Data fusion for air pollution monitoring: satellite, ground based and ensemble models for PM _2.5 prediction

Air pollution emerged as a serious concern globally, causing environmental degradation and human health impacts, resulting in around 9 million mortalities each year. This research study employed a combination of ground-based low-cost sensors (LCS), satellite data, and meteorological data to predict PM2.5 concentrations. This study utilizes Moderate Resolution Imaging Spectroradiometer (MODIS) land surface reflectance data, which fills a gap in existing literature. This study mainly focuses on the Islamabad Capital Territory (ICT) area, where rapid urbanization and high air pollution cause health and environmental implications. By combining LCS and MODIS datasets, ensemble learning models, specifically Random Forest (RF), Extra Trees (ET), and Gradient Boosting (GB), were implemented to enhance the accuracy of PM2.5 predictions. This work has never been done before in the region and particularly utilizing surface reflectance that makes it novel to predict the particulate matter, which is best for data scarce region. The dataset was enhanced with polynomial features and split by 80/20 (80% for training and 20% for testing). A 5-fold Cross-validation was used to assess model robustness, after which the models were tested on 20% data. The results show that GB (RMSE = 14.87 (4.3%), MAE = 9.10(2.6%), and R2 = 0.73) has the best performance, followed by ET and RF. The findings underscore the utility of data fusion in improving air quality monitoring and support evidence-based policymaking in resource-limited settings, especially in areas with no air quality monitoring infrastructure.