Towards Graph-Based Deep Learning for Map Generalization: Insights from Building Footprints Simplification and Aggregation

Research Approach

This study explores the use of graph-based deep learning to automate map generalization, specifically focusing on the simplification and aggregation of building footprints. The researchers propose a unified framework that treats building polygons as graphs, where vertices are represented as nodes and boundaries as edges. By reformulating the generalization process, the authors define simplification as a node movement prediction task (adjusting vertex positions) and aggregation as a link prediction task (determining how adjacent buildings should be merged).

Methodology

To evaluate this approach, the authors tested three common graph neural network (GNN) architectures—GCN, GAT, and GraphSAGE—on multi-scale building datasets (1:10,000 to 1:25,000 and 1:10,000 to 1:15,000). The data preparation involved transforming building polygons into graph representations and generating ground truth labels by calculating node displacement vectors and updating adjacency matrices to match target map scales. The models were trained using a combination of intrinsic geometric features (such as turn angles and segment lengths) and relative spatial features (such as distances to boundaries).

Key Findings

The results indicate that while graph-based deep learning is a viable path for automated generalization, the two tasks present different levels of difficulty. GraphSAGE outperformed other architectures in link prediction, showing promise for automated aggregation. However, the study reveals that precise node movement prediction remains a persistent challenge. Overall, the findings suggest that aggregation is significantly more complex than simplification, as it requires the model to capture higher-level spatial relationships and maintain topological coherence between multiple entities.

Automated map generalization is a critical bottleneck in cartography, traditionally relying on rigid, rule-based algorithms that struggle with the diversity of urban morphologies. By moving toward a data-driven, graph-based approach, this research provides a methodological foundation for creating more flexible, intelligent systems that can adapt to various geographic contexts without the need for manual parameter tuning.