Best interpolation method for a particular project

2018
8
11-19-2019 06:46 AM
MikeO_Shell1
New Contributor II

I'm trying to find the best interpolation method for  a project I'm working on.  I've tried the basic settings with spline, IDW & kriging and they all appear ok at first glance, but when I look more closely they don't follow the points exactly like they should.  For example, I will have a point or cluster of points with a value of 59, and a neighboring point or cluster of points with a value of 60.  When I adjust the symbology to 59.5 I would think the line would be drawn evenly between the two points or clusters of points every time.  In some cases the line is drawn between the two, but in other cases it is not.  The resulting map is "generally" correct, but with a lot of little errors here and there.  I'm not real familiar with how interpolation works exactly, and I want to know if one of the methods would be more accurate or have less errors for the particular project I'm working on. I will have 20,000+ points over an area the size of Florida (if that even matters).

0 Kudos
8 Replies
SteveLynch
Esri Regular Contributor

Mike

As far as exact interpolation, please reference the following image. 2 cells of a raster, white equals the cell center and red is the sample location.

Predictions are made at the cell center and that value is assigned to the whole cell. 

When predicting the value for the upper cell, 1395 will carry the most weight and therefore the cell's value will not be equal to 1395. However, 3199 falls exactly at the center of the cell and therefore that cell's predicted value will be equal to 3199.

The smaller the cell size the closer the predicted cell value will be to that of the sample point

Regards

-Steve

MikeO_Shell1
New Contributor II

Thank you.  That gave me a better understanding of the importance of cell sizes.  I guess I can experiment with the different methods while keeping the cell sizes low in each.  Are there any adverse effects of having cell sizes as low as possible?

0 Kudos
SteveLynch
Esri Regular Contributor

Are there any adverse effects of having cell sizes as low as possible?

- processing time

- size of output raster

- at some point you are not gaining anymore information

You have to think about the scale at which you phenomena changes. Say you decide on 30 meters, that means that each 30x30 meter area has the same value.

If you have access to Geostatistical Analyst you can perform cross validation to decide which interpolator produces the best results.

MikeO_Shell1
New Contributor II

Thanks again.  When you say 30x30 meter area, are you referring to the cell size or the spacing of the points?  I will look into cross validation.

0 Kudos
SteveLynch
Esri Regular Contributor

yes, cell size.

0 Kudos
MikeO_Shell1
New Contributor II

FYI, I figured out part of the issue.  My points were random and scattered.  So, I created a grid with evenly spaced points and ran a standard Kriging again and it turned out much more accurate.  Still not perfect, but much better.  Thanks again for your help.  I'll keep trying to tighten it up further.

0 Kudos
SteveLynch
Esri Regular Contributor

good.

However, if your points are on a grid you don't need to interpolate. You can use the PointToRaster tool, that is, if each cell has at least one point.

HeatherSmith
Esri Contributor

Hi Mike,

The intention of geostatistics is to create predicted values for areas where you have no sampled data. IDW is an exact method, which means that your output values will match your input values exactly. However, most often, this will produce less reliable predictions than inexact methods like kriging. This lesson explains this concept and explores a few different interpolation methods.

If you're not interested in creating a prediction surface and your goal is instead to display your data as a raster, then Steve's suggestion of the PointToRaster tool is probably easier.

If you are interested in interpolation and geostatistics,  there is a learning path here which starts with lessons that assume no prior knowledge. They also explain how to use cross validation to assess your results as Steve suggested above.