The Influence of Data Classification Methods on Predictive Accuracy of Kernel Density Estimation Hotspot Maps
Nenad Milic1, Brankica Popovic2, Sasa Mijalkovic1, and Darko Marinkovic1
1Department of Criminalistics, University of
Criminal Investigation and Police Studies, Serbia
2Department of Informatics and Computer Science, University of Criminal
Investigation and Police Studies, Serbia
Abstract: When it comes to hot spot
identification, spatial analysis techniques come to the fore. One of such
techniques, that has gained great popularity among crime analysts, is the Kernel
Density Estimation (KDE). Small variation in KDE parameters can give different
outputs and hence affect predictive accuracy of hotspot map. The influence
these parameters have on KDE hotspot output sparked many researches, mostly
analyzing the influence of the cell size and bandwidth size. Yet, the influence
of different classification methods applied to calculated cell values,
including the choice of threshold value, on the KDE hotspot predictive accuracy
remained neglected. The objective of this research was to assess the influence of different classification methods to KDE
predictive accuracy. In each KDE computation, calculated cell values were
divided into five thematic classes, using three the most common (default)
classification methods provided by Environmental Systems Research Institute (ESRI)
Geographical Information System (Arc GIS) (equal interval classification,
quantile classification and natural breaks classification) and incremental
multiples of the grid cells’ mean. Based upon calculated hit rates, predictive
accuracy indices and recapture rate indices and taking into account the
necessity that mapping output should satisfy some operational requirements as
well as statistical rules, this research suggest that incremental mean approach
with hotspot threshold of 3 and above multiples of the grid cell’s mean, should
be used.
Keywords: Crime mapping, hot spot, kernel density, classification methods.