The creation of super-resolution aerial imagery is a very promising, but challenging task that has attracted attention in recent decades. Deep learning models help to solve this problem, but blurry texture and contour artifacts lead to reduced human perception of the resulting images. This study aims to improve the semantic aspect of super-resolution images with a specific learning strategy, when the generator of generative adversarial network (GAN) is learned using semantic features extracted by semantic segmentation network. Thus, the summarized loss function becomes more complicated. The proposed approach has been trained and tested on the UAV annotated DOTA, UAVid, and VisDrone datasets with good segmentation and perception results.