It is difficult to quantify geographic features with a certain character or pattern, especially taking spatial variability and heterogeneity into account. Landscape exhibits various patterns and processes in different scales. However, it seems that CycleGAN is able to extract some general features from spatial distribution of city structures.
To examine whether it is possible for CycleGAN to extract city styles and transfer the potentially learned style to other place, Corvallis is used as a sample for generating satellite images. View the real Corvallis Here.
Different levels of details are generated for single tile with different cities' style. For example,the fake image with Beijing’s style has most detailed information of tall buildings and shadow around them.
To examine the stability of the training model, fake satellite images are generated in a relatively large-scale area in Corvallis. Tiles covering Corvallis are put together using mosaic method. View Corvallis with different cities' styles.