Module 1: Comparing Methods for Crime Hotspot Analysis

The first week of Applications in GIS kicked off with crime analysis using ArcGIS Pro. Crime statistics data for Washington D.C. and Chicago for the years of 2017 and 2018 were provided to examine any spatial patterns and to create crime hotspot maps. 

The original Washington D.C. 2017 crime data was queried to select rows that were classified as BURGLARY or ASSAULT and the selection was exported as two new feature classes to create two hotspot maps. For both maps, a spatial join (from the burglary/assault table to the census tracts table) was completed to create a new field that showed the number of burglaries or assaults in each census tract. The crime rate was calculated by dividing the number of burglaries or assaults by the number of household units and multiplying by 1000.

For the burglaries map, the census tracts were set to show the crime rate for each census tract. Five classes were created using a natural breaks classification method and a graduated color scheme. Extremely high crime rates due to lower household numbers were excluded using the Advanced Symbol Options and using SQL expressions to exclude these rates. The hatched areas included the values that were excluded from the main class categories. A snapshot of the map is below: 

For the kernel density assault hotspot map, the kernel density tool was used on the exported assault data, with a cell size of 100 ft and a search radius of 1320 ft. The results were displayed as graduated colors, with 6 classes based on the mean assault value. The mean assault rate value was determined using the symbology statistics, in which the lowest class was set to equal the mean, followed by the mean*2, mean*3, etc. Values of zero were excluded using a SQL expression in the Advanced Symbol Options. The final assault hotspot map snapshot is below: 

Next, point crime data for Chicago was analyzed using three different methods to create three homicide hotspot maps. The first map used a Grid-Based Thematic method, in which grids cells of 1/2 miles was spatially joined to the homicide data and the resulting top 20% homicide counts were used to create the hotspot map below: 

The same data was evaluated using a kernel density method. The output cell size was set to 100 square miles and a search radius of 2630 to create the hotspot map below: 

The last method used the Cluster and Outlier Analysis (Anselin Local Moran's I) tool. The results show different types of clusters, specifically low-low, low-high, high-low and high-high cluster types. To create the hotspot map, the clusters categorized as high-high were exported as a separate feature class to create the map below: 

Each method produced different results for the same dataset. These hotspot maps could be applied towards predicting future crime hotspots, since past crime hotspots tend to be a good indicator of future crime locations. In order to compare the efficiency of each method in predicting the next year's homicide hotspots, the 2017 hotspot maps were compared to the actual 2018 homicide locations data for the Chicago area. The table below was created to compare:

Hotspot Technique	Total Area (sq mi) in 2017	Number of 2018 homicides within 2017 hotspot	% of all 2018 homicides within 2017 hotspot	Crime density (2018 homicides  within 2017 hotspot per sq mi)
Grid Overlay	15.46	159	26.99	10.28
Kernel Density	25.80	256	43.46	9.92
Local Moran’s I	51.02	347	58.91	6.80

Based on the evaluation, the Local Moran’s I method has a larger area, but a smaller density, which would mean having to spread out the police resources over a larger area. The Grid Overlay method has the most condensed area, but a lower accuracy (2018 percentage). Therefore, Kernel Density is the best for predicting future homicides because it provides a more targeted area with a higher 2018 percentage than the grid overlay but still has a higher density than the Local Moran’s I map.