In semantic segmentation of satellite images, the task to acquire accurate label is tedious and time consuming. Automation of such tasks turns out to be challenging due to the fact that existing labels is generally quite coarse and noisy. Weakly-supervised learning recently proposed by Adrien and Hicham et al [1] shows some improvement on accuracy but is still unsatisfactory in practice. In this work, we propose to constrain the segmentation network such as U-Net [2] on multi-level convolution neural networks (CNNs) for binary urban semantic segmentation in satellite images to filter noisy labels.