Graph convolutional network for modeling corridor-based urban transformation: Revealing spatial coherence in land-use change in Erbil.

Abstract

Residential-to-commercial transformation often occurs selectively along particular urban corridors, yet the structural mechanisms underlying this spatial concentration remain insufficiently understood. This study examines whether the historical configuration of Erbil's street network explains corridor-based land-use change between 2004 and 2024. A reconstructed 2004 street network comprising 279 nodes was analyzed using weighted degree, betweenness, closeness, PageRank, and straightness centrality. Logistic Regression and a two-layer Graph Convolutional Network (GCN) were trained using identical predictors and evaluated through a stratified 70:30 split, 30 repeated random splits, conventional classification metrics, and spatial-coherence indicators. In the main split, the GCN achieved higher accuracy, precision, and F1-score, whereas Logistic Regression obtained slightly higher recall and ROC-AUC. Repeated validation confirmed that Logistic Regression remained more sensitive to transformed nodes, while the GCN produced fewer connected components, lower fragmentation, and higher largest-connected-component ratios. Weighted straightness emerged as the strongest and most consistent predictor of commercial conversion. The findings demonstrate that historical street-network structure meaningfully shapes later corridor transformation and that model evaluation should extend beyond node-level accuracy to include spatial continuity. The proposed framework offers a transferable tool for anticipating commercial expansion and supporting proactive corridor planning decisions.