Interpretable and Transferable Models to Understand the Impact of Lockdown Measures on Local Air Quality

Johanna Einsiedler, Yun Cheng, Franz Papst, Olga SaukharXiv 2020 · arXiv

Abstract

This study introduces a framework utilizing Generalized Additive Models (GAMs) and transfer learning to isolate the impact of COVID-19 lockdowns on local air quality. By disentangling meteorological influences from human activity, the researchers quantify pollution reductions across Europe and China, demonstrating that transfer learning can effectively model short-term policy shifts with limited data.

Introduction

Measuring the precise impact of lockdowns on air quality is challenging due to the confounding effects of weather. Simple year-over-year comparisons (e.g., 2020 vs. 2019) fail to account for variations in wind, temperature, and humidity, which significantly influence pollutant concentrations. Furthermore, standard machine learning models often require extensive datasets, making them difficult to apply to the brief windows of initial lockdown periods. This research addresses these gaps by providing a weather-compensated, data-efficient modeling approach.

Methods

The researchers employed a two-stage modeling strategy based on Generalized Additive Models (GAMs). First, "Pre-Lockdown" (pre-LD) models were trained on historical data (2018–2019) to capture the relationship between meteorological variables (e.g., wind speed, precipitation) and air quality. Second, a "Lockdown" (LD) model was developed using transfer learning. By fixing weather-related parameters—assuming atmospheric physics remain constant—and only retraining time-dependent variables (proxies for human mobility), the team could accurately model the lockdown period despite the "small data" constraint.

Methodological framework for transfer learning using GAMs.

The study utilized diverse datasets from China (Beijing, Wuhan), Switzerland, and Austria, incorporating hourly meteorological data and mobility reports from Apple and Google for validation.

Key Results and Insights

The models revealed that weather often masks the true effect of policy changes. For instance, in Switzerland, a warmer spring in 2020 would have led to overestimating lockdown benefits without weather compensation. Conversely, the model identified that a spike in PM10 in Europe during lockdown was caused by a Sahara dust event rather than local emissions.

Comparison of NO2 and PM2.5 levels in 2019 vs 2020 across pre-, during, and post-lockdown periods.

Key findings include:

  • NO2 Reductions: Significant drops were observed, ranging from -52.8% in Wuhan to -22.6% in Lower Austria.
  • Particulate Matter Paradox: While Wuhan saw a 50% drop in PM2.5, Beijing experienced only a 10% reduction. This is attributed to increased "oxidation capacity" in the air as NO2 levels fell, accelerating the formation of secondary particulates.
  • Transfer Learning Efficiency: The authors demonstrated that as little as four weeks of data is sufficient to adapt a pre-existing model to new social conditions.

Conclusion

This framework provides a robust tool for policymakers to evaluate the upper bounds of air quality improvements achievable through mobility restrictions. Future work will focus on modeling complex chemical interactions between pollutants and incorporating land-use data for broader spatial predictions.

Summary updated

Abstract

The COVID-19 related lockdown measures offer a unique opportunity to understand how changes in economic activity and traffic affect ambient air quality and how much pollution reduction potential can the society offer through digitalization and mobilitylimiting policies. In this work, we estimate pollution reduction over the lockdown period by using the measurements from ground air pollution monitoring stations, training a long-term prediction model and comparing its predictions to measured values over the lockdown month.We show that our models achieve state-of-the-art performance on the data from air pollution measurement stations in Switzerland and in China: evaluate up to -15.8% / +34.4% change in NO2 / PM10 in Zurich; -35.3 % / -3.5 % and -42.4 % / -34.7 % in NO2 / PM2.5 in Beijing and Wuhan respectively. Our reduction estimates are consistent with recent publications, yet in contrast to prior works, our method takes local weather into account. What can we learn from pollution emissions during lockdown? The lockdown period was too short to train meaningful models from scratch. To tackle this problem, we use transfer learning to newly fit only traffic-dependent variables. We show that the resulting models are accurate, suitable for an analysis of the post-lockdown period and capable of estimating the future air pollution reduction potential.
BibTeX
@misc{einsiedler2020interpretable,
  author = {Johanna Einsiedler and Yun Cheng and Franz Papst and Olga Saukh},
  title = {{Interpretable and Transferable Models to Understand the Impact of Lockdown Measures on Local Air Quality}},
  year = {2020},
  eprint = {2011.10144},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2011.10144},
}
Figure