Skip to main content

LUP Student Papers

LUND UNIVERSITY LIBRARIES

Modelling of Ground PM2.5 Concentrations in Stockholm, Sweden: A Machine-Learning Based Approach

Rönnbäck, Eric John LU (2026) NGEK11 20261
Department of Earth and Environmental Sciences (MGeo)
Abstract
Accurate estimation of fine particulate matter (PM2.5) is essential for epidemiological exposure assessment, yet monitoring networks can be too sparse to capture spatial variation across urban areas. Existing modeling approaches address this gap using either physics-based dispersion models, which require detailed emission inventories, or machine learning models that rely on satellite aerosol optical depth (AOD), which suffers from frequent missing retrievals at high latitudes. This study develops a Random Forest model for predicting daily PM2.5 concentrations in the greater Stockholm area (2014 - 2019) using publicly available meteorological reanalysis data and spatial land use predictors. Validated through leave-one-station-out (LOSO)... (More)
Accurate estimation of fine particulate matter (PM2.5) is essential for epidemiological exposure assessment, yet monitoring networks can be too sparse to capture spatial variation across urban areas. Existing modeling approaches address this gap using either physics-based dispersion models, which require detailed emission inventories, or machine learning models that rely on satellite aerosol optical depth (AOD), which suffers from frequent missing retrievals at high latitudes. This study develops a Random Forest model for predicting daily PM2.5 concentrations in the greater Stockholm area (2014 - 2019) using publicly available meteorological reanalysis data and spatial land use predictors. Validated through leave-one-station-out (LOSO) cross validation across eight monitoring stations, the model achieves an R² of 0.84 and an RMSE value of 1.57 μg/m. Decomposition into temporal and spatial components reveals that the model's skill is overwhelmingly temporal (temporal R² = 0.84), while its ability to differentiate between stations on a given day is poor (spatial R² = -0.76) consistent with high inter-station correlations (mean r = 0.89) indicating that Stockholm's stations largely measure the same regional air mass. Meteorological predictors, particularly planetary boundary layer height and wind components, account for over 97% of model importance, while spatial predictors contribute less than 3%. These results demonstrate that RF-based PM2.5 prediction is feasible in a low-concentration Nordic environment without satellite AOD, but that the resulting model captures regional temporal variation rather than local spatial differentiation (Less)
Please use this url to cite or link to this publication:
author
Rönnbäck, Eric John LU
supervisor
organization
course
NGEK11 20261
year
type
M2 - Bachelor Degree
subject
language
English
id
9248115
date added to LUP
2026-08-12 10:51:12
date last changed
2026-08-12 10:51:12
@misc{9248115,
  abstract     = {{Accurate estimation of fine particulate matter (PM2.5) is essential for epidemiological exposure assessment, yet monitoring networks can be too sparse to capture spatial variation across urban areas. Existing modeling approaches address this gap using either physics-based dispersion models, which require detailed emission inventories, or machine learning models that rely on satellite aerosol optical depth (AOD), which suffers from frequent missing retrievals at high latitudes. This study develops a Random Forest model for predicting daily PM2.5 concentrations in the greater Stockholm area (2014 - 2019) using publicly available meteorological reanalysis data and spatial land use predictors. Validated through leave-one-station-out (LOSO) cross validation across eight monitoring stations, the model achieves an R² of 0.84 and an RMSE value of 1.57 μg/m. Decomposition into temporal and spatial components reveals that the model's skill is overwhelmingly temporal (temporal R² = 0.84), while its ability to differentiate between stations on a given day is poor (spatial R² = -0.76) consistent with high inter-station correlations (mean r = 0.89) indicating that Stockholm's stations largely measure the same regional air mass. Meteorological predictors, particularly planetary boundary layer height and wind components, account for over 97% of model importance, while spatial predictors contribute less than 3%. These results demonstrate that RF-based PM2.5 prediction is feasible in a low-concentration Nordic environment without satellite AOD, but that the resulting model captures regional temporal variation rather than local spatial differentiation}},
  author       = {{Rönnbäck, Eric John}},
  language     = {{eng}},
  note         = {{Student Paper}},
  title        = {{Modelling of Ground PM2.5 Concentrations in Stockholm, Sweden: A Machine-Learning Based Approach}},
  year         = {{2026}},
}