Applications of generative methods in cartography on the example of creating pictorial maps

Research on the use of generative methods in cartography, visualization, and map generation is still in a very early phase.Schetinger et al. (2023) describe the potential of using generative text-to-image methods for the different phases of visualization, starting with 1) identifying relevant data, 2) converting data into visualized formats, 3) visualization and diagram recommendation, and 4) interaction and personalization of visual designs.The first experiments on map creation using a text prompt were conducted by Kang et al. ( 2023) with DALL-E 2. Prompt inputs of the following structure were used: "A {MapType} of {Region} on {Place} with {Description}".When comparing manually and AI-generated maps, four potential error categories were identified: inaccuracies, misleading information, unexpected features, and nonreproducibility.Observed inaccuracies can be due to unclear borderlines, inconsistent place shapes, and truncated content.Zhang et al. ( 2024) present a "MapGPT prototype" that uses voice input instead of interactions to control the map production process.In the background, 59 map modules are utilized for various map production tasks.Initial approaches in cartography attempt to use Neural Style Transfer (NST) to transfer graphic image styles to map representations.Bogucka and Meng (2019) experimented with NST to transform emotional image representations onto a topographic base map of Berlin.Christophe et al. (2022) used NST to transfer historical map styles (maps by Cassini, 18 th century and Etat-Major, 19th century) to orthophotos.With PictoAI (Drews et al. 2024), a custom GPT tool 1 integrated with DALL-E 3 was developed for the derivation of thematically appropriate pictograms.