A method for compiling satellite image map geographic objects based on vector map data via deep learning.

Map compilation is a fundamental component of cartography. Apart from vector maps, satellite image maps also require compilation for specific purposes, such as enhancing visual clarity, concealing sensitive information, and ensuring consistency with corresponding vector representations. This paper proposes an automated method for satellite image map compilation based on deep learning and guided by vector map data, which consists of several key steps. First, scale-matched vector maps and satellite image maps are aligned and partitioned to generate training samples in the form of paired vector tiles and satellite image patches. Second, a base satellite image generation model with an encoder-decoder-type deep learning architecture is constructed and trained on these sample pairs to learn the mapping from vector data to realistic satellite imagery. Third, geographic objects designated for compilation are identified, and transfer learning is performed to fine-tune the base model, which improves its sensitivity to regions requiring modification. Subsequently, specific compilation operations-deletion, insertion, distortion, and displacement-are defined; corresponding vector features are then edited using analogous transformations before being input into the trained model to generate updated satellite images reflecting the intended changes. The proposed method enables both selective and operation-diverse compilation of geographic objects in satellite image maps. Experiments conducted using real-world datasets verify the capability of the proposed method to compile linear and polygonal geographic objects through various operations.