Bayesian Universal Kriging
Also known as: BUK, Bayesian kriging with trend, Bayesian spatial interpolation with covariates, stochastic universal kriging
Bayesian Universal Kriging (BUK) extends classical universal kriging by placing prior distributions on trend coefficients and spatial covariance parameters, then propagating full posterior uncertainty into predictions. It interpolates spatially referenced continuous data while simultaneously estimating large-scale deterministic trends driven by covariates and small-scale stochastic spatial dependence, yielding prediction intervals that honestly account for both parameter and interpolation uncertainty.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use Bayesian Universal Kriging when you have georeferenced continuous measurements (e.g., pollution levels, soil properties, temperature) and meaningful spatial covariates (elevation, land-use type) that explain large-scale trends, and when honest uncertainty quantification is important — for instance in environmental risk assessment or resource estimation. It is especially appropriate when sample sizes are small to moderate and classical variogram estimation is unreliable, or when prior scientific knowledge about correlation ranges should be incorporated. Avoid it when computational cost is prohibitive (large n > 10,000 without approximations), when you need instantaneous predictions in real time, or when spatial trend is absent and ordinary kriging suffices.
Strengths & limitations
- Provides full predictive distributions, yielding credible intervals that account for uncertainty in both trend coefficients and covariance parameters.
- Naturally incorporates spatial covariates as fixed-effect trend components, improving prediction where explanatory variables are informative.
- Prior information about plausible correlation ranges or nugget ratios can be encoded formally, reducing estimation instability in sparse datasets.
- Posterior predictive checks allow model fit to be assessed rigorously within the Bayesian workflow.
- Handles non-stationary means properly by modelling the trend separately from the spatially correlated residual.
- Computationally intensive: MCMC requires inverting the n×n covariance matrix at each iteration, scaling as O(n³), making exact inference impractical for n beyond a few thousand.
- Sensitive to prior specification, especially for range and nugget parameters when data are sparse or irregularly distributed.
- Requires the user to choose and justify a parametric covariance family (Matérn, exponential, etc.) whose misspecification can bias predictions.
- Convergence of MCMC chains must be diagnosed carefully; poorly mixing chains produce unreliable posteriors.
- More complex to implement and interpret than classical universal kriging, demanding familiarity with Bayesian computation.
Frequently asked
How does Bayesian Universal Kriging differ from classical Universal Kriging?
Classical universal kriging treats variogram parameters as fixed known quantities (estimated separately via REML or method-of-moments) and then conditions on them. BUK places prior distributions on those parameters and marginalises over them, so prediction intervals are wider but better calibrated — they capture uncertainty from parameter estimation as well as interpolation uncertainty.
What covariance function should I choose?
The Matérn covariance family is usually preferred because its smoothness parameter ν allows you to control how differentiable the spatial process is. If ν is also treated as unknown, it can be estimated from the data, though joint inference on ν and the range φ requires care due to identifiability constraints.
Can Bayesian Universal Kriging handle very large datasets?
Exact BUK with MCMC scales as O(n³) per iteration and becomes impractical for n beyond a few thousand. Approximation strategies include predictive-process models, nearest-neighbour Gaussian processes (NNGP), and low-rank spatial basis functions, all of which reduce computational cost while retaining the Bayesian uncertainty propagation.
Do I need informative priors to get good results?
Not necessarily, but weakly informative priors — for example, constraining the spatial range to be at most the diameter of the study region — prevent the sampler from exploring implausible parameter values and improve MCMC efficiency. Fully diffuse priors on variance and range can produce improper or very slowly mixing posteriors.
When should I prefer Bayesian Ordinary Kriging over Bayesian Universal Kriging?
Choose Bayesian Ordinary Kriging when the mean is approximately constant (stationary mean assumption is reasonable) and no useful spatial covariates are available. BUK adds complexity and requires meaningful covariates; if trend covariates are weak or collinear, the simpler ordinary kriging variant will perform comparably with less risk of over-parameterisation.
Sources
- Diggle, P. J., & Ribeiro, P. J. (2007). Model-Based Geostatistics. Springer. ISBN: 978-0387329079
- Banerjee, S., Carlin, B. P., & Gelfand, A. E. (2015). Hierarchical Modeling and Analysis for Spatial Data (2nd ed.). CRC Press. ISBN: 978-1439819173
How to cite this page
ScholarGate. (2026, June 3). Bayesian Universal Kriging. ScholarGate. https://scholargate.app/en/spatial-analysis/bayesian-universal-kriging
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Bayesian Ordinary KrigingSpatial analysis↔ compare
- Co-krigingSpatial analysis↔ compare
- Geographically Weighted RegressionSpatial analysis↔ compare
- Ordinary KrigingSpatial analysis↔ compare
- Spatial AutocorrelationSpatial analysis↔ compare
- Universal KrigingSpatial analysis↔ compare