Explainable PSO-optimised machine learning models for multi-pollutant air quality forecasting in major African cities with transfer learning

Poor air quality is responsible for deaths, particularly in African cities, where rapid urbanisation and rising vehicle emissions are increasing. The current study employs the hybrid Particle Swarm Optimisation–Machine Learning (PSO–ML) framework for air quality prediction in selected African cities, namely, Cairo, Johannesburg, Kinshasa, Lagos and Nairobi. The study contributes novel insights to environmental modelling by integrating advanced machine learning, metaheuristic optimisation, and transfer learning and domain adaptation techniques to address the issue of air quality prediction in African cities. Key variables such as particulate matter (PM 2.5 and PM 10 ), nitrogen dioxide (NO 2 ), sulphur dioxide (SO 2 ), carbon monoxide (CO) and ozone (O 3 ) were used in this study. The performance of XGBoost using PSO optimisation and PSO optimisation with Random Forest was compared with the performance of standard XGBoost and RF using several metrics such as the R 2 and Mean Absolute Error. From the obtained results, the PSO models outperformed the baseline algorithms. Comparing cross-city testing, the PSO-XGBoost models performed very well when trained with the Lagos dataset and tested with the Nairobi dataset, demonstrating the applicability of transfer learning. The algorithm garnered an R 2 value of 83% when trained with Lagos and tested with Nairobi under the PSO-XGBoost model, whilst the PSO-RF followed second with an R 2 value of 81% when trained with Lagos and tested with Johannesburg. Moreover, the PSO-XGBoost had 80% when Lagos and Cairo were tested with Johannesburg and Nairobi, respectively. In addition, the domain adaptation through fine-tuning revealed significant improvements in predictive skill for all source-target pairs, demonstrating that adapting models to different African cities is a useful step. The greatest gains were from Lagos to Nairobi (16.4%). Moreover, the SHAP results highlight the critical role of the SO 2 factor in predicting PM 2.5 . The enhanced machine learning offers a flexible and powerful tool that can be utilised to support data-driven environmental management and contribute to Sustainable Development Goals in health, climate action, and sustainable urban development. It also provides a pragmatic approach to estimating air quality in metropolitan regions with few monitoring facilities by training predictive models on data-abundant cities.