چكيده لاتين
Due to rapid urbanization and industrialization, air pollution has become one of the fundamental challenges in countries and large industrial cities. Air pollutants are considered major contributors to human diseases. Among them, particulate matter with an aerodynamic diameter less than 2.5 micrometers (2.5PM) is recognized as the primary pollutant in the country, which infiltrates the respiratory system through inhalation and leads to respiratory and cardiovascular diseases, reproductive disorders, and central nervous system and cancer. Accurate spatiotemporal modeling of air pollutant concentrations, especially for air quality management and exposure assessment in epidemiological studies, is crucial. On one hand, the rapid advancement of computational technologies and access to relevant air quality data has enabled researchers to propose complex models using deep learning for modeling various air pollutants. Deep learning models have proven their efficiency in modeling complex and dynamic problems such as air pollution in various studies . On the other hand, measuring the concentration of 2.5PM locally is considered the gold standard, but this process is time-consuming and expensive. Moreover, air quality monitoring stations can be used to analyze air pollution in their vicinity, but this does not cover the entire area. Therefore, retrieved Aerosol Optical Depth (AOD) products from satellites as well as satellite imagery of other pollutants affecting 2.5PM concentrations have the potential to complement ground-based monitoring stations by providing spatially and temporally resolved exposure estimates. In the first step of this research, data preparation and completion were addressed. In the second step, spatial-temporal prediction methods for the concentration of 2.5PM at different time intervals were implemented and evaluated based on deep learning networks such as Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), and a combination of these two networks (ConvLSTM), as well as Multi-Layer Perceptron (MLP). Subsequently, the results obtained from the models were compared for the same prediction time lag and month. The structure of the employed models was designed considering the input data, and an optimization search was conducted through trial and error. Finally, the impact of the data used in this research was examined using gradient analysis. The data used in this study includes monthly concentrations of 2.5PM, meteorological parameters, remote sensing data, as well as parameters such as roads, power plant locations, and population density in the year 2021. The results of the experiments indicate that the ConvLSTM model outperformed other examined methods. This model is also able to estimate 91% (R2=0.91) of the variations in 2.5PM concentration for one-month ahead prediction and predict the pollution level. The summary of the results of this research indicates that the simultaneous use of various data types and the ConvLSTM deep learning network can be effective in modeling and predicting pollution by training spatial and temporal relationships present in the time series data.
