Select (pure) training areas from each class. Then run supervised classification using, for instance, the maximum likelihood classifier. You may not get 100% separation of these 2 classes because there might be some spectral similarities. Should that happen, you can do post-classification processing using ancillary data, or consider using expert system as suggested by Cyril. Good luck.

I don't know why you classified with the unsupervised classifier. But if you have the option to perform supervised classifier, you should select training areas from barren and settlement part of your image and run the supervised classifier. However, there are many more possible solutions, but the one described above is the simplest.

one of the straight forward option is to implement such a task using supervise classification, but you should have training and validation data.

Hi,

You can do Alessandro's suggestion and increase the iteration.

The best way to separate these confused classes is to run an expert system classification with the use of ancillary data.

Increasing the number of clusters in an unsupervised classification can help but has limitations in that it will not entirely remove the problem

Select (pure) training areas from each class. Then run supervised classification using, for instance, the maximum likelihood classifier. You may not get 100% separation of these 2 classes because there might be some spectral similarities. Should that happen, you can do post-classification processing using ancillary data, or consider using expert system as suggested by Cyril. Good luck.

If  you use multispectral image, you may use also various band combinations to classify your image or calculate indexes such as build-up index..

I think you should use all suggested techniques.because areas you mention are often hard to separate

You may classify the datasets in more than 64 classes using unsupervised classification with principal axis, true cold option and 20 iterations. After that you can merge the classes using class attributes table by selecting the colors on swiped original satellite image.

Increase the number of required classes and reclassify.  Else use some supervised classification (Like MLC)

It will be worthwhile to read Cihlar (2000) where supervised and unsupervised classification methods are compared (section 3.2 pages 1101 - 1104). Though this paper is focused on regional scale mapping, some principles are applicable at other scales as well. 

Cihlar J (2000). Land cover mapping of large areas from satellites: Status and research priorities, International Journal of Remote Sensing, 21:6-7, 1093-1114

URL: http://dx.doi.org/10.1080/014311600210092

B) As others have suggested, one way to resolve the problem you are facing is by generating large number of clusters. Another possible approach is through multi-stage iterative classification methods. Wyman et al. (2001) has provided a good description of an iterative classification method (p. 1156-1157).

Wayman et al. 2001. Landsat TM-based forest area estimation using iterative guided spectral class rejection. PE&RS 67(10): 1155-1166.

URL: http://info.asprs.org/publications/pers/2001journal/october/2001_oct_1155-1166.pdf

this is usually not an easy task for mixed pixels especially with large pixel size.  however, I found the most efficient way is semi-manual digitizing and separating them, then you add them later to main classified map using your GIS skills!

New version of ENVI 5.5 has added this option in the software environment for users to edit classified images easily.

As suggested before, first use a hybrid supervised-unsupervised classification, then the semi-automated method by editing your classes manually using ancillary data in GIS.....,

This link may be helpful for you......watch entire video....

https://www.youtube.com/watch?v=eg0H_-mQO24