Spatial interpolation is the estimation of a variable at locations where it has not been observed, using measurements from other locations and assumptions about how that variable varies across space. It is a fundamental operation in spatial analysis, geographic information systems (GIS), and geostatistics. Examples include estimating rainfall between weather stations, constructing elevation surfaces from surveyed points, and mapping groundwater levels between wells. The resulting surface is a model of spatial variation, not a set of additional observations.
Principles and mathematical formulation
Suppose measurements are available at locations . The objective is to estimate the value at an unsampled location , written as .
Many interpolation methods take the form
where the weights depend on the prediction location, the observations used, and the chosen method. Weights often sum to one, allowing the estimator to reproduce a constant surface, but this condition is not universal. Some methods include an additional trend model or use nonlinear relationships.
A common justification is spatial autocorrelation: nearby observations often resemble one another more than distant observations do. However, proximity alone does not guarantee similarity. A mountain ridge, coastline, geological fault, or abrupt land-use boundary may separate locations with substantially different conditions. Spatial dependence may also be stronger in one direction than another, a property called anisotropy.
Interpolation and extrapolation
Interpolation generally concerns estimation within a region supported by observations. Extrapolation extends predictions beyond that support. For scattered points, the convex hull—the smallest convex region containing all sample locations—is a useful geometric reference, but it is not a sufficient measure of support. A prediction inside the hull can still be poorly constrained if it lies within a large sampling gap or across a physical barrier.
The distinction matters because a method that behaves well among observations may become unstable or unreliable beyond their geographic or environmental range.
Exact and smoothing methods
An exact interpolator reproduces the observed values at their locations. A smoothing interpolator may depart from individual measurements to represent an underlying spatial process or reduce the influence of measurement noise.
Exactness is not necessarily evidence of greater accuracy. If measurements contain error, forcing a surface through every observation can reproduce noise rather than the underlying field. Whether a method is exact can depend on its settings and on whether the prediction target is a noisy observation or a latent, noise-free process.
Main methods
Nearest-neighbor interpolation
Nearest-neighbor interpolation assigns each prediction location the value of its closest observation. It produces piecewise-constant regions separated by abrupt boundaries.
The method is simple, computationally inexpensive, and does not create new numerical values. It can be appropriate for categorical labels or for applications requiring preservation of original values. However, abrupt changes between regions are usually unrealistic for smoothly varying physical quantities. For categorical phenomena, proximity-based assignment is a classification rule rather than evidence that the true boundary lies midway between observations.
Inverse distance weighting
Inverse distance weighting (IDW) estimates values as a normalized distance-weighted average:
where is the distance from the prediction location to observation , is a power parameter, and is the selected neighborhood. At an observation location, an exact implementation returns the observed value rather than evaluating the singular expression.
Larger powers concentrate influence on nearby observations; smaller powers spread influence more broadly. Results also depend on search radius, neighbor count, and sampling geometry.
Because standard IDW uses nonnegative normalized weights, its estimates remain within the range of the contributing observations. It is straightforward to apply but can produce circular “bull’s-eye” patterns around isolated samples and does not inherently provide a probabilistic prediction uncertainty.
Triangulation and natural-neighbor interpolation
A triangulated irregular network (TIN) partitions sampled locations into nonoverlapping triangles. Linear interpolation within each triangle uses its vertex values to define a plane. The resulting surface is continuous across triangle edges, but its slope generally changes abruptly at those edges.
Natural-neighbor interpolation derives local weights from changes in the Voronoi tessellation when a prediction point is inserted. Standard natural-neighbor estimates use nonnegative weights, preserve observed values, and generally produce smoother transitions than piecewise-linear triangulation. Boundary handling requires attention, particularly outside the convex hull.
Both approaches are common in terrain modeling and other applications involving irregularly distributed points.
Splines and radial basis functions
Spline interpolation fits a surface with specified smoothness properties. Thin-plate splines, for example, balance agreement with observations against a penalty related to surface bending. Interpolating splines pass through the observations, while smoothing splines allow departures from them.
Radial basis function methods represent a surface using functions centered on sample locations, often together with polynomial terms. Particular choices overlap mathematically with spline methods.
These approaches can represent smoothly varying fields effectively, but some configurations produce values above or below the observed range. Such overshoot can be problematic for bounded quantities or terrain features with sharp discontinuities. Smoothing and shape parameters strongly affect the result.
Trend surfaces and regression-based methods
A trend surface represents broad spatial variation using a mathematical function of coordinates, often a polynomial. Regression-based prediction can additionally use explanatory variables such as elevation, distance from the coast, or land cover.
These methods are useful when geography influences the target through identifiable environmental relationships. For example, elevation may explain part of the spatial variation in air temperature.
A common hybrid strategy fits a regression model, interpolates its residuals, and adds the two predictions. Regression kriging is one such approach. Although often grouped with spatial interpolation in applied mapping, regression-based spatial prediction is broader than interpolation based solely on neighboring target observations.
Kriging
Kriging derives prediction weights from a model of spatial dependence and a specified mean structure. Ordinary kriging assumes an unknown constant mean over the relevant modeling domain or neighborhood; universal kriging incorporates a specified spatial trend.
Spatial dependence is often described by a variogram. Under an intrinsic stationarity assumption, the semivariogram is
where is the separation vector. Common model characteristics include a nugget, associated with measurement error or variation below the sampling scale; a sill, representing a limiting semivariance in models with a plateau; and a range describing the distance over which dependence persists.
