Modeling Functional Urban Centrality via Graph Attention Networks and Multi-Source Spatial Data Fusion

Conventional spatial analysis methods heavily rely on static morphology and Euclidean distance metrics, failing to capture functional urban centrality, which is inherently dynamic, flow-driven, and non-Euclidean. This study presents an unsupervised Graph Attention Network Autoencoder (GAT-AE) framework to model and quantify functional urban attractiveness by fusing multi-source geospatial indicators—including high-resolution satellite imagery, points of interest (POI), and VIIRS nighttime lights—with physical street network topology. Discretized using Uber’s H3 hexagonal grid at resolution 9, the territory is structured into a functional spatial graph where edge weights explicitly couple topological shortest-path distances (computed via Dijkstra's algorithm) with gravity-based spatial interaction dynamics. Benchmarked against non-spatial baselines (linear PCA and a dense MLP autoencoder), the GAT-AE demonstrates superior representational capability within its 8-dimensional latent space. Although the model makes a deliberate algorithmic trade-off by accepting a higher raw reconstruction loss (MSE=0.0104) to enforce topological message passing, it substantially optimizes spatial coherence (Global Moran's I=0.973) and cluster separability (Silhouette Score=0.391 vs. 0.263 for PCA; Calinski"-" Harabasz Index=18,419). Bivariate spatial analysis confirms that the attention mechanism effectively isolates complex non-linear urban phenomena—specifically identifying enclaved zones (high physical density with low topological connectivity) and topological anomalies (low physical density serving as high-throughput functional nodes)—that traditional linear models completely miss. Externally validated against localized real estate land values (r=0.473,p