How to Conduct Kernel Density Estimation

[JB2033]

HI GIS folks, 

 

I'm hoping you can help me figure out how to conduct a Kernel Density Estimation on Deer-Vehicle Collision data. 

 

I have some Deer Vehicle Collision data and i'm trying to identify "hot spots" for these collisions within a certain region. 

 

In a "flow chart" style explanation, how would I do this?

 

So far, my understanding is this: 

 

Deer Vehicle Data ---> Apply: Kernel Density Estimation ---> Produce: Kernal Density Raster Surface ---> Apply: Reclassification of Raster Density Surface ----> Produce: A Reclassified Raster Density Surface (AKA Hot Spot Map?) 

 

 

Am I missing any major steps?

 

Thanks for your help!

[Lurker]

yep, that's the way

[Svenne68]

Well a KDE is NOT a hotspot (although many think so and use it as)....It only shows you a relative density and has NO statistic footing....If you want a hotspot/coldspot output that is statistically worthy and reliable you need to use the centrographic tools in Spatial Statistical and especially the Getis-Ord Gi*

[Lurker]

nice adding Svenne68,

 

so after conduct KDE , can we directly use the output data to make hotspot map with those tools? or we need additional steps again?

[Svenne68]

The centrograpic tools and global statistic tools (mean, standard distance, directional distribution, NNI and Getis-Ord General(Gearys C)) will tell you if there is patterns and clusters in your data and how significant they are but NOTHING about where and why. 

 

You might need to manipulate/change/add your data before you can perform a Getis-Ord Gi* test (also one with automatic rendering). This tool will tell you WHERE your coldspots are and your hotspots are and the area that is 'not significant' and it will do it at a 90-95-99% confidence level and give you values about observed distance and expected distance and other values that is statistical proven.

 

KDE and Hotspot analysis is two different things and therefore  NOT comparable although MANY police forces uses the KDE. The KDE will hide your coldspots.....depending on your threshold it will increase/decrease your density.....if use over different polygons/areas it will give you different outputs that are not in the same scale but only relevant to the dataset.....Getis-Ord Gi* does NOT do this. Intended or unintended you could lead he user/reader into misinterpretation with a KDE......this would be almost impossible with the Gi*

 

A good reference is the 'Understanding hotspot' by Chainey et al and published by NIJ in 2005 (National Institute of Justice) 

[Lurker]

enlightening me Svenne68, thanx you