Automated creation of pictorial maps using Retrieval-Augmented Generation (GenRAGmap)

• Applicant: Professor Dr.-Ing. Dirk Burghardt
• Subject Area: Geodesy, Photogrammetry, Remote Sensing, Geoinformatics, Cartography
• Term: since 2025
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 567521318

The aim of the project is to explore the potential of generative artificial intelligence for map production. By combining generative methods and Retrieval-Augmented Generation (RAG), an automated workflow for creating pictorial maps is to be developed. The use case is to create hand-drawn style pictorial maps for tourism in cities that take current information and user preferences into account. The workflow comprises several phases, starting with the integration of external geo and points-of-interest (POI) data sources. Current spatial data from OpenStreetMap, Wikivoyage, Wikidata and geotagged social media content is retrieved to identify relevant tourist attractions and events. This data is fed into the generation process to enable a precise and contextualized representation of the information collected. The second phase of the workflow involves optimizing prompt engineering, in which structured geo-data is converted into narratively formulated texts, which are then processed by generative open-source models such as Stable Diffusion or DeepSeek-VL2. Text-to-image methods will be used to generate pictorial map symbols. The use of natural language processing (NLP) and image generative models will create a seamless connection between textual and visual information. Both pre-trained models will be used and training with specific cartographic or hand-drawn styles will be carried out to create new, customized Low-Rank Adaptation (LoRA) models. Image-to-image-based methods will be used to embed the pictorial map elements in the base map. Research will be carried out into how seamless integration of foreground and background information can be achieved. Another key feature of this approach is the ability to personalize map creation. User requests, such as for "family-friendly attractions" or current events, are supported by retrieving live data and personalized preferences to generate individually tailored, interactive maps. The use of RAG makes it possible to integrate current, relevant and contextualized data in real time, resulting in dynamic and user-specific map creation. The evaluation takes the form of usability studies. For this purpose, pictorial maps are used that are adapted for different cities, topics and user groups. In addition, a comparative analysis is carried out with manually created pictorial maps. User satisfaction and the acceptance of generative map representations are assessed, thus gaining insights for the further development of generative map creation methods.

• DFG Programme: Research Grants