Change Detection with Maps, but with a lot of classes
After going through the tutorial for Change Detection with Maps, I'm testing how to apply this to my project. T1 and T2 (taken 8 months apart with UAV) would be unsupervised classified maps with 5 classes. This imagery will be a pond with both emergent and submerged aquatic vegetation (using NDVI only wasn't a good metric).
Using the Level concept from the tutorial...
I've applied it to my own project where there may be 5 unsupervised classes:
Instead of positive or negative vegetation change, for my project it will be what did the class change to or did it remain the same classes. I have 5 classes that may not change (aka Class 1 remained Class 1), and 20 possible changed classes (Class 1 is now Class 2).
I am testing with NAIP imagery (same resolution/sensor) before my next UAV imagery comes in.
Am I going about this change detection with maps idea correctly?
Was this article helpful?
2 comments
Hi Shannon,
I think this is a workable solution, but maybe somethings to consider.
Depending on your end goal, it may not be optimal to have everything single type of change included, as it number of pairs of classes increases as the square of number of classes in the beginning. Meaning while this is somewhat manageable for 5 classes with 25 class pairs, increase it to 7 classes would mean almost doubling the number of class pairs at 49.
For this kind of upscaling of the number of classes, I would recommend to instead isolate exactly where the change has occurred first, and then extracting the most important bits of information later on.
That being said, I see no issue with your approach from a methodological point of view, as every possibility is taken into consideration. In my experience with vegetation, the most interesting bit is going from vegetation to non-vegetation. The switch from one type of vegetation to another is rare I find, but it may be the case for you.
Another quick comment would be that you should make sure the Class 1 in the first scene is the same Class 1 as the second scene (and for all the other classes), as for unsupervised classification there can be mismatches.
Mike
Thanks so much for the ideas!