1

In classification of the imagery, is there a way of leaving certain image objects unclassified? Say for example, if the classifier is uncertain what class to classify the image object as, does eCognition have functionality to leave the image objects unclassified? I'm not sure if that is clear, but let me know if you all have any ideas. I've looked into hierarchical classification, however, I am using a ML classifier, RF, to do the classification, which is separate from the hierarchical classification algorithm. The eCognition Community site has been down for several weeks now, and there is sparse information across the web, as well as the User Guide, which is not very helpful. Thanks in advance.

1 Answer 1

2

If you are using a random forest classifier then the inputs/outputs should be derived from that and not eCognition. In that case, RF will not be able to decide to leave data unclassified as far as I know. RF works with an ensemble approach and will make more decision trees to compensate. You would need a classifier that can accept a threshold value (of class mean or standard deviation) to allow for null in your classification, but I do not know of one like that.

Depending on what your data look like, you could conduct an additional classification on the confused areas using the hierarchical classification. If for example you are getting significant confusion between agriculture and green spaces, you could classify the whole image (including all possible classes, but grouping agriculture and green space) and then take your mixed/confused class and try and run an additional, separate algorithm on that (not necessarily another RF).

2
  • So in the RF classification, I would obtain broad classes, say, impervious, pervious, water, and vegetation. Then in the secondary classification, using the confused class, say, vegetation, I would then perform, for example, nearest neighbor. The thresholds would differentiate the vegetation, into, in my case, 2 different vegetation species, and leave the rest as vegetation, or what I would consider, the unclassified vegetation species. Does that sound about right? Or would you look into other algorithms? Commented Feb 21, 2020 at 18:11
  • Yes, that's how I envisioned doing it. For most classifiers, you will need to think of your 'unclassified vegetation' as just another broad class that you cannot ground truth or define as well as others.
    – Masjo
    Commented Feb 25, 2020 at 18:31

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.