Ring-based Spatial Transformer: Learning Non-linear Spatial Interactions between Building Distribution and Pedestrian Flow

A novel ring-based SpatialTransformer model has been developed to analyze and predict pedestrian flow based on the distribution of building uses around railway stations. This approach, which applies self-attention mechanisms to learn complex spatial interactions without prior structural assumptions, consistently outperformed traditional Geographically Weighted Regression in predictive accuracy, particularly highlighting the influence of building features in mid-to-outer distance zones.

Why it matters

This research provides an advanced analytical tool for understanding and predicting human movement patterns in urban environments. It offers a more accurate method for assessing how land use and infrastructure planning influence pedestrian activity, which is crucial for urban development, transportation planning, and retail strategy.

Key insights

• A ring-based SpatialTransformer model was proposed to understand the relationship between building distribution and pedestrian flow.
• Concentric ring buffers at 100-meter intervals, extending up to 800 meters from railway stations, were used to define spatial tokens.
• The model utilized self-attention to directly learn non-linear spatial interactions between building uses at different distances.
• Testing around 100 randomly selected railway stations in Tokyo, the model used GPS-derived walking trip counts as the target variable.
• Across 30 independent trials, the SpatialTransformer consistently surpassed Geographically Weighted Regression (GWR) in predictive accuracy.
• SHAP analysis indicated that building features in mid-to-outer distance zones (beyond immediate vicinity) have a dominant influence on pedestrian flow.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2608.14660