You can use either GTs or you can use the existing classified image where you can select the known pixels randomly against each of the land use classes.

Are you classifying high or low resolution imagery. What is your classification precision? Is it detailed (like precision agriculture mapping, detailed urban mapping) or is it broad (like built up versus forests)?

Although not ideal, you can select ground truth pixels from the image you are classifying if you can clearly visually identify your classes of interest on the image. They have to be completely independent of the training samples and you can minimize the subjective bias by using random sampling and more than one analyst perspective. 

If you are classifying low resolution data (e.g. Landsat), try to find out if you can get high res. data over the same area. Less ideally (for academic purposes anyway), you can look up if google earth has high res.image over the area. You must however verify the time stamp on the GE imagery ( I believe using the historical  tool), to make sure it corresponds with your classification image

thank you Julius, here i am trying to classify Landsat-7,8 images and i got those image data but my doubt is that there should be a ground truth image for reference so that we can classify this is  land, this indicates water, forest and so on...i collected some images from  landsatlook viewer but there is no any reference map, if possible please provide me dataset with reference map.

http://landsatlook.usgs.gov/

@chutiya, Sir i am trying to classify image with unsupervised classification where classes is not known....can i define classes based on image observation i mean what type of class can be created just my manually seeing the image?

Normally, you should have Ground Truth Data. But if you don't, you should try to find known points from other sources (e.g. Google Earth) or better from G.P.S. points taken by you. You can take a look at:

Congalton, R. G. and Green, K. 2009. Assessing the accuracy of remotely sensed data: Principles and Practices. 2nd Edition. CRC Press, Taylor&Francis Group, Boca Raton.

I was looking for ground truth images too but could not find . Someone said 

Arcgis can do that

Gautam Kumar , have you published this work? i would love to read it ( i only found your "An Image Processing Tool to Generate Ground Truth Data from Satellite Images using Deep Learning" on "towardsdatascience.com)