Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Improving Local Air Quality Predictions Using Transfer Learning on Satellite Data and Graph Neural Networks

Finn Gueterbock, Raul Santos-Rodriguez, Jeffrey N. Clark

TL;DR

This study tackles the problem of predicting $NO_2$ at unmonitored urban locations by combining satellite-derived data, meteorological features, and sparse ground measurements within a GraphSAGE-based inductive framework. It introduces autoregressive temporal modeling and a transfer-learning scheme pre-trained on London and fine-tuned on Bristol to create high-resolution 'virtual sensors'. The transferred GraphSAGE model achieves notable improvements over Bristol-only baselines, with $NRMSE$ reductions of 8.6% and $Grad{-}RMSE$ reductions of 32.6%, demonstrating the viability of cost-effective, scalable air quality monitoring in data-scarce regions. The approach has practical implications for public health and climate planning and can be extended to other pollutants and larger regions.

Abstract

Air pollution is a significant global health risk, contributing to millions of premature deaths annually. Nitrogen dioxide (NO2), a harmful pollutant, disproportionately affects urban areas where monitoring networks are often sparse. We propose a novel method for predicting NO2 concentrations at unmonitored locations using transfer learning with satellite and meteorological data. Leveraging the GraphSAGE framework, our approach integrates autoregression and transfer learning to enhance predictive accuracy in data-scarce regions like Bristol. Pre-trained on data from London, UK, our model achieves a 8.6% reduction in Normalised Root Mean Squared Error (NRMSE) and a 32.6% reduction in Gradient RMSE compared to a baseline model. This work demonstrates the potential of virtual sensors for cost-effective air quality monitoring, contributing to actionable insights for climate and health interventions.

Improving Local Air Quality Predictions Using Transfer Learning on Satellite Data and Graph Neural Networks

TL;DR

This study tackles the problem of predicting at unmonitored urban locations by combining satellite-derived data, meteorological features, and sparse ground measurements within a GraphSAGE-based inductive framework. It introduces autoregressive temporal modeling and a transfer-learning scheme pre-trained on London and fine-tuned on Bristol to create high-resolution 'virtual sensors'. The transferred GraphSAGE model achieves notable improvements over Bristol-only baselines, with reductions of 8.6% and reductions of 32.6%, demonstrating the viability of cost-effective, scalable air quality monitoring in data-scarce regions. The approach has practical implications for public health and climate planning and can be extended to other pollutants and larger regions.

Abstract

Air pollution is a significant global health risk, contributing to millions of premature deaths annually. Nitrogen dioxide (NO2), a harmful pollutant, disproportionately affects urban areas where monitoring networks are often sparse. We propose a novel method for predicting NO2 concentrations at unmonitored locations using transfer learning with satellite and meteorological data. Leveraging the GraphSAGE framework, our approach integrates autoregression and transfer learning to enhance predictive accuracy in data-scarce regions like Bristol. Pre-trained on data from London, UK, our model achieves a 8.6% reduction in Normalised Root Mean Squared Error (NRMSE) and a 32.6% reduction in Gradient RMSE compared to a baseline model. This work demonstrates the potential of virtual sensors for cost-effective air quality monitoring, contributing to actionable insights for climate and health interventions.
Paper Structure (17 sections, 3 figures, 2 tables)

This paper contains 17 sections, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methods
  4. Data
  5. Surface NO$_2$ Measurements
  6. Satellite Data
  7. Additional Features
  8. Models
  9. Results and Discussion
  10. Impact and Implications
  11. Conclusion
  12. Appendix
  13. Data
  14. Model training
  15. GraphSAGE
  16. ...and 2 more sections

Figures (3)

  • Figure 1: Maps showing the spatial distribution of NO$_2$ sensors locations in (a) Bristol and (b) London accessed during this study. In (a), the two Bristol sensor locations marked in orange are those for which temporal NO$_2$ predictions are presented in \ref{['fig:predictions_subfigures']}.
  • Figure 2: Actual and predicted NO$_2$ values from the transferred GraphSAGE model for two locations in Bristol: (a) Well's Road - a location with typically low NO$_2$ values, and (b) Colston Avenue - a location with typically high NO$_2$ values.
  • Figure 3: NO$_2$ values at time $t$ vs NO$_2$ values at time $t+1$.