Deep learning-based pm forecasting and post-infant mortality assessment in urban areas: a case study in Bangladesh

Abstract

This study investigated the application of advanced deep learning models for forecasting particulate matter (PM) pollution and assessing its associated health impacts in Bangladesh. Specifically, the Temporal Fusion Transformer (TFT), Deep Autoregressive Recurrent Neural Network (DeepAR), Generative Adversarial Network (GAN), and 1D Convolutional Neural Network (1D-CNN) were employed to forecast daily PM concentrations for the next seven days across four urban areas (Chattogram, Rajshahi, Dhaka, and Sylhet) during 2013–2018. The study further quantified post-infant mortality risks attributable to PM₁₀ exposure in these regions. Model performance was evaluated using Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Coefficient of Determination (R²), and Index of Agreement (IOA). TFT demonstrated superior capability, achieving mean RMSEs of 21.23 µg/m³ (PM_2.5) and 34.90 µg/m³ (PM₁₀), outperforming DeepAR by 39% and 40.6%, respectively, and surpassing 1D-CNN and GAN. The attention mechanism in TFT revealed distinct temporal dynamics, with PM_2.5 prediction dominated by temperature and relative humidity indicating inversion-driven accumulation, while PM₁₀ forecasting emphasized wind speed, rainfall, and relative humidity aligning with Bangladesh’s monsoon-driven dispersion and dust resuspension. Health risk analysis, conducted using the World Health Organization’s (WHO) Air Quality Health Impact Assessment Tool (AirQ+), revealed significant post-infant mortality impacts, with the highest attributable mortality per 1000 population observed in Dhaka (42.95%) and Sylhet (24.64%). The strong agreement between observed and forecasted health outcomes validated the reliability of the TFT model.