RNMV‐TFormer: A Morphology‐Aware Geospatial Representation Framework for Urban Function Zone Prediction Integrating Road Network Structure and Multi‐Source Urban Features

Abstract

Accurate prediction of urban function zones is essential for effective spatial planning and urban governance. Although road network characteristics have been widely incorporated into existing studies as key indicators for urban function zone prediction, they are often simplified into distance‐ or density‐based descriptors, such as road density and accessibility. Such simplified representations fail to capture the intrinsic structural semantics of road networks, thereby limiting model performance. To address this issue, this study proposes RNMV‐TFormer, a morphology‐aware framework that integrates road network morphology with multi‐source urban geospatial features. The framework encodes the geometric and topological characteristics of road networks and combines them with residential, commercial, transport‐related, and locational features in a unified grid‐based learning framework. Using multi‐period datasets from Chengdu, China, the proposed method achieved a validation accuracy of 93.8% and a macro‐F1 score of 92.89%, outperforming Random Forest and convolutional neural network (CNN) baselines. Ablation experiments further showed that removing road morphology features reduced validation accuracy and macro‐F1 by 4.45 and 4.89 percentage points, confirming the importance of structural road semantics for urban function zone prediction.