A Data-Driven Supervised Machine Learning Approach to Estimating Global Ambient Air Pollution Concentrations - With Associated Prediction Intervals

Aug 6, 2026

Upload by Finn

A data-driven supervised machine-learning model is presented to estimate global ambient air pollution concentrations at an hourly resolution, addressing a gap where many air quality models typically provide annual resolution. The work targets five key pollutants—NO2, O3, PM10, PM2.5, and SO2—used in the UK Daily Air Quality Index framework. It produces a comprehensive 2022 global hourly map, evaluates cross-country extrapolation feasibility, recommends monitoring-station placement using uncertainty metrics, and identifies the most severely polluted regions and dominant contributors.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages29

Reading Time1 h 13 min

AccessFree to download

Introduction

Air pollution monitoring challenges and need for hourly estimates

Model contributions and outputs

Related Work

Global transboundary drivers of air pollution

Climate change impacts and relevance of high-resolution views

A Data-Driven Supervised Machine Learning Approach to Estimating Global Ambient Air Pollution Concentrations With Associated Prediction Intervals Liam J. Berrisford 1,2,3, ∗ [l.berrisford@exeter.ac.uk](l.berrisford@exeter.ac.uk) [ cs .LG] 15 Feb 2024 3 Hugo Barbosa 1 [h.barbosa@exeter.ac.uk](h.barbosa@exeter.ac.uk) Ronaldo Menezes 1,4 [r.menezes@exeter.ac.uk](r.menezes@exeter.ac.uk) 1 BioComplex Laboratory, Department of Computer Science, University of Exeter, England 2 Department of Mathematics, University of Exeter, England UKRI Centre for Doctoral Training in Environmental Intelligence, University of Exeter, England 4 Department of Computer Science, Federal University of Ceará, Fortaleza, Brazil ∗ Corresponding Author February 19, 2024 1 Introduction Air pollution represents a significant global challenge. Astonishingly, 99% of the global population is exposed to air pollution that exceeds the air quality limits established by the World Health Organization (WHO) [1] . In response, numerous countries have implemented air pollution monitoring systems 1. While direct monitoring through stations isan essential initial step in comprehending air pollution, it is impractical to deploy a monitoring station at every location. This impracticality stems from both logistical challenges, such as the necessity for infrastructure like power lines and data connections, and financial considerations, with a standard monitoring station in the UK costing up to £198,000 [2] . Consequently, models are crucial for bridging spatial and temporal gaps in air pollution concentration measurements. Air pollution models, like those used by the WHO, typically provide data with an annual temporal resolution. However, there is an evident gap in global models that focus on hourly air pollution concentrations. Considering that certain air pollution guidelines, including WHO’s, specifically demand hourly resolution [3], the need for comprehensive, hourly resolved air pollution concentration estimates is paramount for effective decision-making. This paper introduces a datadriven, supervised machine-learning model designed to predict air pollution concentrations at an hourly resolution on a global scale. There are four key outputs from this work: 1. Development of a comprehensive air pollution concentration map, covering the entire globe for the year 2022, at an hourly resolution. This map includes concentrations of NO2 , O3 , PM 10 , PM2.5 , and SO2 , the quintet of pollutants that constitute the Daily Air Quality Index (DAQI) in the UK [4] . 2. Analysis to gauge the feasibility of extrapolating air pollution concentrations from one region to another. This involves addressing the critical question: To what extent can the air pollution data of one country accurately predict the air pollution levels in another? 3. Provision of strategic recommendations for the placement of future air pollution monitoring stations, informed by the uncertainty metrics derived from our model. 4. A comprehensive evaluation of global air quality, identifying regions with the most severe air pollution and pinpointing which of the five DAQI pollutants predominantly contributes to poor air quality in these areas. The complete spatial map of air pollution concentrations can be used downstream in other scientific studies, such as the health implication of particulate matter (PM) [5], or the impact of ozone on vegetation [6] at a resolution not previously possible. The unprecedented resolution of this map enables more detailed and accurate analyses than were previously feasible. Furthermore, our approach to estimating air pollution concentrations from one country based on data from another represents a step towards open data initiatives. This fosters international data sharing, contributing to a more equitable global understanding of air pollution and research [7] . Our findings illustrate that even data-rich countries like the United Kingdom can

FAQs

Why is an hourly resolved global air pollution dataset important?

Several air quality guidelines, including WHO’s, require hourly resolution, while commonly used global models often provide only annual resolution.

Which pollutants are included in the paper’s global concentration mapping?

The map covers NO2, O3, PM10, PM2.5, and SO2, corresponding to the UK Daily Air Quality Index pollutant set.

How does the study support decisions about where to place future monitoring stations?

It uses uncertainty metrics derived from the model to inform strategic recommendations for monitoring-station placement.

Aug 6, 2026

Upload by Finn