Estimating long-term daily PM2.5 concentrations using deep learning over Delhi, India, and analyzing temporal variations in it

In recent years, the rapid increase in the number of PM2.5 pollution episodes has become a common issue, as it causes severe problems in health, the atmosphere-climate system, and air quality. Therefore, monitoring and analyzing trends in PM2.5 has gained significant attention in recent years. However, the lack of continuous daily PM2.5 data makes it challenging to analyze the temporal variations in PM2.5. In this study, the daily PM2.5 concentration is estimated from 2000 to 2020 for the study area using the state-of-the-art Long Short-Term Memory (LSTM) artificial neural network model. LSTM estimates PM2.5 based on satellite-derived aerosol optical depth (AOD), atmospheric temperature, wind speed in horizontal direction, wind speed in vertical direction, relative humidity, atmospheric pressure, and rainfall. Further, the temporal variations in the estimated long-term PM2.5 concentrations are analyzed using the Mann- Kendall trend test. From the comparison of month-wise average of daily PM2.5 concentrations during the first 5 years (2000-2004) and the last 5 years (2015-2019), it is observed that the average PM2.5 concentrations during 2015-2019 are significantly higher when compared to the average PM2.5 concentrations during 2000-2004. The highest percentage change is in January, with values of 130.07%, 203.98%, 127.46%, and 130.04% for stations S1, S2, S3, and S4 respectively. Further, based on the analysis of the trend in the corresponding AQI index, it is observed that during October, November, December, and January, there is an increasing trend in the AQI categories poor, very poor, and severe. On the other hand, the AQI categories good, satisfactory, and moderate are decreasing during March, April, and May.