A thousand dots can tell you where events happened. They do not immediately tell you where the underlying concentration is strongest.
Crime incidents, traffic crashes, trees, shops, disease cases and wildlife sightings can all be represented as points. When many points overlap visually, the map becomes difficult to read. Geography often asks a different question: where is the local intensity of these events highest?
Kernel density estimation, or KDE, converts point observations into a smooth surface by placing a small influence function around each point and adding those influences together.
KDE does not discover a hidden surface. It constructs a useful estimate of concentration from the points and a chosen neighbourhood size.
Quick Read: The KDE Mechanism
POINT EVENTS → KERNEL AROUND EACH POINT → BANDWIDTH → SUM OVER SPACE → DENSITY SURFACE
1. Each Point Contributes Local Influence
A kernel places the greatest weight near an event and progressively less weight farther away until its influence reaches zero or becomes negligible. Overlapping kernels create higher estimated density.
2. Bandwidth Controls the Geography
The bandwidth determines how far each point spreads its influence. A narrow bandwidth reveals fine local concentrations. A broad bandwidth produces a smoother regional pattern. Neither is automatically correct.
3. Too Narrow Can Chase Noise
With a very small bandwidth, almost every cluster of a few points can become a separate peak. The map may describe sampling noise rather than a stable geographic process.
4. Too Broad Can Erase Structure
A very large bandwidth can merge distinct concentrations into one smooth hill. Important local hotspots disappear because the method has averaged them together.
5. Density Is Not Risk
A dense cluster of traffic crashes may occur where traffic volume is highest. That does not prove each journey is unusually dangerous. To estimate risk, the event count may need an exposure denominator such as vehicles, population or time spent in the area.
6. Density Is Not Causation
A hotspot identifies concentration. It does not explain why the concentration exists. Spatial Association and Causation remains the owner of causal interpretation.
7. KDE Is Not Spatial Interpolation
Spatial Interpolation predicts the value of a measured variable between observations. KDE estimates the intensity of point occurrence. One estimates attributes; the other estimates event concentration.
8. Boundaries Can Distort Density
Near a coastline or study boundary, part of a kernel can extend into impossible or unobserved space. Edge correction may be needed when the boundary constrains where events can occur.
9. Weighted KDE
Points can carry weights. A traffic incident with five casualties might contribute more than a minor incident, or a facility point might be weighted by capacity. The resulting surface then represents weighted intensity rather than simple event count.
10. Primary Geography: Where Are the Stickers Densest?
Place dots on a playground map for where pupils stopped during recess. Ask where dots are most concentrated, then imagine each dot as a small fading circle. The overlapping circles reveal the intuition behind KDE.
11. Secondary Geography: Change the Bandwidth
Students can compare two density maps from the same points using different bandwidths. The exercise makes an important principle visible: analytical settings can change the apparent geography even when the events stay fixed.
12. Advanced Geography: Choose Bandwidth With Evidence
Bandwidth can be selected through subject knowledge, cross-validation or formal optimisation. The appropriate spatial scale should correspond to the process being studied rather than simply producing the most attractive map.
13. Singapore Example: Traffic Incidents
A KDE surface can identify corridors where reported incidents concentrate. But traffic volume, junction density and road length must be considered before calling those corridors unusually risky.
14. Singapore Example: Amenities
Mapping clinics, food outlets or community facilities as a density surface can reveal service concentrations. Yet density does not equal accessibility: capacity, opening hours and transport still matter.
15. Ecology Example
Wildlife sightings can be smoothed into apparent habitat-use intensity, but observer effort matters. More sightings beside accessible trails may reflect where people looked rather than where animals prefer to live.
16. Hostile Test: “The Red Area Is the Hotspot”
Ask what bandwidth, denominator, study boundary and sampling process produced the red area. A density surface is a modelled representation, not direct evidence of cause or risk.
17. Where KDE Reasoning Breaks
- Bandwidth blindness: accepting the smoothing distance without justification.
- Density-risk collapse: treating event concentration as individual risk.
- Cause-by-colour: interpreting a hotspot as causal proof.
- Exposure blindness: ignoring population, traffic or observation effort.
- Edge blindness: allowing kernels to spill across impossible boundaries without thought.
- Pretty-map bias: choosing parameters for appearance rather than mechanism.
18. Ten Questions for KDE
- What does each point represent?
- Were all events equally observable?
- What bandwidth was used?
- Why is that bandwidth appropriate?
- Are points weighted?
- Does a boundary constrain events?
- Is density being confused with risk?
- What denominator would change interpretation?
- Does the pattern survive other bandwidths?
- What mechanism could explain the concentration?
19. Where This Fits
Spatial Distribution owns pattern broadly. Spatial Interpolation owns prediction of measured attributes between observations. This article owns the transformation of point events into an estimated intensity surface.
The Idea to Keep
A hotspot map is partly the geography of the events and partly the geography of the smoothing rule used to see them.