RSMT: A Remote Sensing Image-to-Map Translation Model via Adversarial Deep Transfer Learning

Maps can help governments in infrastructure development and emergency rescue operations around the world. Using adversarial learning to generate maps from remote sensing images is an emerging field. As we now know, the urban construction styles of different cities are diverse. The current translation methods for remote sensing image-to-map tasks only work on the specific regions with similar styles and structures to the training set and perform poorly on previously unseen areas. We argue that this greatly limits their use. In this work, we intend to seek a remote sensing image-to-map translation model that approaches the challenge of generating maps for the remote sensing images of unseen areas. Our remote sensing image-to-map translation model (RSMT) achieves universal and general applicability to generate maps over multiple regions by combining adversarial deep transfer training schemes with novel attention-based network designs. Extracting the content and style latent features from remote sensing images and a series of maps, respectively, RSMT generalizes a pattern applied to the remote sensing images of new areas. Meanwhile, we introduce feature map loss and map consistency loss to reinforce generated maps' precision and geometry similarity. We critically analyze qualitative and quantitative results using widely adopted evaluation metrics through extensive validation and comparisons with previous remote sensing image-to-map approaches. The results of experiment indicate that RSMT can translate remote sensing images to maps better than several state-of-the-art methods.