Kernel Density Value Interpretation

3712
11
05-26-2020 10:01 AM
MariaManz
New Contributor

Hello! I am working with shark tagging data, and I am running a kernel density analysis in ArcGIS Pro to determine areas of highest density. My output visually looks correct, however the value outputs are confusing to me. I understand they are the pixel density, but I am wondering how I can translate to the population density of sharks in that area so I can use the numbers as well as the image in my analysis. For example when I run my kernel density in RStudio my output shows that 75% of the population is within a certain area. Any clarification would be very much appreciated! 

0 Kudos
11 Replies
DanPatterson
MVP Esteemed Contributor

You ruled out the help topic descriptions?  For example...

How Kernel Density works—Help | Documentation 


... sort of retired...
0 Kudos
MariaManz
New Contributor

Hi! Thanks for the quick reply, I did look into this and I think the answer I am looking for is in this sentence specifically: "The density at each output raster cell is calculated by adding the values of all the kernel surfaces where they overlay the raster cell center.". I am just. not quite sure how to interpret that. I am new to GIS, and I am not quite sure what they mean by "kernel surfaces" and "raster cell center". Thanks again!

0 Kudos
DanPatterson
MVP Esteemed Contributor

it might help if you had an image of your input points and the options that you specified in the KD tool.

In short, you have observations of something in space... a sample effectively.  What you are trying to do is represent the "density" of those points with respect to the other points and a distance for each cell.  Each cell in the resultant raster is needs a value which is calculated as follows:

Kernel density spreads the known quantity of the population for each point out from the point location. The resulting surfaces surrounding each point in kernel density are based on a quadratic formula with the highest value at the center of the surface (the point location) and tapering to zero at the search radius distance. For each output cell, the total number of the accumulated intersections of the individual spread surfaces is calculated.

From

Differences between point, line, and kernel density—Help | Documentation 

So it important to know the cell size and many other parameters that are specified in the tool

Kernel Density—Help | Documentation 

In the end the result is a measure of the number of occurrences of something per unit area.  If you change the parameters, keeping the data set the same, you will conceivably get a different result


... sort of retired...
0 Kudos
MariaManz
New Contributor

Thanks again for your continued help, I think it is starting to become more clear to me. I have attached a screenshot of my input values, where "Full_Mako" is my data set layer that has the lat and long coordinates included. The rest of the input values were automatically generated by GIS, and the option to change the unit area has disappeared? I do not populate the data to anything because each point represents one location, so it is weighted by one. I have also included the output values. Thanks again! 

0 Kudos
DanPatterson
MVP Esteemed Contributor

Maria Manz

  • Use None if no item or special value will be used and each feature will be counted once.
  • cell size .... 0.1688...
  • method ... planar
  1. I think that your coordinates are in decimal degrees?  Right? 
    • if so is your study area huge? like 42 degrees high and/or wide? (1/250 of the width or height is the default cell size in the units of the data)
  2. Why don't you want to specify a search radius? was this agreed upon? or is this an R default that you are emulating?  Think about this. Do you see clusters in the points?  In what sort of radius are the clusters?
  3. method... use geodesic and revisit your cell size choice, especially if your points are clustered into groupings smaller than your cell size

It is hard to make any other recommendations without the "end goal" being known, especially if this goes beyond a pretty map and you need to make inferences about the population patterns.  So, you can get a map, it will represent 'density per unit area' but the details still depend on the choices made.

If you have a license for 

An overview of the Spatial Statistics toolbox—ArcGIS Pro | Documentation

 

there are other tools there for mapping and examining patterns (ellipse pattern) and clusters or

An overview of the Geostatistical Analyst toolbox—Help | Documentation 

if you are looking at a predictive surface.

 

So many decisions eh ?


... sort of retired...
MariaManz
New Contributor

Yes, my study area spans from about Canada to the Gulf of Mexico, so it is quite large. I didn't specify a study area because I have such a big study area and there are multiple areas of high density or clusters, I provided an image below, and I thought I read somewhere that in that situation it is better to use the default settings. But thank you for your advice! I'll try to look into ways for estimating cell size and search radius! Thanks again!

0 Kudos
DanPatterson
MVP Esteemed Contributor

On a quick note... you might want to remove the mako shark sightings in the pacific ocean and particularly those on land!

These are going to have a big impact on the results you have obtained.  You will need to eliminate the landmass completely with nodata values.  These are skewing your results


... sort of retired...
MariaManz
New Contributor

I think I am finally getting closer to reasonable results, thank you so much for your help! 

0 Kudos
DanPatterson
MVP Esteemed Contributor

Maria Manz‌ post again, if you get acceptable results and show what parameters you used for the inputs.  It is important that you masked/removed the land from the analysis... you don't want to set off alarm bells about onshore shark sitings


... sort of retired...