Synergizing TabNet and SHAP for PM10 Forecasting: Insights From Makkah, Saudi Arabia

Particulate matter (PM10) poses a serious threat to public health by increasing the risk of respiratory issues like asthma and bronchitis, as well as cardiovascular problems such as heart attacks and strokes. In Makkah, Saudi Arabia, the combined impact of dense vehicular traffic, large-scale construction projects, and an arid climate contributes to elevated PM10 concentrations, posing substantial challenges to air quality management and urban sustainability. This study utilizes the TabNet model to estimate PM10 concentrations, taking advantage of its ability to perform sparse feature selection and sequential decision-making to uncover complex relationships among different environmental variables. To improve the predictive accuracy of proposed model, hyperparameter tuning was carried out using the Covariance Matrix Adaptation Evolution Strategy (CMA-ES). The dataset containing meteorological and atmospheric parameters was collected from the Haram station in Makkah over the period from January 2016 to December 2018. TabNet outperformed other machine learning models, achieving a Mean Absolute Error (MAE) of 8.27 and coefficient of determination (R2) of 0.872 on the training set, while attaining an MAE of 9.05 and an R2 of 0.805 on the testing set. Afterwards, SHAP analysis illustrated the relative contributions of various features to PM10 concentrations, identifying atmospheric pressure as the most significant factor, closely followed by humidity and temperature. Lower to medium atmospheric pressure was found to substantially elevate PM10 levels, whereas medium to high humidity and elevated temperatures were likewise associated with increased PM10 concentrations. Furthermore, SHAP interaction plots revealed a moderating effect of atmospheric pressure on the influence of temperature on PM10 levels. These insights highlight the importance of considering both individual and interactive environmental factors when developing air quality models, leading to a deeper understanding of air pollution dynamics in Makkah and supporting more effective mitigation strategies.