تخمین مکانی – زمانی آلاینده‌های منواکسید کربن و دی‌اکسید نیتروژن شهر تهران مبتنی بر داده‌های حاصل از سنجش‌ازدور و داده‌های کمکی

آلودگی هوا یکی از پیامدهای ناهنجار فعّالیت های بشر است که نه‌تنها سلامت انسان را تهدید می‌کند؛ بلکه بر همه عوامل محیط‌زیست ازجمله گیاهان و جانوران تاثیر نامطلوب می‌گذارد. تهران به‌عنوان مرکز اداری، سیاسی و اقتصادی کشور و پرجمعیّت ترین شهر ایران، یکی از آلوده‌ترین شهرهای دنیا به شمار می رود. از مدل‌های خطّی و غیر خطّی متعدّدی تاکنون به‌منظور مدل سازی آلودگی هوا استفاده شده است. در نوشتار پیش رو از ویژگی های مکانی و زمانی مستخرج از تصاویر سنجش‌ازدور و داده های محیطی ایستگاه های پایش آلودگی هوا سازمان محیط‌زیست واقع در سطح شهر تهران پس از پیش‌پردازش‌های لازم به مثابه ورودی مدل استفاده شد. ازمیان ایستگاه های آلاینده سنج موجود در سطح شهر تهران، با درنظرگرفتن پوشش سری زمانی مشترک داده های ثبت‌شده، تعداد هشت ایستگاه انتخاب شد. به منظور انجام فرایند مدل‌سازی از روش شبکه عصبی مصنوعی پرسپترون چندلایه با الگوریتم آموزش لونبرگ – مارکوارت و تابع فعّال‌سازی سیگموئیدی استفاده شد. در پژوهش حاضر از داده های هواشناسی، داده‌های مربوط به غلظت آلاینده‌ها در روزهای قبل، کاربری اراضی و نیز داده‌های مستخرج از تصاویر ماهواره‌ای شامل داده‌های مربوط به پوشش گیاهی و جزایر حرارتی به‌منظور مدل‌سازی غلظت آلاینده‌ها استفاده شد. از روش تبدیل موجک بر روی مقادیر غلظت آلاینده ها در روزهای قبل استفاده گردید و سپس روش انتخاب ویژگی جنگل تصادفی بر روی ویژگی های ورودی مدل اعمال شد؛ همچنین با توجّه به تغییرات مکانی آلودگی هوا سعی بر آن شد که با استفاده از اطّلاعات هفت ایستگاه، مقادیر غلظت آلاینده یک ایستگاه برآورد شود. نتایج حاصل از ارزیابی مدل بیانگر کارابودن مدل ارائه‌شده در تخمین مقادیر بیشینه روزانه غلظت آلاینده بود. منواکسید کربن و دی‌اکسید نیتروژن به ترتیب با خطای 13% و 11.5% به‌صورت زمانی پیش‌بینی شدند؛ همچنین این دو آلاینده به‌صورت مکانی با خطای تخمین کمتر از 17% پیش‌بینی شدند.

Spatio – Temporal Estimation of Carbon Monoxide and Nitrogen Dioxide based on Remote Sensing Data and Ancillary Data in Tehran