Under its assumptions, kriging provides a best linear unbiased predictor and a model-based prediction variance. This does not make it universally superior to other methods: performance depends on the sampling design, mean model, spatial covariance model, and treatment of measurement error. Kriging weights can be negative, so predictions need not remain within the observed range.
The prediction variance is conditional on the fitted model and typically does not fully account for uncertainty in estimated model parameters or unrecognized model misspecification.
Historical development
Spatial interpolation developed through overlapping traditions in surveying, cartography, numerical analysis, and applied statistics. Estimating elevations between surveyed points and drawing contour lines are longstanding cartographic tasks.
During the twentieth century, digital computing made systematic surface fitting and large-scale numerical interpolation increasingly practical. Geostatistics developed particularly through mining and ore-reserve estimation, providing methods that explicitly modeled spatial dependence. Kriging subsequently became widely used in environmental and earth sciences.
The growth of GIS, digital terrain modeling, and automated monitoring expanded the use of interpolation. More recent systems combine spatial statistical models with large environmental datasets, scalable computation, and machine learning methods.
Data preparation and model design
Coordinates and distance
Interpolation requires an appropriate coordinate reference system. Treating longitude and latitude as ordinary planar coordinates can distort distances, especially over large regions or at high latitudes. Depending on geographic extent, suitable choices include an appropriate projected system, geodesic distances, or a model defined on a sphere.
Distance may also need a process-specific interpretation. Along-stream distance can be more meaningful than straight-line distance for some river variables, while coastlines or geological boundaries may limit connectivity.
Sampling geometry and data quality
Dense, well-distributed observations generally support finer spatial detail than sparse or clustered observations. Clusters can exert excessive influence in methods that do not adequately account for sampling geometry. Outliers, coordinate errors, duplicate locations, inconsistent units, and measurements collected at different times can all distort a surface.
Neighborhood design is therefore part of the model. Search radii, maximum neighbor counts, and directional sectors influence which observations contribute to each prediction. Very small neighborhoods may produce unstable local variation; excessively large neighborhoods may obscure local structure.
Spatial support and resolution
A measurement represents a spatial support: a point, soil core, plot, grid cell, or administrative area. Predicting a point value is not the same task as predicting an average over a region. Block kriging, for example, explicitly estimates averages over specified areas or volumes.
Producing a fine raster grid does not create information at that resolution. A one-meter output grid derived from widely spaced observations can give a misleading impression of detail.
Interpolation from aggregated areas also raises different questions from point interpolation. Redistributing population counts among zones, for example, requires areal interpolation and often mass-preserving constraints. These issues relate to the modifiable areal unit problem.
Validation and uncertainty
A plausible-looking map is not sufficient evidence of predictive accuracy. Validation compares predictions with observations withheld from model fitting or collected independently.
Common measures include mean error, which indicates average signed error, and root mean square error:
Mean absolute error is less sensitive than RMSE to large individual errors. Metrics describe different aspects of performance and should be interpreted in relation to the variable’s units and intended use.
Cross-validation must account for spatial structure. Randomly withholding points can produce optimistic results when test observations lie very close to training observations. Spatially blocked validation, separation buffers, or region-based holdouts can better represent prediction into spatial gaps or new regions. The validation design should match the intended prediction task; no single partitioning strategy is appropriate for every application.
Uncertainty arises from several sources:
- measurement error and unresolved small-scale variation;
- sparse or uneven spatial coverage;
- uncertainty in fitted trends and dependence parameters;
- incorrect assumptions about stationarity, barriers, or directional dependence;
- temporal differences among measurements.
Probabilistic predictions can be evaluated using interval coverage and other calibration diagnostics. Methods without an intrinsic uncertainty model can use additional procedures, such as spatial resampling or ensembles, but these procedures also require assumptions. Measurement uncertainty and interpolation uncertainty are related but distinct.
Applications
Spatial interpolation is used in several major fields:
- Terrain modeling: constructing digital elevation models from surveyed elevations or sampled point clouds.
- Meteorology and climatology: mapping precipitation, air temperature, and other variables between observing stations, sometimes with elevation and broader climate fields as predictors.
- Hydrology: estimating groundwater levels and other spatially distributed water-system properties.
- Soil and environmental mapping: representing soil characteristics, salinity, and pollutant concentrations from field measurements.
- Mineral-resource estimation: predicting grades and quantities within geological domains, commonly with explicit treatment of spatial support.
- Ocean and atmospheric observations: reconstructing fields from irregular sampling locations while accounting for depth, time, and physical boundaries.
Interpolation is also used when resampling an existing grid during reprojection or image transformation. In that setting, the source values already occupy a regular grid, and the task differs from constructing a field from scattered observations. Nearest-neighbor, bilinear, and cubic resampling are common examples.
Limitations and interpretation
No interpolation method is best for every variable, sampling design, and prediction objective. Smooth fields, abrupt boundaries, directional processes, and highly skewed measurements may require different models.
A smooth surface can conceal substantial uncertainty, while an exact surface can faithfully reproduce erroneous observations. Deterministic methods do not eliminate uncertainty merely because they do not report it. Conversely, model-based confidence or prediction intervals do not guarantee that all important sources of error have been represented.
Interpolation is primarily predictive rather than causal. A spatial association between a variable and elevation or land cover does not, by itself, establish causation. Spatial prediction also cannot reliably recover features that the observations and auxiliary information do not resolve. Interpreting an interpolated map therefore requires attention to sampling coverage, model assumptions, validation results, and the distinction between estimated and observed values.