Improving the potential of pixel-based supervised classification in the absence of quality ground truth data