Generative Transformer Models for the Cartographic Generalization of Vector Data Incorporating Spatial-Semantic Context Information (Cart2Former)

• Applicant: Professor Dr.-Ing. Martin Kada
• Subject Area: Geodesy, Photogrammetry, Remote Sensing, Geoinformatics, Cartography
• Term: since 2026
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 577232574

Cartographic generalization aims to optimize map content both spatially and semantically while considering cognitive aspects. However, its processes and operators are difficult to formalize and implement algorithmically. At the same time, the availability of large data volumes opens up new opportunities to learn generalization processes using artificial intelligence, particularly deep neural networks. This project develops generative Transformer models and trains them to address cartographic generalization processes for the vector representation of map data. To this end, specific operators for feature extraction, as well as positional, structural, and contextual encodings and adaptive pooling mechanisms, are investigated and further developed. On this basis, Transformer blocks are conceived that enable hierarchical feature extraction directly from geometry coordinates and thereby allow for an efficient embedding of the inputs. These blocks serve as the foundation for designing Transformer architectures that realize generative models for key generalization processes such as shape simplification, object aggregation, and displacement in the case of spatial conflicts for polygonal and polyline geometries, as well as thinning for network data. Special attention is given to autoregressive approaches that generate geometries sequentially, building on previous steps. This facilitates the preservation of shape characteristics and symmetries. Spatial and semantic contextual information is explicitly integrated. In addition, the project investigates how the Transformer models can be pretrained in a self-supervised manner on large amounts of vector-based map data to enhance their transferability. In this way, models are created that can be applied to various downstream tasks without relying on large, annotated datasets.

• DFG Programme: Research Grants