Kriging Interpolation

The variogram is the mathematical engine of Kriging, quantifying spatial autocorrelation—the principle that points closer together are more similar than points farther apart.

• Nugget: Represents microscale variation or measurement error at zero distance.
• Sill: The plateau where the variogram stabilizes, indicating the total variance of the data.
• Range: The distance beyond which data points are no longer spatially correlated.

Fitting a theoretical model (Spherical, Exponential, Gaussian) to the empirical variogram dictates the weights assigned to neighbors during prediction.

Kriging is statistically optimal because it guarantees two critical properties:

• Unbiasedness: The expected value of the prediction error is zero. The model does not systematically over- or under-estimate.
• Minimum Variance: The prediction error variance is minimized, making it the most precise linear estimator possible.

Unlike simpler methods like Inverse Distance Weighting (IDW), Kriging weights are derived from the data's inherent spatial structure, not an arbitrary power function.

A unique advantage of Kriging over deterministic interpolators is the generation of a Kriging Variance map alongside the prediction map.

• This variance quantifies the uncertainty of the prediction at every unmeasured location.
• High variance indicates sparse sampling or high nugget effect.
• In Radio Environment Mapping (REM), this allows a cognitive radio to assess the reliability of a predicted spectrum hole before transmitting, directly informing risk-aware dynamic spectrum access decisions.

Kriging adapts to different statistical assumptions about the spatial field:

• Ordinary Kriging: Assumes an unknown but constant local mean. It is the most common variant, relying solely on the variogram and neighboring data.
• Universal Kriging: Accounts for a deterministic spatial trend (drift) across the region, such as a gradient caused by a large-scale path loss slope. It models the trend with a polynomial function and applies Kriging to the residuals.

For REMs, Universal Kriging is often superior when a strong propagation trend exists across the mapping area.

The prediction is computed by solving a system of linear equations that minimizes the error variance subject to an unbiasedness constraint.

• Weights (λ): Assigned to each known neighbor based on the variogram model.
• Lagrange Multiplier (μ): Enforces the constraint that all weights sum to 1.
• Matrix Inversion: The core computational cost involves inverting an N+1 x N+1 covariance matrix, where N is the number of neighboring points.

For real-time REM applications, local Kriging (using a moving search neighborhood) is preferred to keep the matrix size computationally tractable.

Co-Kriging extends the univariate model to leverage cross-correlation between a primary variable and one or more secondary variables.

• In RF Sensor Fusion, a sparse network of high-fidelity spectrum analyzers (primary) can be supplemented by a dense network of low-cost, noisy energy detectors (secondary).
• The cross-variogram models the joint spatial variability, allowing the dense secondary data to improve the prediction accuracy of the primary variable.
• This is critical for fusing heterogeneous sensor data in a Radio Environment Map (REM) to achieve high resolution at lower deployment costs.

The foundational prerequisite for Kriging interpolation. A variogram quantifies the degree of spatial dependence between RF measurements as a function of their separation distance. It models how signal similarity decays over space, typically exhibiting a nugget (measurement error), sill (maximum variance), and range (distance where correlation vanishes). Without an accurate experimental variogram fitted to a theoretical model (spherical, exponential, or Gaussian), Kriging weights are statistically invalid.

A Bayesian alternative to Kriging that provides identical predictions under specific covariance function choices but offers a more formal probabilistic framework. GPR defines a prior distribution over functions and updates it with observed sensor data to produce a posterior predictive distribution. Key advantage: it natively quantifies epistemic uncertainty at every interpolated point, making it ideal for REM applications where knowing how little you know in unmeasured regions is as critical as the predicted power level.

Extends static Kriging by incorporating the time dimension into the covariance model. Instead of relying solely on spatial neighbors, this technique leverages recent historical measurements at a single sensor location to improve current estimates. The spatio-temporal variogram models how correlation decays across both space lag and time lag simultaneously. Critical for tracking moving emitters or rapidly changing spectrum occupancy where a purely spatial snapshot becomes stale within milliseconds.

A deep learning alternative that models distributed sensors as nodes in a graph with edges representing spatial proximity. Message-passing layers aggregate information from neighboring nodes to predict spectrum values at unobserved locations. Unlike Kriging, GNNs can learn complex, non-linear spatial dependencies directly from data without requiring manual variogram modeling. However, they sacrifice the statistical interpretability and confidence intervals that Kriging provides natively.

A signal processing paradigm that exploits the sparsity of spectrum occupancy in the frequency domain to reconstruct wideband maps from sub-Nyquist samples. While Kriging interpolates in the spatial domain, compressed sensing reconstructs in the frequency domain by solving an L1-minimization problem. The two techniques are complementary: compressed sensing can recover the spectrum at individual sensors, while Kriging fuses those sparse reconstructions into a continuous spatial map.

A specific REM layer that isolates the large-scale, log-normal signal variation caused by terrain and building obstructions, distinct from distance-dependent path loss. Kriging is particularly well-suited for constructing shadow fading maps because the phenomenon exhibits strong spatial correlation—two receivers behind the same building experience similar attenuation. The variogram of shadow fading typically has a range of tens to hundreds of meters in urban environments, directly informing sensor deployment density.