Air pollution is one of the most important consequences of human activities, which not only threatens human health but also negatively affects all elements of the environment, including plants and animals. Tehran, the capital of Iran, and the administrative, political and economic center of the country, is no exception which is constantly struggling with these hazard. So far, many linear and nonlinear models have been applied to model air pollution. In this research, 8 pollutant measurement stations distributed over Tehran were selected according to the availability of their recorded data. In order to provide a model predicting pollutants, spatially and temporally, the combination of spatial and temporal features extracted of remote sensing data and environmental data was modeled using multilayer perceptron artificial neural network. The input data include meteorological data, topography, traffic index, population data, air pollutant concentrations for the last days, and land use map. In addition, vegetation cover, distance from heat islands, and the land surface temperature derived from remotely sensed data were used as remotely sensed attributes. In order to increase the accuracy of modeling, wavelet transform and feature selection methods were used on input attributes of the model. Random forest feature selection method was applied on the input data in order to reduce the number of input attributes,. The results of the model evaluation indicated that the model was efficient in estimating the concentrations of pollutants. Temporally, carbon monoxide and nitrogen dioxide were predicted with error estimation of 13% and 11.5%, respectively. Besides, these pollutants were spatially predicted with the estimation error less than 17%.  Extended English AbstractIntroduction: Each air pollutant has its own temporal and spatial characteristics, based on its concentration and type. Therefore, several studies have been conducted by deterministic and statisticalempirical methods for spatiotemporal modelling of air pollutants, till now. The statisticalempirical methods include linear and nonlinear models for which there are advantages and disadvantages. Since nonlinear methods such as artificial neural network (ANN) are able to find the complex nonlinear relationships between dependent and independent variable, they are usually applied for statistical nonlinear modelling. There are different variables that have been used as independent variables for modelling of air pollutant concentrations in different studies. According to the literature, there are no many studies examining the possibility of the use of spatiotemporal models of air pollutant concentrations for the places and times for which they are not developed; therefore, this study aims to assess the performance of the spatiotemporal models to predict air pollutant concentrations for the other times and places. As, ANN multilayer perceptron efficient performance has been proved by the other studies, this method has been applied for spatially and temporally modelling carbon monoxide and nitrogen dioxide as the main air pollutants in Tehran. Materials and Methods: The study area of this research is Tehran. This city with 800 km2 is between 35° 34´ to 35° 49´ latitude north, and 51° 04´ to 51° 36´ longitude east. This city locates in the southern side of Alborz Mountain and northern boundary of central desert of Iran. Environmental spatial data were used along with satellite imagederived data in this study. Environmental data include air pollutant concentrations, meteorological data, traffic data, land use map and population data. Landsat8 image was applied as remotely sensed data in this study. Moreover, wavelet transformation is used on the air pollutant concentration data of four days before the prediction date. Among 156 generated attributes, 42 attributes were extracted by random forest feature selection. 3048 samples of maximum daily pollutant concentrations were used to model pollutant concentrations for 8 pollutant measurement stations around Tehran in 18 months.  These samples were randomly divided into three portions including 70% for training, 15% for validation and 15% for test. For spatial modelling, samples of seven stations (90% for training and 10% for validation) were applied for modelling and the samples of one station were used for test of the model, and it was repeated for each station, separately. Root mean square error, standard error of estimation, coefficient of determination, and error percentage were used to assess the models.Results and Discussion: According to the results, temporal models were able to predict the variation of the carbon monoxide and nitrogen dioxide concentrations with coefficient of determination equals to 0.93 and more than it. As the results showed the performance of the model predicting the maximum concentration of the nitrogen dioxide with error percentage of 11.46% was better than that predicting carbon monoxide concentration. Besides, the findings indicated that the spatial models predicted the maximum concentration of carbon monoxide with highest and lowest accuracy for Salamat Park and Cheshmeh stations, respectively. In addition, while, maximum concentration of nitrogen dioxide was predicted with the lowest error percentage at Shokoofeh station, it was predicted with the highest error percentage at Shahid Beheshti University station. It was shown that nitrogen dioxide concentration was predicted more accurate than carbon monoxide concentration which is due to the higher variation of carbon monoxide concentration compared to nitrogen dioxide. Moreover, since, the sources of the pollutants differ at each station, the performance of the models vary for different stations. Furthermore, the accuracy of the models depends on the accuracy, number and density of the air pollutant measurement stations; this can be another reason for the variation in the performance of the spatially predicting models at different stations.   Conclusion: In this study, ANN multilayer perceptron was fed by environmental and remotely sensedderived attributes, was applied for maximum daily concentrations of carbon monoxide and nitrogen dioxide, spatially and temporally. Nitrogen dioxide concentration was predicted better than carbon monoxide concentration as nitrogen dioxide concentration variation was lower than the other pollutant. Therefore, ANN multilayer perceptron performs better for predicting the pollutant which is more stable in the air. Besides, ANN model performs better for the temporally modelling than the spatially one. In addition, the architecture of ANN is different for spatially and temporally models. Finally, the models derived for different stations perform differently in terms of the accuracy.