If you have 400 points of accuracy data, your confusion matrix should have the 400 points spread between the classes. Example:
Urban 50 60
Non-Urban 90 200
Notice that the 4 numbers add up to 400. Only the diaganol is where an urban pixel in the reference data was also classified as urban (50) and your classification of non-urban areas was also matched with a non-urban reference pixel (200). 250/400 = overall accuracy. Do a search on producer's and user's accuracy to figure them out and then also Kappa Coefficient. But I would also recommend you look at techniques for comparing your classification results with your reference data to create a matrix where all the data in your results add up to 400.
The squared R approach is favored by many remote sensing analysts, but I think you are trying to do the approach I discussed above. I hope this helps.