Space-Time Universal Kriging
Also known as: STUK, spatiotemporal universal kriging, space-time kriging with trend, universal kriging in space-time
Space-Time Universal Kriging (STUK) is a geostatistical method that interpolates a continuously varying phenomenon across both space and time while explicitly modelling a deterministic trend component. It generalises Universal Kriging to the joint space-time domain, producing unbiased optimal predictions and associated uncertainty estimates at unobserved space-time locations.
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 Space-Time Universal Kriging when you have point-referenced observations collected at irregular space-time locations and need interpolated maps or time-series at unsampled locations, and when the variable has a discernible spatial or temporal trend (e.g., temperature gradient, temporal seasonality) that must be accounted for rather than assumed away. It is well-suited for environmental monitoring, climate reanalysis, air quality mapping, and epidemiological surveillance with moderately sized datasets. Avoid it when the trend is poorly specified or highly nonlinear (prefer nonparametric alternatives), when data are gridded and abundant (machine-learning or deep-learning interpolators may be faster), when the dataset is very large (the n x n kriging system becomes computationally prohibitive without approximations), or when observations are not spatially referenced.
Strengths & limitations
- Provides optimal (minimum-variance) unbiased predictions at any unsampled space-time location.
- Produces a full prediction uncertainty map (kriging variance) alongside the interpolated values.
- Handles non-stationary means by explicitly modelling spatial and temporal trend components.
- Flexibly accommodates asymmetric space-time covariance structures via product-sum or metric variogram models.
- Well-established theoretical foundations (BLUP) with implementations in R (gstat), Python (PyKrige), and ArcGIS.
- Computationally expensive: solving the kriging system scales as O(n^3) in the number of observations, making it impractical for large datasets without approximations such as local neighborhoods.
- Requires careful specification and fitting of the space-time variogram, which demands sufficient data density across both dimensions.
- Results are sensitive to misspecification of the trend model; an incorrect trend can bias predictions systematically.
- Assumes second-order stationarity of the residual process, which may not hold in heterogeneous landscapes.
- Collapses to ordinary kriging if the trend is omitted, so model selection (OK vs. UK) is an additional step.
Frequently asked
How does Space-Time Universal Kriging differ from Space-Time Ordinary Kriging?
Ordinary Kriging assumes a constant but unknown mean and enforces a single unbiasedness constraint (weights sum to 1). Universal Kriging assumes a spatiotemporal trend modelled by known basis functions and adds one unbiasedness constraint per trend term. STUK should be preferred when a clear gradient or trend is visible in the data; if the mean appears stationary, Ordinary Kriging is simpler and equally valid.
What space-time variogram models are commonly used?
The most common choices are the product model (separable: spatial and temporal variograms multiply), the product-sum model (adds spatial and temporal components plus their product, allowing more flexibility), and the metric model (combines space and time into a single scaled distance). The product-sum model is widely recommended as a balance between flexibility and identifiability.
How large a dataset is too large for this method?
The full kriging system requires inverting an n x n matrix, so datasets above roughly 5,000-10,000 observations become slow without local neighbourhood approximations. In practice, most implementations use a moving window (local kriging) that retains only the nearest n_local observations, reducing computation substantially while accepting a small loss of global optimality.
Can I use Space-Time Universal Kriging for forecasting future time points?
Yes, as long as the trend model can be extrapolated (e.g., a fitted polynomial or harmonic in time) and the variogram is fitted to historical data. However, prediction variance grows rapidly as the temporal target moves beyond the observation window, so uncertainty bounds must be communicated clearly. For long-range forecasting, dynamic spatiotemporal models (e.g., state-space or machine-learning-based) are generally preferred.
What software implements Space-Time Universal Kriging?
The R package gstat (with the spacetime package for data handling) provides full support including space-time variogram modelling, fitting, and kriging. Python users can use PyKrige with custom covariance functions. ArcGIS Pro offers empirical Bayesian kriging with spatiotemporal extensions. MATLAB's Statistics and Machine Learning Toolbox also supports kriging with trend.
Sources
- Kyriakidis, P. C., & Journel, A. G. (1999). Geostatistical space-time models: A review. Mathematical Geology, 31(6), 651-684. DOI: 10.1023/A:1007528426688 ↗
- Graler, B., Pebesma, E., & Heuvelink, G. (2016). Spatio-temporal interpolation using gstat. The R Journal, 8(1), 204-218. link ↗
How to cite this page
ScholarGate. (2026, June 3). Space-Time Universal Kriging. ScholarGate. https://scholargate.app/en/spatial-analysis/space-time-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.
- Geographically Weighted RegressionSpatial analysis↔ compare
- Ordinary KrigingSpatial analysis↔ compare
- Space-Time KrigingSpatial analysis↔ compare
- Space-Time Ordinary KrigingSpatial analysis↔ compare
- Universal KrigingSpatial analysis↔ compare