Hello Jad Tahouri

For calssification

I would create a testing dataset based on visual interpretation of satellite images, use can use ROI option in ENVI. Then use it for testing accuracy instead of the ROI used for classification.

Try to select well distributed and unbaised ROI and independent of locations used in training the classfication.

For modeling:

This answer may be very subjective to the modelling approach.

1) if your model produce probabilities or soft results, go with ROC curve analysis.

https://towardsdatascience.com/understanding-auc-roc-curve-68b2303cc9c5

2) If you want to evaluate with exact prediction, used confusion matrix, sensitivity, specificity, precision.

https://towardsdatascience.com/accuracy-paradox-897a69e2dd9b

However, evolution of model performance is subjective, because "Models are abstraction of real systems".

Hope this answer will help you

Best regards