Commuting Flow Prediction using OpenStreetMap Data

Official implementation of the paper:

Commuting Flow Prediction using OpenStreetMap Data, Computational Urban Science, Springer, 2025

Authors: Kuldip Singh Atwal, Taylor Anderson, Dieter Pfoser, & Andreas Züfle (George Mason University, Emory University)

---

Abstract

This repository contains the official implementation of the study "Commuting flow prediction using OpenStreetMap data." We propose a method to broaden the utility of state-of-the-art commuting flow prediction models by using globally available OpenStreetMap (OSM) data. While existing high-performing models often rely on location-specific proprietary datasets, limiting transferability, this approach leverages building types—specifically residential and non-residential classifications—derived from OSM to predict commuting flows. Our experiments demonstrate that models using these OSM-derived features achieve prediction accuracy comparable to those using region-specific data, while enabling application in data-poor regions through transfer learning.

---

Key Features

• OSM-Only Dependency: The model features are derived exclusively from OpenStreetMap (building types, road networks) and basic census data (population), eliminating the need for proprietary datasets like PLUTO.

• Building Type Indicator: Incorporates the count, density, and area of residential and non-residential buildings as key indicators for commuting mobility, which significantly improves prediction accuracy over standard OSM features.

• Transfer Learning Capability: The approach allows a model trained in one region (e.g., NYC) to be successfully transferred to another (e.g., Fairfax County) where ground truth commuting data may be unavailable, explaining up to 62.1% of flow variation.

• Model Benchmarking: Includes implementations comparing the proposed GMEL-OSM approach against Deep Gravity, XGBoost, and Random Forest models.

---

Methodology

1. Data Preprocessing and Building Classification

The pipeline begins by processing OSM data to extract building footprints.

• Building Classification: A machine learning approach is used to classify OSM building footprints into Residential and Non-residential types based on their geometric and topological features.

• Feature Generation: We derive nine specific input features for flow prediction:

  • Count, density, and area of residential buildings (3 features)
  • Count, density, and area of non-residential buildings (3 features)
  • Region population and population density (2 features)
  • Distance between census tracts (1 feature) calculated via Open Source Routing Machine (OSRM)

---

2. Flow Prediction Models

This repository supports the evaluation of several models using the generated features:

• GMEL (Graph Attention Networks): A geo-contextual multitask embedding learner that captures spatial dependencies between origin and destination regions.
• Deep Gravity: A deep neural network approach inspired by the classic gravity model.
• Baseline Models: XGBoost (Regression tree gradient boosting) and Random Forest.

Note: The original repositories for the key models used in this study can be found here:

• GMEL: https://github.com/jackmiemie/GMEL
• Deep Gravity: https://github.com/scikit-mobility/DeepGravity

---

3. Evaluation

The models are evaluated using standard metrics including Root Mean Square Error (RMSE), Coefficient of Determination (R²), and Common Part of Commuters (CPC). The repository also includes notebooks for visualizing:

• Choropleth Maps: Visualizing relative prediction errors for inflows and outflows across census tracts.
• Scatter Plots: Comparing ground truth vs. predicted commuters (log-log scale).
• Histograms: Analyzing the distribution of relative prediction errors.

---

Dataset

The model is trained and validated on two distinct study areas:

1. New York City (NYC), USA: An urban environment with high transit density.
2. Fairfax County, Virginia, USA: A suburban environment used to test transferability.

Data Sources:

• Input Data: OpenStreetMap (OSM) and U.S. Census TIGER/Line shapefiles.
• Ground Truth: Longitudinal Employer-Household Dynamics (LODES) Origin-Destination Employment Statistics (2015).

---

The main Jupyter notebook, Flow_Prediction.ipynb, contains the pipeline for:

• Processing ground truth flows from LODES
• Analyzing predictions using standard OSM features
• Analyzing predictions using GMEL/Building Type features
• Generating visualization metrics (bins, deltas) and plots (histograms, scatter plots, maps)

---

Citation

If you use this code, methodology, or data in your research, please cite:

@article{atwal2025commuting,
  title={Commuting flow prediction using OpenStreetMap data},
  author={Atwal, Kuldip Singh and Anderson, Taylor and Pfoser, Dieter and Z{\"u}fle, Andreas},
  journal={Computational Urban Science},
  volume={5},
  number={1},
  pages={2},
  year={2025},
  publisher={Springer}
}

---

Acknowledgements

This work is supported by the National Science Foundation Grant No. 2109647 titled “Data-Driven Modeling to Improve Understanding of Human Behavior, Mobility, and Disease Spread”.

Computing resources were provided by the Office of Research Computing at George Mason University.