Accurate urban building footprint data are essential for sustainable urban planning (SDG 11). This study introduces a novel framework that overcomes the dual challenges of extracting detailed features from medium-resolution Sentinel-2 imagery and the inherent quality issues of crowdsourced OpenStreetMap (OSM) data. We achieve this by integrating Graph Neural Networks (GNNs) with a rigorous, multi-source data assessment pipeline.

Our objective was to evaluate an UrbanGraphSAGE GNN architecture for segmenting building footprints in Algiers. A foundational component was the creation of a high-confidence ground truth dataset by cross-validating OSM data against Google Open Buildings and Overture Maps, followed by a temporal stability analysis. Our model departs from standard CNNs by first segmenting the imagery into superpixels, which then serve as nodes in a graph. The GNN classifies these nodes by learning from both their spectral features and their spatial context, allowing it to model complex urban morphologies effectively.

The results confirmed the framework's robustness. The final augmented model achieved a strong Test F1-Score of 0.7579 and an excellent recall of 0.9192. This high recall is critical for creating comprehensive urban inventories, as it minimizes the number of missed buildings.

This study validates a powerful framework for leveraging GNNs and rigorously assessed open data for urban monitoring. The methodology offers a scalable and low-cost solution for creating reliable building footprint datasets, providing a valuable tool for planners in rapidly urbanizing cities.