عنوان مقاله [English]
Aerosols or airborne particulate matter (PM) with different sizes from both natural and anthropogenic emission sources have substantial influences on climate, environment and human health. Particulate matter is characterized by the size (PM10 andPM2.5; particles having an aerodynamic diameter less than 10 and 2.5 μm). However, due to limited spatial coverage and high operational cost, in situ observations are insufficient to capture high resolution, tempo-spatial variation of PM concentration, especially for many developing countries such as Iran. Satellite remote sensing technology, on the other hand, provides a cost-effective way for epidemiological studies and PM monitoring. at various scales by measuring satellite-derived aerosol optical depth (AOD), especially for places where ground-level monitoring is not available. Satellite-derived AOD is related to ground-level PM concentration and can be empirically converted into PM mass. Therefore, a number of empirical models have been developed to predict ground-level PM2.5 or PM10 concentration from various satellite-derived AOD products, such as Moderate Resolution Imaging Spectroradiometer (MODIS) and the Multiangle Imaging Spectro Radiometer (MISR). There are a few studies about the AOD-PM relationship in Iran and much fewer studies use MISR instrument. level2 product of AOD retrieved by MISR with improved resolution of 4.4 km, is a suitable instrument for estimating particular matter. Therefore, the main aim of this study was investigating empirical linear and nonlinear regression models along with artificial neural network model for estimating particulate matter (containing PM2.5 and PM10) using AOD from the MISR 4.4 km aerosol product. In this regard, Tehran one of the largest industrial cities of Iran during 2016 and 2017 was selected as the study area.
To assess ground level PM2.5 and PM10, a number of data sets from various sources were collected for this research, including 2 years daily PM2.5 and PM10 mass concentration at 13 air quality aground stations in Tehran which had sufficient records. Also we used 3 meteorological stations including Geophysic, Mehrabad, and Shemiran. Ground-based meteorological parameters (a total of five parameters), including surface wind speed (SPD), surface temperature (ST), visibility (Vis), and surface relative humidity (SRH), and dew point temperature (Td) were obtained from the meteorological stations.
MISR instrument onboard the EOS-Terra satellite measures the reflected and scattered sunlight from the Earth in four spectral bands of 446, 558, 672, and 866 nm at each of the nine viewing angles (nadir, ±26.1, ±45.6, ±60.0, and ±70.5°). It has a much narrower swath of about 380 km compared with that of about 2330 km of MODIS. The level 2 MISR AOD has the spatial resolution of 4.4 km and the temporal coverage is about once per week in mid latitude. The latest global MISR AOD product (version 22) for 2013 can be downloaded from the Atmospheric Sciences Data Center at NASA/Langley Research Center (http:// eosweb.larc.nasa.gov).
Since the data stations are dispersed, PM monitoring sites were matched with meteorological stations based on the neighboring. On the other hand, as MISR pixels are not distributed equally all over the region, we match them to closest PM monitoring site. Previous studies show that the distance between AOD pixels and closest PM monitoring ground stations must be within 25 km.
In order to retrieve the ground-level PM value proper conversion should be made first. Related researches have shown that the meteorological conditions (such as relative humidity) can strongly impact models for the AOD–PM relationship, as particle extinction properties can change substantially with different vertical mixing and aerosol hygroscopic growth. Thus, |there is a number of methods which proposed to use meteorological factors to improve the relationship between AOD and ground-level PM. In this study, we developed linear and nonlinear models using MISR AOD values coupled with meteorological parameters for estimating daily ground-level PM concentration:
where PM is daily ground-level PM2.5 or PM10 mass concentration (μg/m3), and AOD is MISR-derived AOD (unit less). c1 is the intercept Regression coefficients. c2–c7 are associated with predictor variables, including AOD, visibility (km), surface wind speed (m/s), surface temperature (°C), surface relative humidity (%), and dew point(°C).
In some studies, an exponential function is used for visibility, relative humidity and dew point temperature. Furthermore, surface meteorological conditions including surface temperature and surface wind speed are employed in the model to further amend the corrected AOD to obtain a better relationship between AOD and PM. Here is the none linear model:
As the third approach, neural network algorithms - black-box models of artificial intelligence- were employed. We used the artificial neural network (ANN) algorithm to model ground-level PM concentration based on the meteorological variables and satellite-derived AOD data. We used a back-propagation neural network (BPNN) algorithm to build the ground-level PM estimation model for predicting the PM concentration. The estimation of the ground-level PM concentration model consists of six in the input layer, six neurons in the hidden layer, and one neuron in the output layer. The seven parameters in the input layer include, surface temperature, wind speed, relative humidity, visibility, dew point temperature and MISR AOD products. The neuron in the output layer is PM concentration.
Cross validation (CV) of all models was conducted by leaving 10 percent of entire PM monitoring site out, fitting the model without this part, and predicting daily PM from 2016 to 2017 for the left out data with the fitted model. This was iterated so that every data was left
out one at a time, and the associations between the predicted values and the observed values were assessed to examine model performance.
Linear and nonlinear regression models had similar predictability for ground-level PM10 (respectively correlation coefficients of 41% and 45%) and PM2.5 (respectively correlation coefficients of 46% and 42%). Artificial neural networks improved the estimation of surface PM2.5 and PM10 significantly and the correlation coefficients improved to 60% and 65%. for PM2.5 and PM10, respectively. It is clear that the neural network model enhanced the regression models result about 19 to 23 percent. The result can use in the regions which there is not sufficient air quality stations.