterns.
Lossless compression preserves encoded values, while lossy compression changes them. Whether such changes are acceptable depends on the intended analysis, not merely on visual appearance.
The Open Geospatial Consortium develops standards relevant to geographic data access and interoperability. A Web Map Service commonly supplies rendered map images; those images should not be assumed to preserve the original analytical values of the source raster.
Applications and limitations
Geospatial rasters are particularly suited to spatially extensive phenomena: elevation, vegetation, soil properties, atmospheric conditions, and ocean variables. They support land-cover mapping, agricultural monitoring, numerical weather prediction, disaster assessment, and land-use planning. Time-indexed rasters enable the study of seasonal variation and long-term environmental change.
Their regular structure also creates limitations. Fine grids can require substantial storage and computation, especially across many bands and dates. Grid boundaries approximate irregular geographic features, and a single cell may contain several land-cover types. Such mixed pixels complicate classification and interpretation.
Results can change with grid size and placement, an expression of issues related to the modifiable areal unit problem. Spatially adjacent values are often correlated, so the number of cells should not automatically be treated as the number of independent observations.
Meaningful comparison between datasets requires more than matching file dimensions. Coordinate reference systems, grid alignment, time periods, units, vertical references, valid-data masks, and the spatial meaning of values must also be compatible. Metadata describing these properties, together with processing history and quality information, are essential for interpreting a raster as geographic evidence rather than merely as an image.