Estimation of PM10 variations in the Southeastern and Eastern Anatolia regions of Türkiye using remote sensing and statistical models

Abstract

The significance of air quality extends to both indoor and outdoor settings, with poor air quality directly impacting the quality of life. Particulate Matters smaller than 10 μm in diameter (PM10) are of particular concern. Penetration to the airways is easier with these types of particles. Particulate Matters (PMs) can contribute to both respiratory and cardiovascular diseases (i.e. asthma, emphysema, and lung cancer). Air quality in Türkiye is affected by dust coming from Sahara Desert and Arabian Peninsula. Our study aims to research PM10 variations in the Southeastern and Eastern Anatolia regions of Türkiye from 2014 to 2016 using remote sensing and statistical models. The Moderate Resolution Imaging Spectroradiometer (MODIS) derived Level-3 Aerosol Optical Depth (AOD) data, air quality data and meteorological variables of 15 ground-observed stations present in the Southeastern and Eastern Anatolia regions of Türkiye were used in this study to assess PM10 estimations in the study area. The PM10 ground observations used in the study for the years 2014, 2015 and 2016 were received from the Republic of Türkiye Ministry of Environment and Urbanization and Climate Change. Hourly PM10 and meteorological data were converted to daily data and days having more than 30% missing values were not included into the study in order to find out PM10 variations. Moreover, statistical methods such as Multiple Linear Regression (MLR) model, Linear Mixed Effect (LME) model and Machine Learning (ML) techniques were used to anticipate the Particulate Matter (PM) concentration in the region using satellite-based AOD and meteorological variables such as temperature, wind speed, relative humidity, precipitation and atmospheric pressure. Since some cities are located on the same grid in the AOD study area map obtained from MODIS in the PM10 estimation studies, 12 out of 15 stations were included in the statistical analysis to avoid extra PM10 calculations. The results indicated that aerosol pollution in the region as a result of transported dust from the Arabian Peninsula in spring time was high enough and the PM10 concentration in the cities close to the borders of Iraq and Suria such as Hakkari and Mardin was observed more than 200 µg/m3 in spring of 2015. Both local heating emissions and long-range dust transport could have significantly impact on PM10 levels in Southeastern Anatolia in study period. In addition, stubble burning in those years may also be effective in the high PM10 levels in the region in the fall. Significant amounts of dust, pollution, or biomass combustion make contribution to higher atmospheric aerosol concentrations. MODIS data from the AOD maps of 2014, 2015 and 2016 show that the Southeastern and Eastern Anatolia regions were exposed to aerosol pollution. The results of statistical models for prediction of PM10 indicated that PM10 was very dependent on AOD and temperature. The statistical parameters such as Correlation Coefficient (R), R-Squared (R2) and Root Mean Square Error (RMSE) were calculated to understand performances of models applied in the study. In the MLR method, PM10 was estimated by including only one meteorological factor other than AOD. Each meteorological factor was used with AOD in the equation respectively in the MLR method. When PM10 was predicted by using all meteorological factors respectively, it showed that among these parameters, temperature affected PM10 values more. Based on the outcomes, performance of all statistical models was improved when AOD values with all meteorological parameters were used in estimating PM10. For example, the R, R2 and RMSE values of MLR for averaged data for summer season were calculated about 0.69, 0.47 and 8.21 with best performance in the study period, while winter accounted to the lower performance with R, R2 and RMSE values of 0.40, 0.16 and 25.10, respectively. MLR for all 12-station data gave its best result in autumn season with an R, R2 and RMSE values of 0.57, 0.33 and 42.46 respectively. For the LME model, the random effect parameter selected AOD enabled the model to have R, R2 and RMSE values of 0.51, 0.26 and 36.60, respectively for all study period. Atmospheric pressure also had a random effect in the LME model, with R, R2 and RMSE values of 0.86, 0.73 and 28.26, respectively for all studied years. AOD and atmospheric pressure had a significant impact on improving the LME model results. For whole study period Extreme Gradient Boosting (XGBoost) having 0.73, 0.54 and 17.88 values as R, R2 and RMSE, respectively for averaged data and having 0.69, 0.44 and 32.13 values as R, R2 and RMSE, respectively for non-averaged data making a moderate performance in estimating PM10 levels. Among all statistical methods, Random Forest performed the best in terms of fitting the regression line, with R, R2 and RMSE values of 0.97, 0.93 and 17.90, respectively, when averaged variables are included in the model, and R, R2 and RMSE values of 0.97, 0.93 and 29.26, respectively, when all variables are included in the model. While MLR depended on factors like quantity of meteorological variables included, its performance also was affected by which meteorological variable is chosen with AOD in PM10 estimation. In this study, the LME model also showed variability in performance depending on which meteorological variable was selected as a random effect for PM10 estimation. XGBoost's performance was moderate and it also had better results than MLR method for overall. In conclusion, satellite data with meteorological variables gives us the best performance when it is introduced to the Random Forest Model in order to forecast PM10 in study area. The correlation between PM10 and AOD is influenced by weather conditions, local pollutant emissions and the chemical composition of aerosols. Ground-based monitoring data is commonly used in health effect research. Because satellite data is readily available and inexpensive, AOD images can be used to estimate PM10 via using Machine Learning methods which processes factors that affect abundance of PM10.