# Do integer values in a Confidence Raster correspond with a specific probability %?

I have been using the Maximum Likelihood Tool (Spatial Analyst extension, ArcMap 10.3.1) to classify a number of rasters into "Vegetation" and "Bare Ground" classes. I'd like to have some sort of quantitative measure for how reliable each of my classifications are (to report along with my data), so I've been generating a Confidence Raster for each image.

I understand that the Confidence Raster assigns each pixel an integer value (1-14, highest-to-lowest) that corresponds with the certainty of that pixel's class assignment. But do these integer certainties correspond to a specific probability/percentage?

In the example image discussed in the Desktop Help documentation, it says that "Value 1 has a 100 percent chance of being correct," "Value 5 has a 95 percent chance of being correct," but pixels with Value 14 "have a 0.005 percent chance of being correct." This scale doesn't make sense to me, so I doubt that each Level indicates a discrete 1% drop in likelihood. Also, I haven't seen these statements supported in any other journals or forums I've found online, so I'm hesitant to assume that's the end of it.

Based on the help for the Maximum Likelihood Classification tool, and the text in the help page How Maximum Likelihood Classification Works, I'm pretty sure the codes match these values. The help page should list these - I dropped a request on their web page to include a table like this.

``````14 <= .005
13 >0.005 to 0.01
12 >0.01 to 0.025
11 0.025 (etc)
10 0.05
9  0.1
8 0.25
7 0.5
6 0.75
5 0.9
4 0.95
3 0.975
2 0.99
1 0.995 to 1
``````

No, the values do not correspond to a probability but rather a confidence region. The 14 values are, a rather arbitrary, set of nominal representations of each of the predefined confidence regions.

Analogously, think of a linear regression line with a set of confidence envelopes, with each incremental envelope indicating less certainty in the estimate. In this case the larger the value the higher the classification uncertainty, to the point of random (14).