05-20-2019 09:00 PM
I have a thematic image from machine learning classification, I want to know how much accuracy of mine. I used Check Accuracy tool from the toolbar menu. I choose the image and stuck at the GLT Viewer and MAA Tool. thought that the concept of this function is to check accuracy an image depends on reference data such as T&E File/AOI/something like that? or how it supposed to be? what and how I can make a T&E file? is it strongly needed? and is there menu to create them?
Solved! Go to Solution.
05-20-2019 12:36 PM
Check Accuracy button actually starts the MAA tool (Metric Accuracy Assessment tool) for checking the geometric accuracy of an image based on ground control points.
For checking the accuracy of classification, what you need to use is the accuracy assessment tool availabe at Raster tab -> Classification group -> Supervised drop down menu.
We also have an spatial model operator named Assess Classification Accuracy and you may want to use that as part of the model that does the classification.
05-23-2019 07:29 PM - edited 05-23-2019 02:13 AM
OK, Thanks for the information,
I tried the accuracy assessment for my thematic image.
Before we load an image to do an assessment for the first accuracy assessment, the accuracy assessment viewer is empty (picture "Accuracy assessment empty window"), then I load an image, I did an assessment to the image that I've load and success so I closed the viewer.
But, when I loaded the same image because I want to do an assessment with other random points, the accuracy assessment viewer isn't empty (picture Accuracy assessment window").
Does that commonly happen to the image that has been assessment? how to clear that viewer?
And the last, I learned from a tutorial and help about accuracy assessment erdas imagine for a thematic image, I got one main concept about it, accuracy assessment is a tool for check the accurate a method depends on thematic image and sample point. The software has its own sample point (I chose "Random Point") and then they have each class type for each point, so as a user, we input a class of each point which class they were classified in the image to the "References" Column.
But, we can input the class of each point if we know where the point's location in the image. from the tutorial I see when the user input the class into the references column, he clicks the point in Accuracy Assessment Viewer and then it's directly going to the point in the image. How I could do that?
Here I attach the tutorial source that I've been followed.
05-24-2019 04:35 PM
If you load the same image for accuracy assessment with different random points, the software opens an Attention dialogue stating the classified file is already opened and gives the option “Close the file and open another”. If NO is chosen, the Accuracy assessment window remains as is. If YES is chosen, you’ll have the option to save the point table first and then select the thematic image again. At that point the Accuracy assessment window will be cleared.
Or after your first accuracy assessment is completed, you can simply Select All rows in the Cell Array, then Delete Selection to clear the Accuracy assessment window. Your previously loaded thematic image will still be active.
After adding the random points, the class they were classified in the image should not be added to the "Reference" column. The classes added to the Reference column should come from your ground truth image (or other ground truth reference source). In the tutorial, the user selects the same image for both reference and classified image. This example uses the classified image as ground truth to verify the same classified image. That’s really not the intended use. I assume the presenter of the tutorial is using the same image for the purpose of demonstrating how the accuracy assessment tool works.
Clicking the Reference point in the Accuracy Assessment Viewer does not drive to the point location in the image. It’s not obvious from the tutorial, but the presenter is manually driving and zooming to the reference point icon in the viewer.