Spatio-temporal modelling of particulate matter concentrations using satellite derived aerosol optical depth over coastal region of Chennai in India

The present study primarily focuses on describing aerosol optical depth (AOD), its distribution pattern and seasonal variation, and modelling Particulate Matter Concentrations in Chennai. The frequency distribution of AOD and PM2.5 demonstrates that AOD can be used as a proxy for estimating PM2.5 in the study region as the occurrence of AOD almost resonates with that of PM2.5. The seasonal variation of AOD and PM2.5 revealed that during the winter (October-January) and summer (February-May) seasons, AOD reasonably followed the trend of PM2.5. However, during the monsoon period, AOD showed random variations. Different models like linear and non-linear regression models and machine learning models such as random forest (RF) have been developed for PM2.5 estimation. The model's performance in different stations and seasons has been assessed. The effect of meteorology and other factors in the model has also been assessed. From linear and non-linear model analysis, AOD was a significant parameter in estimating PM2.5. The Random Forest model was the stable model for the study region, with a model R-2 of 0.53 and an RMSE of 15.89 mu g/m(3). The inclusion of meteorological parameters like relative humidity, wind speed, and wind direction decreased the error in prediction by 17.45 mu g/m(3). The seasonal and spatial analysis indicates that the prediction capability of models varies with stations and seasons. The best performing model was found to be Model RF, and the model could explain about 53.14% of the variability in PM2.5 concentration occurrence in the study region with a prediction error of 15.89 mu g/m(3).

Automated summary

The present study primarily focuses on describing aerosol optical depth (AOD), its distribution pattern and seasonal variation, and modelling Particulate Matter (PM) concentrations in Chennai. The study finds that AOD can be used as a proxy for estimating PM2.5 in the region, and different models have been developed for PM2.5 estimation. The Random Forest model performs the best, explaining about 53.14% of the variability in PM2.5 concentration, with a prediction error of 15.89 μg/m^3.