Flexible style transfer from remote sensing images to maps

Style transfer has emerged as a prominent technique for transferring stylistic elements between images (e.g., a reference image and a map). However, current methods face two challenges when applied to create image maps, especially when the reference image and map are not spatially aligned (e.g., covering different regions). These challenges include aligning the semantic elements between maps and remote sensing images, and then balancing the photorealistic textures with cartographic symbolism to maintain cartographic quality. To address these challenges, we propose a flexible style transfer method from remote sensing images to maps, relaxing the requirement of strict spatial alignment between remote sensing images and maps. Our approach enables the generation of image maps with adjustable stylistic results, offering a balance between photorealism and symbolization. First, we analyze the semantic of the input map and the reference imagery including semantic classes and semantic relationships encoded by colors. Then we implement hierarchical control and parameter interpolation to enable style matching. We also compare the transfer results of our method to those of the baseline image style transfer methods across four aspects including visual similarity, graphic discriminability, semantic consistency, and overall readability. The evaluations show that our approach significantly enhances cartographic quality by flexibly balancing photorealism and symbolization, while offering the flexibility to generate image maps with varying preferences.