Research map / Graph learning / Knowledge and temporal graphs
Estimating On-Road Transportation Carbon Emissions from Open Data of Road Network and Origin-Destination Flow Data
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes HENCE, a hierarchical heterogeneous graph neural network that estimates county-level on-road transportation carbon emissions from open road network and OD flow data instead of hard-to-collect vehicle-mileage statistics.
Problem
On-road transportation carbon emission estimation normally requires hard-to-collect vehicle miles traveled statistics, making large-scale estimation costly; the paper asks whether open road network and OD flow data can predict regional emissions accurately.
Approach
A three-level hierarchical graph (road network, community, region) with heterogeneous community- and region-level graphs whose spatial links (road connectivity) and OD links (travel demand) are fused by attention, trained to regress county-level DARTE emissions.
Key results
- HENCE achieves R^2 of 0.7502 (2015) and 0.7865 (2017), beating the best baseline Random Forest by 11.83% and 7.37% on R^2.
- Removing the region-level graph lowers R^2 by 17.46% on average in the ablation study.
- In temporal transfer (train 2015, test 2017), HENCE reaches R^2 0.7452 versus 0.6549 for the best graph baseline EGAT.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use HENCE instead of the strongest baseline, Random Forest | Regional on-road carbon emission estimation · DARTE 2015 | R^2: 0.6708 -> 0.7502 | Helped |
| use HENCE instead of the strongest baseline, Random Forest | Regional on-road carbon emission estimation · DARTE 2017 | R^2: 0.7325 -> 0.7865 | Helped |
| use HENCE instead of Random Forest | Regional emissions prediction under spatial transfer · West-east transfer, 2017 | R^2: 0.7008 -> 0.7811 | Helped |
| use HENCE instead of Random Forest | Regional emissions prediction under spatial transfer · North-south transfer, 2017 | R^2: 0.6298 -> 0.7267 | Helped |
| use HENCE instead of Random Forest | Regional emissions prediction under temporal transfer · Train 2015, test 2017 | R^2: 0.5644 -> 0.7452 | Helped |
| add spatial links to the heterogeneous graphs (ablation) | HENCE · DARTE 2015 and 2017 | R^2: 0.652 -> 0.750; 0.719 -> 0.787 | Helped |
| add OD links to the heterogeneous graphs (ablation) | HENCE · DARTE 2015 and 2017 | R^2: 0.689 -> 0.750; 0.706 -> 0.787 | Helped |
| add community-level modeling (ablation) | HENCE · DARTE 2015 and 2017 | R^2: 0.722 -> 0.750; 0.761 -> 0.787 | Helped |
| add region-level modeling (ablation) | HENCE · DARTE 2015 and 2017 | R^2: 0.601 -> 0.750; 0.668 -> 0.787 | Helped |
Surprising findings
- Plain Random Forest on aggregated features outperforms several graph learning baselines such as EGAT, EGNN, SortPooling and DiffPool on the in-distribution task.
Limitations
- Region-level attributes are refreshed once per epoch, a computational-cost compromise that may lag the joint representation.
- Evaluated only on US counties with a single emission database (DARTE) and two years of data.
Experimental setup
Datasets: DARTE on-road carbon emission database OpenStreetMap road network LODES origin-destination commuting flows
Benchmarks: DARTE 2015 county emissions DARTE 2017 county emissions
Metrics: R^2 MAE RMSE
Baselines: SVM Random Forest EGAT EGNN SortPooling DiffPool RFN
Builds on: EGAT DiffPool SortPooling EGNN
Closest papers by meaning
- Hugging Carbon: Quantifying the Training Carbon Emissions of AI Models at Scale (ICML 2026 · similarity 0.86)
- City-Scale Pollution Aware Traffic Routing by Sampling Max Flows Using MCMC (AAAI 2023 · similarity 0.84)
- OSM+: Billion-Level Open Street Map Dataset for City-wide Experiments (ICML 2026 · similarity 0.84)
- Towards Dynamic Spatial-Temporal Graph Learning: A Decoupled Perspective (AAAI 2024 · similarity 0.82)
- Satellites Reveal Mobility: A Commuting Origin-destination Flow Generator for Global Cities (NeurIPS 2025 · similarity 0.81)
- BO4Mob: Bayesian Optimization Benchmarks for High-Dimensional Urban Mobility Problem (NeurIPS 2025 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.