A Benchmark of Deep Neural Network Models for PM2.5 Forecasting Using Multimodal Data: A Case Study in Africa and the Middle East

Abstract

ABSTRACT Air pollution forecasting is essential for protecting public health and shaping effective environmental policy, yet it remains especially challenging in regions such as Africa and the Middle East where monitoring networks are sparse, and data availability is limited. This study addresses these challenges by developing a unified deep learning framework that integrates ground-based air quality observa tions, meteorological reanalysis data, and satellite-derived aerosol information to forecast daily PM2.5 concentrations. Using diverse urban environments as testbeds, the models were rigorously benchmarked across multiple error and correlation metrics to evaluate their ability to capture pollu tant dynamics under varying climatic and data conditions. The results demonstrate that convolutional architectures, particularly a one-dimensional CNN, achieved the most reliable forecasts, consistently outperforming both simpler baselines and more complex recurrent and graph-based models. These findings highlight not only the value of combining heterogeneous data sources but also the effective ness of lightweight, efficient architectures for operational forecasting. By providing systematic benchmarks in data-scarce contexts, this study supports the development of scalable early-warning systems and offers actionable insights for evidence-based environmental management and policy.