Changing overall accuracies in RF and ML classifiers

Dear all,

I am using SNAP to do classification by using Random forest and Maximum likelihood classifiers. One thing I’ve figured out is that any time that I do classification and calculate overall accuracies; overall accuracies change although it is not much (maybe 2 or 3 percent) but it is weird.

Any body know why this is happening?


Are you applying Random Forest algorithm with exactly the same parameters on the exactly same image very time and the exactly same training data?

Thanks for answer.
yes but I have 7 classes and I choose almost 2300 pixels for any class for training but I choose '10000 ’ in ‘number of training samples’. This can be the reason that I am getting different answers?
Maybe I should choose (7*2300=16100) pixels in ‘number of training samples’ and by this way, the result can be in somehow fixed. Is not it?

And another thing:
I do not know how RF works here? did it chose randomly pixels for all classes OR it chose 2300 for any class?

The number of training samples is the number of pixels inside your polygons (the ones used to train the image) that the algorithm select for training. you can leave this as is, or you try to increase this number a bit and see if you get better results.

In number of trees in theory, the more trees you have the better the accuracy of your results. However, at some point, the improvement decreases as the number of trees increases. So, try and use a number of trees ~100.