Interpolation
THE ART AND SCIENCE OF CALCULATING VALUES OF PLACES BASED
ON NEARBY DATA.
Not to be confused with extrapolation
Which is the art and science of calculating values for somewhere that is outside
the extent of known data.
But back to interpolation
At its simplest, think drawing isolines from point data. Elevation, temperature,
pressure, etc. Those of you that had me in 301 did some of this. It’s manual and
subjective, but does the job.
If you’re looking at this blankly, google drawing isolines and get up to speed.
Because the rest of this is, basically, getting the computer to draw those isolines for you.
I’ve added a couple of textbook chapters to the week 6 folder. Be sure to read them!
And ask questions where things don’t make sense. (don’t worry about the math – but
pay close attention to the diagrams.
Burroughs has the best descriptions of spatial interpolation I’ve seen. But the other one
has a lot of great figures demonstrating what happens to the same data when using
different interpolation techniques.
Two basic types
Discrete – we use this when NOT working with continuous surfaces
For example, I got milk prices for every business (that sells milk) in Ellensburg. I
want to interpolate. Why, who knows, but still…. You cannot buy milk inbetween
these locations, so you cannot estimate the prices there. What you can do is build
thiessen polygons which represent the area closest to each milk selling location. In
short, the cost of the nearest milk. We tend to do this in vector.
Continuous – think temperature, humidity, elevation, etc. All things that exist
everywhere. With the exception of isolines, we do this in raster.
Let’s now look at these interpolation methods in a bit more detail. Don’t forget those
two book chapters.
Thiessen Polygons (also known as Voronoi
polygons)
Yes, you can do this in Pro.
https://pro.arcgis.com/en/pro-app/tool-reference/analysis/create-thiessen-polygons
.htm
http://www.gitta.info/Accessibilit/en/html/UncProxAnaly_learningObject4.html
https://www.caliper.com/glossary/what-is-a-thiessen-polygon.htm
This is the only valid interpolation technique for discrete (non-continuous) data.
Continuous interpolators
Remember, we are working off an input point file and producing a raster output
file.
They can be
Local: Which means that the interpolation uses only a few points
Global: which means the interpolation algorithm uses all the input points at once.
And
Exact: meaning the output layer perfectly matches the input points (the raster under the
input points has the same value as the input point).
Inexact: meaning the output layer doesn’t match the input points (some sort of
averaging is used)
Global interpolators
Rarely (never?) used. But worth mentioning. They are better described in the
articles.
Trend surface analysis. Think grabbing a piece of paper and fitting it to your data.
First order, the paper is flat. Second order, the paper has one curve, Third order, two
curves, etc….
Universal Kriging. Don’t worry about it.
Local interpolators
Except for discrete data and contour lines, this is what is used!
Three main options:
Splines (exact)
Weighted Average or inverse distance (inexact)
Kriging (exact).
Spline interpolator
In 2D space, think fitting a line to all the input points. In 3D, think fitting a piece
of paper to all the input points.
In other words, your output surface perfectly fits the input points.
About all you can do is play around with how curvy it is inbetween points. But
really, just take the defaults if you go with this.
Splines in Pro:
https://pro.arcgis.com/en/pro-app/tool-reference/3d-analyst/how-spline-works.htm
Inverse Distance Weighting (IDW)
The closest analogue is creating contour lines. In this case, the algorithm looks at
a small neighborhood and calculates an average position – weighted by proximity
to input points. The closer the input point to the output raster pixel, the more
important it is.
User choices include the search radius, how many points to use, and the distance
weighting. Basically, your radius and # of points used will depend on the density
of your input points. The denser the points, the smaller the search radius.
Because it always uses multiple points (and thus some sort of an average), it is
inexact.
Note – read the Pro page:
https://pro.arcgis.com/en/pro-app/tool-reference/spatial-analyst/how-idw-works.htm
Kriging.
Mathematically and conceptually the most complex of the lot. But… can be the
best as you can get estimates of the accuracy of your interpolation. Exact
interpolator
Read the chapters and the Pro help:
https://pro.arcgis.com/en/pro-app/tool-reference/3d-analyst/how-spline-works.htm
The main kicker here is understanding the range and sill
Basically, I can try to explain the usefulness of kriging using a graph. You
plot semivariance (don’t worry about it) against distance (the distance
between your input points). In an ideal world, you get an upward curve,
then it flattens out. In the calculations, the software calculates the
relationships between every two input points – so it’s a processing pig.
The range is, statistically, the distance over which two points have a
statistical relationship to each other. Consider elevation – if two points are
5m apart, they are probably related. If they are 500 miles apart, they
probably have no relationship whatsoever. This statistic trys to get at that
relationship within your dataset.
If you don’t have a nice range and sill, your dataset is probably too sparse
and the accuracy measurements that come out of kriging will be low. If
you have a well defined range and sill, your accuracy will be high!
Figures swiped from the Pro page.
Here’s another excellent reference with nice pics:
https://gisgeography.com/kriging-interpolation-prediction/
OK Bob. But which do I use?
Welp… if you understand kriging and all the assumptions are met…. And you get
a good range and sill, I recommend this.
If you don’t meet the above criteria, I generally recommend IDW – as you are
getting more of an average surface. And you’re hoping the trends shown are
reasonably accurate.
Much like if you have a crummy DEM, I recommend the ESRI slope algorithm. But if
you have a good one, Max DH slope ftw.
And last….
This is a hard topic to teach with links and texts…. So, ask questions on the
discussions! I’ll do my best.
At one time, I understood kriging (my PhD advisor told me to explain and justify
using it in my dissertation. So.. Off to the Stacks I went. Figured it all out.
Realized my input data wasn’t good enough to support kriging. And just went
with IDW. It was kinda disappointing, frankly.
