PM2.5 performance analysis under varying imputation strategies for incomplete sensor data

Air pollution remains a primary global health concern, with one of its most hazardous components, fine particulate matter PM 2.5 , being recognized as a significant contributor to premature mortality worldwide. Thus, forecasting PM 2.5 concentrations is essential for policymaking and mitigating their harmful effects. Low-cost sensor (LCS) technology has advanced air quality monitoring by providing real-time PM 2.5 measurements at finer spatial and temporal granularity due to their higher deployment density, enabling more comprehensive data for predictive modelling. However, these sensors can experience significant data gaps, necessitating effective imputation methods to improve the accuracy of PM 2.5 predictions. In this study, we focus on reconstructing missing PM 2.5 readings by evaluating the impact of different imputation strategies on PM 2.5 forecasting accuracy using data from eight LCSs deployed across Edmonton, Alberta, from January 2023 to December 2024. The dataset includes 7 features: temperature, relative humidity, wind speed, wind direction, latitude, longitude, and altitude, integrated from both LCS sensors and adjacent air quality monitoring stations. We evaluate the efficacy of Kriging as a spatial imputation method for PM 2.5 forecasting, comparing its performance along with mean imputation and k-nearest neighbours, against a no-imputation benchmark to assess whether imputation techniques provide meaningful improvements in forecasting accuracy. Additionally, we compare its performance across data with varying levels of missingness. Our comparative analysis examines how these imputation techniques affect various prediction models, including Random Forest, Extreme Gradient Boosting, Convolutional Neural Networks, Long Short-Term Memory (LSTM), and LSTM with an Attention mechanism. Our findings indicate that although differences among imputation methods were modest, Kriging consistently yielded the best results, particularly at 90% data availability for the deep learning models. The LSTM-AM architecture produced the best overall results in a short-horizon setting, \(R^2\) = 0.9602, MAE = 1.7815 \(\mu\) g/m \(^3\) , and RMSE = 4.5208 \(\mu\) g/m \(^3\) , suggesting an optimal combination of Kriging imputation with LSTM-AM modeling for near-term PM 2.5 forecasting applications.