Quantifying uncertainty in the spatial prediction of PM _2.5 using a deep learning algorithm

Air pollution, particularly particulate matter 2.5 (PM2.5), poses significant health risks in urban areas like Tehran, Iran. Despite efforts to predict PM2.5-prone areas, uncertainty remains challenging, hindering effective mitigation strategies. This study addresses this gap by investigating uncertainty in PM2.5 prediction in Tehran. To achieve this goal, a database containing 11 factors influencing PM2.5 occurrence was established for modelling purposes. This study employed a deep learning approach called convolutional neural network (CNN) to model and predict regions prone to PM2.5 risks. Shannon entropy and Dempster–Shafer methods were used to assess uncertainty in predicting PM2.5-prone areas, while empirical Bayesian kriging (EBK) interpolation and Gaussian noise were applied to investigate uncertainty in input factors. Additionally, Bayesian neural network (BNN) algorithm was introduced to assess uncertainty in model parameters. Evaluation results of PM2.5-prone area maps using receiver operating characteristic (ROC) curves demonstrated 97.7% accuracy for the CNN algorithm in predicting susceptible areas, 95% accuracy for the BNN-CNN algorithm in parameter-calibrated model prediction and 82.7% accuracy for the CNN algorithm in noisy data prediction scenarios. The uncertainty analysis revealed that the central areas of Tehran exhibited higher reliability in predicting PM2.5-prone areas, while western and eastern regions, especially at the borders, showed higher uncertainty.