A survey on deep learning-based map generation via style transfer of visible light remote sensing image

Abstract

Map generation from remote sensing imagery is a critical task in urban planning and Earth observation. Recently, deep learning-based style transfer methods have emerged as a transformative approach due to the ability to handle complex image features and significantly enhance production efficiency. Unlike existing surveys that broadly cover general image-to-image translation, this paper provides a uniquely targeted review focusing explicitly on the intersection of style transfer algorithms and geographical map synthesis. We systematically trace the evolutionary path from traditional map-making methods and conventional image style transfer techniques to advanced deep learning models, explicitly linking how general feature extraction architectures have been adapted for the strict spatial constraints of remote sensing data. Furthermore, we critically analyze the evolution of generative adversarial networks (GANs)-based style transfer methods and the application in generating map images, categorizing them into text-free and text-annotated generation tasks – a crucial distinction representing the latest innovative frontier in this field. In this paper, the inherent disadvantages of current methods for generating map images from remote sensing imagery, such as incomplete retention of topological map information, blurred road edges, and unstable GAN training, are comprehensively evaluated. Finally, possible solutions and future research directions are proposed to offer new perspectives for advancing highly accurate and practically applicable map image generation through remote sensing style transfer.