Development of a Novel Imputation Framework for PM2.5 Particle Data in Pakistani Cities using Machine Learning and Statistical Techniques

Ahmad Alsaber^1*Muhammad Khan^2*JIAZHU PAN²Amani Alshatti³Alison J. Gray²

• ¹American University of Kuwait, Safat, Kuwait
• ²University of Strathclyde Department of Mathematics and Statistics, Glasgow, United Kingdom
• ³The Public Authority for Applied Education and Training, Safat, Kuwait

The absence of PM2.5 data in environmental monitoring systems, due to sensor and communication failures, system maintenance issues, and monitoring gaps, can hinder assessments relevant to public health and the development of policies aimed at improving air quality. Consequently, this study aims to analyse five imputation methods: (1) Bayesian regressions (BR), (2) K-Nearest Neighbors (KNN), (3) missForest, (4) Predictive Mean Matching (PMM), and (5) Random Forest (RF), considering data from daily measurement of PM2.5 collected from May 2019 to December 2024 across monitoring sites in the Pakistani cites of Islamabad, Karachi, Lahore and Peshawar. The data was used to simulate three missing mechanisms, namely MCAR, MAR and MNAR, with missing rates ranging from 5% to 25%. The performance of each imputation technique was evaluated using Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) criteria to assess the accuracy of each imputation technique. The results confirmed that imputation accuracy under the MAR mechanism consistently produced the lowest errors as the missing rates increased. Among all the tested methods, missForest and KNN were the most optimal and had the best performance at all missingness levels, and, more importantly, retained most of the temporal structure, range, and variability of the data. The relative performance of the missForest in RMSE and MAE values among the various missing rates and mechanisms was lowest, thus presenting it as the best imputation approach when analyzing PM2.5 data in this scenario.