In my opinion you will have to mask out cloud covered pixels. I have attached a paper that appears to have solution to identification of cloud covered pixels ( Otsu's threshold) in first half of their methodology.  Hope this helps.

http://www.a-a-r-s.org/acrs/administrator/components/com_jresearch/files/publications/SC02-0828.pdf

Sujata,

Perhaps some thing to try is use only non-visible bands in the classification because the nir, but especially the swir bands,  will more than likely penetrate thru the thin clouds and not see them at all.  For vegetation, the swir bands are typically highly correlated to the visible bands and might end up generating very similar, or perhaps, better results.

Good Luck,

Pat

If this patch of clouds remains after the atmospheric correction, you may try to mask the cloudy areas and use another clear image, close in time, for these areas. This solution indeed depends on the study purpose.

this image shows a part of uttarakhand state near about tehri dam, its need atmospheric correction. but after atmospheric correction the image can not produce any Indices... 

https://www.youtube.com/watch?v=suKHU-unEFA

Hello Again,

I am the person who developed the COST and DOS1 atmospheric correction methods and wanted to follow up with a short comment (you can see papers related to these in my research gate profile).  Keep in mind that most, probably all, atmospheric correction methods apply an additive, and in some cases like the COST method also a multiplicative, correction using the same value for the ENTIRE image.  They are NOT pixel-by-pixel corrections so the thin clouds will NOT be removed.  Since the correction methods are entire image based ones they generally will not change the relationships of the various spectral clusters in the image and will basically just move the clusters around in a linear fashion.  Since the swir bands (bands 5 and 7 in Landsat 5 and 7 systems), and somewhat the nir band, will penetrate thru this type of thin clouds I still recommend that you investigate using these bands only to see what kind of results you will get (i.e., do not use the visible bands in the analysis).

Pat

I see the Semi- Automatic classification plugin in Q-gis as being easy and effective. You could try that to. It works quite well.