Using deep generative models to estimate PM2.5 concentration from satellite AOD data collected from Tehran, Iran

The concentration of particulate matters has recently increased in the metropolitan area of Tehran resulting in many severe hazards for both the environment and citizens. Particulate matters (PM) with a diameter less than 2.5 microns (PM2.5) are considered to be one of the most dangerous types of pollution. Estimating the concentration of these particles in Tehran is challenging due to the existence of various sources of pollution and the lack of sufficient ground stations. Aerosol optical depth (AOD) data retrieved from satellite imagery can be an alternative. However, AOD are not easily convertible into surface pollution and requires the development of appropriate models such as those based on data-driven approaches and machine learning techniques. Thus, the present study seeks to create a model to estimate the concentration of PM2.5 in Tehran employing deep generative models and in-situ measurements, meteorological data, and AOD data extracted from MODIS satellite imagery. Reviewed literature has proved the ability of deep learning techniques to solve regression and classification problems. Deep learning techniques are divided into various categories, one of which is based on the generative models seeking to reconstruct the input features. In this way, high-level and efficient features can be employed to explore the relationship between PM2.5 and AOD. Thus, the present study has investigated the potential of deep generative models for estimating PM2.5 concentration from high resolution AOD data retrieved from satellite imagery. 

Materials and Study Area:

 As a metropolitan area suffering from air pollution particularly in winters, the capital city of Iran, Tehran was selected as the study area. PM2.5, the main source of pollution in Tehran, is mainly emitted from vehicles and especially old urban public transportfleet.Aerosol data collected by Aqua and Terra sensors of MODIS and retrieved by Multi-angle Implementation of Atmospheric Correction (MAIAC) algorithm were used in the present study. Meteorological data were obtained from the global ECMWF climate model, and the concentration of PM2.5 was measured at air quality monitoring stations. Data were collected for a time interval of January 2013 to January 2020.

The present study has investigated the potential of deep generative models used to provide an estimate of PM2.5 concentration based on satellite AOD data. To reach such an aim, three types of deep generative neural networks, deep autoencoder (DAE), deep belief network (DBN) and conditional generative adversarial network (CGAN) were developed. Moreover, the performance of deep generative modes was compared with linear regression techniques as typical models used to explore the relation between PM2.5 and AOD data. Finally, the most accurate model for the generation of high resolution (1km) PM2.5 maps from AOD data was selected based on the performance of models.

 Results and Discussion:   

The accuracy of each developed model was evaluated using the test data and the obtained results were compared with results obtained from other basic linear regression models. Accuracy evaluation indicated that the developed deep autoencoder (DAE) combined with support vector regression led to the highest correlation (R2 = 0.69) and lowest RMSE (10.34) and MAE (7.95) and thus, can be potentially used for high resolution estimation of PM2.5 concentration. Next was the developed deep belief network which with a performance close to DAE demonstrated its potential capability to estimate PM2.5 concentration from satellite AOD data. The CGAN network acted less accurately in the estimation of PM2.5 concentration as compared to other deep generative models, but outperformed the linear regression algorithms on the test data. To sum up, findings indicated that deep generative models have outperformed classical linear regression techniques used for high resolution estimation of PM2.5 from satellite AOD data. Among the linear methods, the highest accuracy was achieved by the Lasso algorithm with an RSME of 12.14 and MAE of 9.46 on the test data which showed the significance of regularization for the improvement of performance in linear regression algorithms. Nevertheless, the accuracy of linear regression techniques was much lower than deep generative models.

Finally, DAE was selected as the best model for the estimation of PM2.5 concentration across the study area and high resolution maps of PM2.5 concentration were generated using the developed model. Investigating the daily PM2.5 maps generated for two days with different air quality conditions (clean and polluted) demonstrated the efficiency of the developed DAE for PM2.5 modeling.