TITLE:
Using Machine Learning Models to Predict Daily PM10 Concentration in the Wet Savanna of Lamto Station in Côte d’Ivoire
AUTHORS:
Touré E. N’Datchoh, Money Ossohou, Adama Bamba, Yoman C. E. Etien, Kouakou Kouadio, Fidèle Yoroba, Madina Doumbia, Mohamed L. Kassamba-Diaby, Sylvain Gnamien, Véronique Yoboué
KEYWORDS:
Lamto, Prediction, PM10, Machine Learning, Air Pollution, Performance
JOURNAL NAME:
Open Journal of Air Pollution,
Vol.14 No.4,
November
5,
2025
ABSTRACT: Despite its significant impacts on climate, environment, and public health, air pollution monitoring in Africa is still sparse. This study developed a machine learning framework of four models—namely SARIMAX, Random Forest, XGBoost, and LightGBM—that were trained on a daily basis using 25 meteorological variables from ERA5 reanalysis to reproduce daily PM10 surface concentrations in Lamto, Côte d’Ivoire. A total of 2225 daily recorded concentrations from 2017 to 2023 were used, with 80% (1780 days) allocated to training the models and 20% (445 days) for prediction. An assessment of ERA5 reanalysis data against in-situ measurements highlighted its ability to reproduce the pattern of meteorological variables such as temperature and relative humidity, despite some biases. Results show that all models were able to reproduce the observed PM10 variations, although they slightly overestimated concentrations during the main wet season (March-April-May-June-July) and underestimated high pollution events in the main dry season (November-December-January-February). Finally, performance analysis revealed that the Random Forest model outperformed, with an R2 of 0.78 and RMSE of 23 µg/m3, outperforming the other models. This framework successfully demonstrates the utility of machine learning for air quality prediction in West Africa, with potential for future improvements through bias correction and model combination.