Space-Time Kernel Density Estimation (ST-KDE)
Also known as: ST-KDE, spatiotemporal kernel density estimation, space-time KDE, 3D kernel density estimation
Space-Time Kernel Density Estimation extends classical KDE into three dimensions — two spatial and one temporal — to reveal how the intensity of point events (crimes, accidents, disease cases) varies continuously across both geographic space and time. It produces a smooth probabilistic surface that highlights where and when events concentrate most densely.
Key highlights
- Produces a continuous, intuitive density surface that captures smooth gradients in event intensity without imposing arbitrary administrative boundaries.
- Jointly captures spatial and temporal clustering, revealing whether hot spots are stable, migrating, or seasonal.
- Non-parametric: no distributional assumption about the underlying event process is required.
- Flexible bandwidth choice allows adaptation to datasets with heterogeneous event densities.
- Outputs are directly map-ready and communicate findings clearly to non-technical stakeholders.
Intuition
This section is available to Pro members. Upgrade to Pro
How it works
This section is available to Pro members. Upgrade to Pro
When to use it
Use ST-KDE when you have a georeferenced point-event dataset with reliable timestamps and you want a continuous, non-parametric picture of how event intensity shifts across both space and time simultaneously. It is well-suited for crime analysis, epidemiology, accident mapping, and ecological movement studies with sufficient event density. Avoid it when the dataset is very sparse (fewer than ~50 events), when timestamps are imprecise or missing, when the study area has strong physical barriers that a smooth kernel cannot reflect, or when you need formal hypothesis tests rather than exploratory visualization — in those cases, use Getis-Ord Gi* or space-time scan statistics instead.
Strengths & limitations
- Produces a continuous, intuitive density surface that captures smooth gradients in event intensity without imposing arbitrary administrative boundaries.
- Jointly captures spatial and temporal clustering, revealing whether hot spots are stable, migrating, or seasonal.
- Non-parametric: no distributional assumption about the underlying event process is required.
- Flexible bandwidth choice allows adaptation to datasets with heterogeneous event densities.
- Outputs are directly map-ready and communicate findings clearly to non-technical stakeholders.
- Results are sensitive to bandwidth choice: too small produces noisy, spiky surfaces; too large over-smooths and masks real clusters.
- Assumes spatial and temporal kernels are separable (independent), which may not hold when events have complex space-time interaction.
- Does not provide formal statistical significance tests for clusters; only exploratory density estimates.
- Edge effects near the boundary of the study area can inflate or deflate density estimates; correction methods (e.g., edge correction weights) add complexity.
- Computationally expensive for very large datasets or fine-resolution grids.
Common pitfalls
This section is available to Pro members. Upgrade to Pro
Applications
This section is available to Pro members. Upgrade to Pro
Frequently asked
How do I choose the spatial and temporal bandwidths?
Start with cross-validation or least-squares cross-validation (LSCV) to select optimal bandwidths data-driven. Also run sensitivity analysis across a grid of bandwidth combinations and report how hot spot locations shift. Domain knowledge — e.g., typical patrol radius or disease incubation period — can usefully anchor the search range.
Is ST-KDE the same as running KDE separately for each time period?
No. Running KDE independently for each time slice treats time as a categorical grouping variable and ignores smoothing across the temporal dimension. ST-KDE applies a continuous kernel in the time axis, borrowing information from temporally nearby events and producing a fully smooth three-dimensional density surface.
Can ST-KDE be used to predict future hot spots?
ST-KDE is primarily a descriptive and exploratory tool. It characterises past event density and can be extrapolated cautiously for near-future risk mapping (e.g., predictive policing), but it has no explicit forecasting mechanism. For formal prediction, combine ST-KDE with prospective space-time scan statistics or spatiotemporal regression models.
What software implements ST-KDE?
The spatstat package in R supports space-time KDE; ArcGIS Pro offers space-time cube tools; QGIS via the STKDE plugin; and Python via the scipy and scikit-learn libraries for custom implementations. The Nakaya-Yano original implementation is available as a standalone tool referenced in their 2010 paper.
How many events do I need for reliable ST-KDE estimates?
As a practical rule, aim for at least 50-100 events spread across the study area and time window. Very sparse datasets produce highly unreliable density estimates regardless of bandwidth. For thin event counts, consider aggregating to coarser spatial or temporal units, or switching to a parametric space-time model.
Sources
- 1.Nakaya, T., & Yano, K. (2010). Visualising crime clusters in a space-time cube: An exploratory data-analysis approach using space-time kernel density estimation and scan statistics. Transactions in GIS, 14(3), 223-239.
- 2.Kernel density estimation. Wikipedia.
You have read it. What now?
Cite this page
ScholarGate. (2026, June 3). Space-Time Kernel Density Estimation. ScholarGate. https://scholargate.app/spatial-analysis/space-time-kernel-density-estimation