Well, a confusion matrix is a comparison of the classification results with the "truth." So, if you don't have a ground truth image, or ground truth regions of interest, then there really isn't any "truth" with which to compare the classification results. Do you have some kind of data that you would consider to be the right answers, to which you can compare the classification results? - Peg Exelis VIS

I don't think we have any data to use as the truth, apart from a few maps. But the map isn't detailed enough. Can we prepare the ground truth on our own, I mean is there any method through which the ground truth can be prepared?

The usual way to collect ground truth is to go on the ground in the study area, and collect information about which of your classes really exists in some set of points. Then you can create a set of ROIs (one for each class) containing those points, and use those as your ground truth. If your maps are very accurate, then you could create some "ground truth" ROIs. You would create a point ROI for each of your classes, with a set of points you find on your maps for each class. You would be assuming, though, that the map is 100% accurate. Another option, if you can get your maps in digital form, is to put them in ENVI classification format (one pixel value per class), and use them as "ground truth" images. Again, you would be assuming that the map is 100% accurate. Peg Exelis VIS