Checking accuracy of supervised classification method in ArcGIS Pro

Question

I have a question about my image classification method in ArcGIS Pro 2.4.

I'm looking at a field of grain sorghum and doing a project that requires me to define certain features within the image for a conditional extraction. I began with this (multispectral) raster:

Using ArcGIS Pro's image classification wizard, I used training samples to classify the image into 5 different classes: Soils, shadows, Leaves, grain heads, and ground targets (coordinates are blanked out).

After computing the NDVI (and other indices), I then used a conditional statement to make a third raster to only include the "leaves" class, which is the only feature in the image I want to use.

How can I check the statistical accuracy of this method?

I've been looking into this, but being new to image classification, I'm not quite sure how to do this. This would be helpful so that I can compare it to other methods.

Answer 1

I found that in order to do this, you have to compute what's called a "confusion matrix." To do this accurately in a raster, you must have ground-truth points that you can create in a software like ArcGIS. I made about 50 points where I knew for certain that my points were my features, and then I added their features in an attribute table. From there, you can compute your matrix and look at the overall kappa value.