Detecting shadows in high-resolution satellite images is a challenging task due to the fact that shadows can easily be mistaken for low reflectance soil or water and that such images have limited spectral bands. In this work, we propose a semantic level shadow segmentation by using generative adversarial networks and created a dataset of pre-processed images for training, validation and test. In this way, we trained a generator network that produces shadow masks with condition on a satellite image patch and tries to fool a discriminator, which is trained to discern if a given mask comes from the ground truth or from the generator model. The results achieve an accuracy of 95.85% and a Kappa coefficient of 91.76%, which is superior to the compared methods.