Does the method always use the closest three points?

No. The natural-neighbor set depends on the Voronoi geometry and may contain a different number of samples.

Is the estimated value guaranteed to be physically correct?

No. It follows the interpolation rule; unmeasured variation can differ from that estimate.

Does it automatically work outside the sampled region?

Standard natural-neighbor interpolation is defined within the sample convex hull; software may provide a separate extrapolation choice.

AI Glossary / Computer vision and multimodal AI / Intermediate

Natural neighbor interpolation

Definition

Natural neighbor interpolation estimates values between scattered sample locations using weights derived from their spatial arrangement. In the standard Sibson construction, the weights come from how a new query point would divide the existing Voronoi regions. The method produces a locally weighted estimate while respecting the sample values and reproducing linear variation under its usual assumptions.

What it means

A scattered dataset does not lie on a neat rectangular grid. Measurements might come from irregular survey positions or samples selected at convenient locations. Natural neighbor interpolation uses the geometry of those positions to identify which samples influence a query.

Natural neighbors are not simply a fixed number of closest points. They are determined through the Voronoi arrangement around the query, and the number can vary. The method also differs from nearest-neighbor interpolation, which copies one sample value, and from inverse-distance weighting, which uses an explicitly chosen distance-based weighting rule.

How it works

Construct the Voronoi diagram of the sample positions, where each cell contains locations closest to one sample. Imagine inserting the query point. Its new Voronoi cell takes area from some existing cells. Those affected samples are its natural neighbors.

In two dimensions, each neighbor's weight is the fraction of the query cell's area taken from that neighbor's former cell. Normalize the weights and combine the corresponding sample values. The weights are nonnegative and sum to one in the ordinary interior construction. Boundary cases need appropriate treatment, and extrapolation outside the sample convex hull is a separate policy rather than an automatic consequence of the same bounded-cell formula.

Three practical examples

1. A mapping team converts irregularly located elevation samples into a continuous surface for visualization. The interpolation estimates unmeasured locations and does not add new surveyed detail.

2. An imaging experiment reconstructs a scalar field from scattered measurements. Natural-neighbor weights provide a geometry-based alternative to selecting an arbitrary fixed neighborhood size.

3. A design tool interpolates values attached to control points, such as a smooth intensity field. The distribution of controls influences where detail can be represented reliably.

Think of it this way

Think of adding a new service center to a map divided into nearest-center territories. The new center takes portions of territory from nearby existing centers. The share taken from each can determine that center's influence on the new estimate. This analogy explains the geometry of the weights; it does not mean the estimated quantity actually spreads according to service-area boundaries.

Worked example

Suppose inserting a query point creates a Voronoi cell with total area 10 square units. It takes areas 2, 3, and 5 from three existing sample cells. The corresponding natural-neighbor weights are 0.2, 0.3, and 0.5.

If the measured values at those samples are 10, 20, and 30, the interpolated value is 0.2 × 10 + 0.3 × 20 + 0.5 × 30 = 23. Because the weights are nonnegative and sum to one, the result lies between the smallest and largest contributing values.

These areas are illustrative inputs to the calculation, not a Voronoi diagram computed from unspecified coordinates. In practice, the sample geometry determines them. If all three measured values were 7, the estimate would also be 7, demonstrating preservation of a constant field. Preserving these mathematical properties does not establish that 23 is the true unmeasured physical value.

Illustrative case study

Fictional case: A landscape-design team receives elevation measurements concentrated near accessible paths. A natural-neighbor surface looks smooth across the site, but a sparsely sampled drainage channel appears shallower than field observations suggest.

The team overlays the measurement locations on the surface and identifies where interpolation spans large gaps. It collects additional points around the channel and marks unsupported areas in the visualization. The interpolation is useful for representing the available samples, while the project avoids interpreting smoothness as evidence of dense measurement coverage.

Common misconceptions and mistakes

Natural-neighbor interpolation is not nearest-neighbor copying. Its weights can involve several samples and vary smoothly with the query, subject to the method's properties. It is also not a trained AI model simply because it estimates missing values.

Do not assume it discovers sharp discontinuities absent from the samples. A continuous interpolation can bridge a real boundary unless the model or data handling accounts for it. Duplicate positions with conflicting values require resolution. Rescaling coordinate axes can change the geometry and therefore the neighbors and weights, so units and anisotropy should be chosen deliberately.

Business use

This method is useful for turning irregular measurements into maps, grids, or fields without selecting a global polynomial form. Its geometric weighting can be easier to explain than an opaque fitted model, particularly when the measured locations are shown alongside the result.

Assess sample coverage and hold out some measured points for validation where appropriate. Compare interpolation error in dense and sparse regions rather than relying on a single average. State the extrapolation policy and avoid displaying unsupported regions as equally reliable. For physical measurements, keep coordinate units, calibration, and acquisition context linked to the interpolated output.

Check your understanding

Normalized natural-neighbor weights are 0.25 and 0.75, with sample values 8 and 12. What is the weighted estimate?

Answer: 0.25 × 8 + 0.75 × 12 = 11. The estimate remains within the contributing value range, but its physical accuracy still depends on the samples and field.

Related terms

Delaunay triangulation
Convex hull
Voronoi diagram
Sibson coordinates
scattered data
nearest-neighbor interpolation
inverse-distance weighting
extrapolation

Learn next

Plot irregular sample locations and interpolate the same values using nearest, linear, and natural-neighbor methods. Inspect the results near data points, across large gaps, and at the convex hull. Add one new measurement and observe how the local geometry changes the natural-neighbor estimate.

Convex hull

Sources

Primary implementation references: CGAL, “2D and Surface Function Interpolation,” https://doc.cgal.org/latest/Interpolation/index.html ; MathWorks, “Interpolating Scattered Data,” https://www.mathworks.com/help/matlab/math/interpolating-scattered-data.html . Relevant Voronoi-weight and interpolation-property descriptions were reviewed. The cell-area calculation and landscape case are original examples; the stated areas are illustrative rather than computed from a hidden dataset.

Previous: Natural-language user interface Next: Natural sort order