Deep learning approaches in air pollution prediction and forecasting—a systematic review

Rapid urbanization and industrialization have led to environmental pollution, resulting in adverse health concerns and ecosystem degradation. Among these, the most serious widespread threat is air pollution, which has a direct impact on human health, climate change, and ecological balance. Over the past few years, Deep Learning (DL) and Machine Learning (ML) models have offered promising solutions for predicting and monitoring air pollution. The objective of this study is to explore the most recent advances in air pollution prediction using the emerging ML and DL approaches, with a focus on single-pollutant and multi-pollutant in diverse locations. This comprehensive review is conducted through a systematic approach following the PRISMA guidelines. Out of 5,066 peer-reviewed studies published from 2020 to 2024 on air pollution prediction and forecasting, 50 articles with a focus on DL and ML approaches for air pollution prediction across various regions were chosen for the study. Various methodologies, pre-processing techniques, and statistical methods emphasize the use of historical data in predicting air pollution. The findings reveal that particulate matter (PM), particularly PM 2.5 , is the most frequently predicted pollutant. Models such as XGBoost and CatBoost demonstrated high performance on structured data, achieving impressive R 2 (up to 99%) and low RMSE (0.6–2.5) in Air Quality Index (AQI) prediction. Hybrid DL models have demonstrated strong predictive performance across various studies. However, significant challenges persist in model generalization, data quality, multi-pollutant prediction and dependence on historical data. This review examines current trends, the key complications, and outlines future directions for scalable and sustainable air quality prediction systems with a focus on emerging DL and ML approaches.