# How to identify concentrated areas with multiple point values to produce a "heat map" appearance

I have roughly 1.6 million points statewide with values ranging from 1 - 4. I'm looking to identify concentrations of points based on their values (1,2,3,4). I know there are several different routes you could take on this approach, but I am struggling to find one that produces a visually clean output. I am wanting to produce a map that would look similar to a "heat map." Here are a few of the different methods I've tried so far.

1) Converting all the points into a raster and using the most frequent cell assignment. The output is what I would expect, but I'm not a fan of the pixelated look. I used the smoothing tool and that cleaned up the raster cell edges, but I'm still not truly satisfied with the look of the end product.

2)The second approach I've tried is neighborhood point statistics using majority as the statistical type. This approach seems very subjective depending on the cell size and neighborhood size. The output however appears to be more of what I was hoping the end product would look like. However when looking at the individual point data, some of the output looks questionable.

Can anyone think of a more appropriate approach that would allow for more statistical analysis with the end product looking similar to a heat map? I've been looking into the hot spot analysis tool. I don't think that tool will work with my data since I have multiple point values. I am using ArcGIS 10.2 info.

• What do the points represent?
– Aaron
May 13, 2014 at 21:54
• The points represent households that are categorized as having a family. A value of 1 would indicate a household with a high level of education and income. The higher the value, the lower the education level and income would be. May 14, 2014 at 16:48

How about a geostatistical kriging or Gaussian process regression analysis? (same thing)

This would take into account the differing values of each point and the distance from each other, resulting in an output raster that can be adjusted under the display options to be represented exactly like a heat map.

I would suggest spherical kriging, all it would take is a single shapefile containing all the points, their locations of course, and the value each represents. then krig point shapefile by the values.

You can use Focal Statistics to derive meaningful information from your data. The following example shows how focal statistics can be used to represent the majority value of tree diameters within a forest. Focal statistics calculates various statistics using a moving window approach. The image to the left shows raster pixels with values from 1 - 9 representing tree radii. The image to the right shows the results of focal statistics. These are the steps I used to do the analysis:

1. Convert vector point data to raster Point to Raster (Conversion)
2. Convert floating point raster to integer (if needed) Int()
3. Run Focal Statistics. In this case, I calculated "Majority" within a 100m^2 rectangular moving window. You will need to determine the appropriate window size for your analysis.
4. Overlayed the results at 40% transparency on aerial imagery of the forest.

If the 1-4 values are ordinal, the same approach you use for heatmaps would probably work here.
Run Global Moran's I at different intervals to determine the peak band of spatial autocorrelation in the scale appropriate to the analysis you are doing. Then use that as a fixed distance band for Getis-Ord Gi*. Finally, run IDW against your Getis-Ord Gi* results.

Since your points are already weighted, this is perfect for Getis-Ord Gi*.

If, instead, 1-4 are categorical, you are probably going to have to look at some pretty complicated indicator kriging. Best bet would be to use absence presence on every category separately, e.g. first is split between all 1s versus 2,3,4 combined, second analysis is all 2s versus 1,3,4 combined, etc. Combining those results will be the hard part. You might to do conditional rasters based on the highest probability category at each location, or create a threshold probability for presence/absence and then use combined categories. For example, if only category 1 is above the threshold, use one color, it only category 2 use another, and if both are above the threshold use an intermediate color between the two. The cartography could get very complicated for that one with 4 categories though; probably impossible to display meaningfully.

• I can't get the Global Moran's to work. It keeps bombing out and saying memory error and I have plenty of memory. I tried using the default value for the distance band and the same thing happened. Have any ideas why this is happening? May 16, 2014 at 14:55