Objectives  The rapid development of cartographical and communication technologies makes the public free to create, publish, edit and share their map image resources and products with various platforms and tools. These map images are remarkably ubiquitous in terms of map content, mapping standards, and other aspects, which poses a big challenge to establishing massive well-annotated map image data. Thus, although the state-of-the-art deep learning methods have made a breakthrough in recognizing the standard‍ized maps, they are still intrinsically unqualified to effectively address map image recognition and understanding due to the inadequate well-labeled map images.

  Methods  This paper summarizes the progress and challenges regarding map image recognition and discusses the theoretical configurations and potential geospatial artificial intelligence (GeoAI) techniques for efficient map image recognition and understanding. We propose the map features for deep learning models to represent map image content. Then, we explore the promising methodologies for map image content recognition and the possible semantic analysis methods for map image understanding. Subsequently, we prospect several implementations regarding map image recognition and understanding and their future potentials.

  Results  In our opinion, further investigation on theories and methods for map image representation is essential. Moreover, fully utilizing the values of map im‍ages depends upon recognizing the explicit content (map image perception) and mining the hidden semantics (map cognition).

  Conclusions  We hope our exploration can contribute to the cartographical community offering a GeoAI and data representation integrated perspective on map image utilization.