پديد آورندگان :
حجتي، مجيد دانشگاه تهران - گروه سنجش از دور و سيستم اطلاعات جغرافيايي , درويشي بلوراني، علي دانشگاه تهران - گروه سنجش از دور و سيستم اطلاعات جغرافيايي , علوي پناه، كاظم دانشگاه تهران - گروه سنجش از دور و سيستم اطلاعات جغرافيايي , كياورز، مجيد دانشگاه تهران - گروه سنجش از دور و سيستم اطلاعات جغرافيايي , بداق جمالي، جواد سازمان هواشناسي - گروه هواشناسي
كليدواژه :
ذرات معلق , شبكه عصبي مصنوعي , موديس , MODIS , PM10 , ANN , MLP , AOD
چكيده لاتين :
Abstract Particular matters which are referred as PM contents in air are one of the main indexes in air pollution studies. PMs are consisted of water, solid particles with aerodynamic radius less than 10 micrometer or less than 2.5 micrometer. The main aim of this study is to find possibility of the usage of MODIS images for estimation of PM10 concentration of dust storms. To achieve this goal Aerosol Optical Depth (AOD) of MODIS sensors is used. Until now there has been lots of studies which used empirical relations in order to model the behavior of PM10 in dust storms. Some of these studies have combined methodology parameters such as temperature, wind direction and wind speed and etc. AOD data of MODIS is in five separate bands. Most of current studies have not focused on which of these bands are more useful than others and they simply used 555 nanometer band to estimate PM10 concentrations. In this study at first step the days with dust storms in 2013 in Khuzestan providence in Iran has been chosen. To choose these days the visibility of these days is used. Then MODIS AOD data are downloaded for these days. In the next step the correlation between PM10 data and MODIS AOD data is calculated and best band with higher correlation is selected. The best band in this step was 476 nanometers. To increase the accuracy of the model wind speed, wind direction, temperature, humidity and pressure for selected days are used in model. Two MLP (Multilayer Perceptron) and (Radial-Basis Function) RBF models are used in this study and their results has been compared with each other. Results showed that MLP has more accurate values than RBF and RMSE value of this model is 78 microgram/m3 (PPM). Other results of this study is that the 476 nanometer band of MODIS shows better results than other bands at estimation of PM10. 1- Introduction Concentrations of suspended particles are one of the main indicators of air quality. suspended particles are formed of a combination of liquid and solid particles with aerodynamic diameter less than 10 micrometers, or 2.5 micrometers. A major source of these particles in natural areas are places such as sea salt, dust, volcanic ash or resources caused by human activities such as fossil fuels, industry and transportation. In another division for mechanical particles they are sorted into categories such as dust or a chemical compound (such as SO2 and NOx). AOD is one of the parameters that helps us to estimate air quality using satellite imagery. AOD can be defined as reduction of the amount of sunlight absorbed by the aerosol particles Passive imaging is capable of providing measurement data for the AOD. AOD products can be obtained from satellite sensors such as TOMS, SEAWIFS, OMI, POLDER, MERIS, MODIS. Algorithms such as deep blue, dark object are developed. 1- Materials and methods AOD which is derived from MODIS satellite is produced with a spatial resolution of 10 × 10 km at Nadir and 20 × 40 km in the corners of the image. This product is produced using the dark target algorithm and is called MODIS MOD04. MODIS products for aerosols is freely available. Climatic factors affecting the PM are characterized into two general categories. Factors that directly influence on particles (such as speed and direction of air flow, Earth's surface temperature and precipitation) and the factors that influence indirectly on PM (such as temperature, humidity weather, clouds form and barometric pressure). Data which are used in this model includes the relative humidity (RH), wind speed (SWD), wind direction (SWD), the average temperature (T), and the Earth's surface pressure (SP) which have been collected from local meteorological stations. Study Area Ahvaz, the center of Khuzestan province, is one of the major cities in Iran. Its geographical location is in 31 degrees and 20 minutes’ north latitude and 48 degrees 40 minutes’ east longitude, in the plain of Khuzestan. This city is 18 meters above sea level. There are large industrial plants, offices and industrial facilities, the South Oil Company and National Iranian Drilling Company in Ahvaz, because of these facilities Ahvaz has turned into one of the most important industrial centers in Iran and this has caused many immigrants to the Ahwaz. It is one of the cities where most dust storms happen. The origin of most of these dust storms are Iraq and Syria, and some recent storms are with a domestic source. 2- Discussion of results The model of the PM10 content in dust storms are primarily based on data from meteorological stations and MODIS MOD04 products. Wind speed and direction, relative humidity, and the PM is collected by ground stations in accordance with the time specified and selected satellite overpass. At the first step days with dust storms are specified. These days are specified using the reports about dust storms and also visibility of the ground stations in study area. Because MODIS AOD values are calculated in different bands, in the first stage between AOD values which are extracted from MODIS different bands and values of PM associated from ground stations a correlation is calculated. The correlation between AOD MODIS band 3 and PM10 is 0.493 which is the highest value and this bond is used for model. Modelling After determining the best band to use MODIS AOD with the highest absolute correlation with PM10 levels, in order to increase the precision of the PM10 prediction in the final model the climate parameters were used in the modeling process. neural network model Recursive neural network model has input layer, hidden and output layers. In the current model in order to estimate the values of PM10 five input layer and an output layer exists. As the input layer parameters including temperature, relative humidity, wind direction, wind speed and the amount of AOD and output is the corresponding neural network model of PM10 per station at the time of imaging. The model input data sets from a variety of data, including satellite data, weather and so on. This random data network by 60% for training, 20% for validation and 20% for testing was used. Examples of training for an independent set of data stored, validation are used to find errors occurred during training. This will prevent excessive training. The test data in order to calculate the predicted values were then used to calculate RMSE and MAE. The testing date is another independent set of data which is stored in the neural network and used for the final evaluation. samples are not used in the production model. In this study, two Multilayer Perceptron Neural Network (MLP) and Radial Basis Function (RBF) models are used to predict and then the results are compared with each other. Validation In order to validate the model used in this paper, the two indicators of RMSE and MAE are used The results show that artificial neural network MLP values of RMSE and MAE values is less than the other. Values of 78.853 for RMSE and 22.776 for MAE are calculated for MLP model. So using this model to estimate the values of PM10 ground gives better results.
