Variogram Estimation

The Best Linear Unbiased Predictor that uses the modeled variogram as its core input. Kriging computes optimal weights for neighboring measurements by solving a system of linear equations derived from the spatial autocorrelation structure captured by the variogram.

• Ordinary Kriging: Assumes an unknown but constant local mean
• Universal Kriging: Accounts for a deterministic spatial trend
• Co-Kriging: Incorporates secondary correlated variables (e.g., terrain elevation) to improve RF prediction accuracy

The variogram parameters—nugget, sill, and range—directly control how quickly the influence of a sensor measurement decays with distance.

A Bayesian non-parametric alternative to Kriging that provides both a predicted mean and a full posterior variance at every unobserved location. The covariance function (kernel) plays the same role as the variogram in classical geostatistics.

• Matérn kernel: A flexible covariance family that generalizes the exponential and squared-exponential models
• Uncertainty quantification: The predictive variance naturally identifies regions where additional sensor deployments would most reduce mapping error
• Hyperparameter learning: Kernel parameters can be optimized via maximum marginal likelihood rather than manual empirical variogram fitting

GPs are particularly valuable when confidence intervals are required for regulatory compliance in dynamic spectrum access.

Extends variogram modeling into the time domain by treating temporal separation as an additional dimension. The space-time variogram quantifies how quickly RF measurements decorrelate both geographically and chronologically.

• Separable models: Assume the space-time covariance factors into independent spatial and temporal components
• Non-separable models: Capture complex interactions where spatial correlation structure itself evolves over time
• Product-sum models: A practical compromise that combines marginal spatial and temporal variograms with a multiplicative interaction term

This is essential for predictive REMs that forecast spectrum occupancy minutes or hours into the future for proactive resource allocation.

The data-driven first step where the experimental variogram is computed directly from observed RF measurements. Sensor pairs are binned by separation distance, and the average squared difference of their measurements is plotted.

• Matheron's estimator: The classical method, sensitive to outliers in skewed RF power distributions
• Cressie-Hawkins robust estimator: Uses square-root absolute differences, more resistant to anomalous measurements from intermittent interference
• Bin width selection: Too narrow creates noisy variograms; too wide obscures fine-scale spatial structure
• Isotropic vs. anisotropic: Directional variograms detect when spatial correlation depends on orientation (e.g., along a street canyon vs. across buildings)

The process of fitting a parametric theoretical model to the empirical variogram points. The fitted model must be conditionally negative definite to ensure valid Kriging variance calculations.

• Spherical model: Linear behavior near the origin with a finite range, common for phenomena with a clear zone of influence
• Exponential model: Asymptotically approaches the sill, suitable for RF propagation with long-tailed spatial dependence
• Gaussian model: Parabolic behavior near the origin, indicating very smooth spatial variation
• Nugget effect: The discontinuity at zero distance, representing measurement error and microscale variation below the sensor spacing resolution

Weighted least squares is the standard fitting method, giving more influence to bins with more sensor pairs.

A leave-one-out validation technique that assesses variogram model quality by temporarily removing each sensor measurement and predicting it from its neighbors using Kriging with the candidate variogram.

• Standardized residuals: Should have zero mean and unit variance if the variogram model is well-calibrated
• Q-Q plots: Compare residual distributions against theoretical normality to detect systematic bias
• Mean squared error ratio: The ratio of actual to predicted Kriging variance should approach 1.0 for a correctly specified model

Poor cross-validation statistics indicate the need to revisit variogram parameter selection or consider anisotropic or non-stationary model extensions.