Multi-spatial urban function modeling: A multi-modal deep network approach for transfer and multi-task learning
Understanding dynamics of urban land-use is crucial for comprehending urban spaces and evaluating planning strategies. A range of data-driven models based on the representation learning of multiple data sources have focused on extracting spatially explicit characteristics at the feature level for ur...
Saved in:
| Main Authors: | , , , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2025-02-01
|
| Series: | International Journal of Applied Earth Observations and Geoinformation |
| Subjects: | |
| Online Access: | http://www.sciencedirect.com/science/article/pii/S1569843225000445 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Understanding dynamics of urban land-use is crucial for comprehending urban spaces and evaluating planning strategies. A range of data-driven models based on the representation learning of multiple data sources have focused on extracting spatially explicit characteristics at the feature level for urban function inference. However, they commonly pay no attention to the systematic relationships between urban land-use and transportation as the core components of urban systems. Consequently, while performing relatively well in urban land-use recognition, these models cannot transfer across various urban tasks and have almost no generalizability in integrated urban modeling. Guided by the theory of integrated land-use and transport modeling, this study proposes a multi-modal deep learning model to leverage the systematic relationships between urban components. First, spaces of urban places, urban forms, urban flows, and urban locations are conceptualized from the interactions between land use and transportation systems and represented by multi-source heterogeneous spatial features. Second, to account for both direct and indirect interactions in these spaces, a Deep & Wide network is introduced to fuse different modalities of spatial features. Using Shenzhen city as a testbed, extensive experimental results show that our approach improves accuracy by 13.3% compared to state-of-the-art models. We further validate the superior generalizability of our approach across various urban tasks, such as predicting urban land-use, housing prices, and population density, over other baselines. Enabling such a doubly-informed framework of urban theory and AI, this study provides a pilot demonstration for the new scientific paradigm of AI for Urban Science and Modeling. |
|---|---|
| ISSN: | 1569-8432 |