Comparative evaluation of artificial neural network (ANN) applications in analysing factors influencing air pollution in metropolitan areas: A case study of Iran

Effective management of air pollution in Isfahan metropolis requires robust, operational forecasting tools. Aiming to develop an efficient and cost-effective model for the simultaneous prediction of gaseous (CO, NO₂, SO₂) and non-gaseous (UVAI) pollutants, this study leveraged the capabilities of Artificial Neural Networks (ANN). The primary innovation of this research lies in its strategic and hybrid approach to input variable selection. In addition to traditional environmental variables (such as NDVI and distance from point sources), the model incorporates critical anthropogenic and urban variables, specifically population density and building density. This integration of data aims to account for the dynamic effects of local wind patterns induced by the city's physical fabric within the prediction model. The dataset comprised 2400 data points covering spatial and temporal variables. After normalization, the data were split into training (70 %) and test (30 %) sets. Model optimization parameters included 20,000 epochs, 10 hidden-layer neurons, and a weight-decay penalty (0.01) to mitigate overfitting. The model's quantitative performance was excellent. The coefficient of determination (R 2 ) for UVAI prediction was 0.999, indicating exceptional accuracy. Additionally, R 2 values were calculated at 0.98 for NO₂, 0.94 for SO₂, and approximately 0.90 for CO. These results, coupled with near-zero values for Mean Squared Error (MSE) and Root Mean Squared Error (RMSE), confirm the model's high capability for the early and reliable prediction of pollutants. Variable Importance in Projection (VIP) analysis revealed that population density and building density were the most significant predictors of urban pollution, highlighting the critical importance of controlling urban development to improve air quality. However, limitations regarding temporal data and the inherent non-dynamic nature of ANNs in fully modeling atmospheric phenomena remain fundamental constraints. Consequently, future research proposals focus on expanding temporal datasets (specifically including multi-year data for long-term validation) and transitioning towards hybrid models, such as CNN-LSTM. These approaches aim to process time-series dynamics better and maximize model accuracy across various atmospheric scenarios. Ultimately, the proposed ANN model provides a robust decision-support tool for short-term air quality management in Isfahan. • Objective and Methodology: The study aimed to analyze and forecast air pollution parameters in Isfahan Metropolis using an Artificial Neural Network (ANN) model . • Input Data: Input variables included DEM, NDVI, road/building density, population density, distance to parks, and industrial pollutant sources . The output pollutants were four key parameters: CO, NO₂, SO₂, and UVAI . • Model Training: The data (2400 samples) were split into 70% training and 30% testing sets. The ANN model was trained with 2000 iterations, 10 hidden units, and a weight decay of 0.010.010.01. • Model Performance: The model demonstrated very high predictive performance: R2R^2R2 values reached 0.999 for UVAI , 0.98 for NO₂ , 0.94 for SO₂ , and 0.90 for CO . • Input Variable Importance and Policy Recommendations: Variables related to human and urban activities (population and building density) were found to have the most significant impact. Recommended solutions include developing green spaces, improving public transportation, reducing reliance on private cars, relocating polluting industries, and enhancing fuel standards.