Ground truth data is very helpful to verify classifications. If you had hyperspectral imagery you could determine crops by spectral signatures but that is not the case here. If you have access to someone who can visually check the crops that would be the best but there are other means as well such as local farmer cooperatives or, other agricultural organizations that may have access to that information.

If you had/have ground-truthed points that you used in the creation of your classification you could also consider doing a cross-validation evaluation of accuracy (if visiting the sites is not feasible for some reason). This is typically done by taking a random sample of your ground-truth points for each class and removing them from the training set and then using them as ground-truth points for the accuracy assessment. For example, you could use 80% of the sites per class for training and classification and then the other 20% for the cross-validation evaluation (the percentage you would use could vary I supposed, depending on how many ROIS you have for each class). You would need to create 2 sets of ROIs, one for training and one for accuracy for each class, which can be a pain as ENVI doesn't have a built-in tool for taking a random sample from each class seperately. ENVI does have an option under Classification -> Post-Classification -> Confusion Matrix -> Using Ground Truth ROIs for this type of cross-validation.