Dataset Description
This repository contains the processed geospatial data used in the research paper "A Data-Centric Approach to Urban Building Footprint Extraction using Graph Neural Networks and Assessed OpenStreetMap Data."
The dataset focuses on a complex urban area in Algiers, Algeria, and addresses the challenge of label noise in Volunteered Geographic Information (VGI). It includes a high-confidence ground truth layer generated through a multi-source cross-validation pipeline (integrating OpenStreetMap, Google Open Buildings, and Overture Maps) and temporal NDVI stability analysis.
Contents:
Sentinel-2 Composite: A 15-channel multi-spectral composite image (resampled to 10m resolution) containing RGB, NIR, SWIR, and spectral indices (NDVI, NDBI, NDSI).
Curated Ground Truth Mask: A binary raster mask where 1 indicates a high-confidence building footprint and 0 indicates non-building. This mask was generated using the "Algorithm 1" reliability scoring method described in the paper.
Gold Standard Validation Set: A manually labeled vector dataset (GeoJSON) for a specific high-density urban patch, used for rigorous performance evaluation.
Superpixel Segments: Pre-computed segments generated using the Segment Anything Model (SAM) adapted for remote sensing.
This dataset is designed to benchmark Graph Neural Networks (GNNs) and other deep learning models on African urban morphologies.