Application of Land Use Regression Model to Predict Pollutants of NO2, CO, PM10 (Case Study: Tehran City)

Introduction 
Air quality in large cities is today one of the major problems and challenges in developed and developing countries. Occurrence and intensity of air pollution in the cities are influenced by a variety of factors such as pollution sources, meteorological factors and chemical reactions between the pollutants. There are different models to predict the air pollution concentrations in cities. These models can be classified into two groups: the models based on dispersion method and the models based on land use regression (LUR). The first research on LUR model was introduced by SAVIAH project sponsored by the European Union. This study was a multicenter project in Huddersfield and London (UK), Bilthoven (Netherlands), Prague (Czech Republic) and Warsaw (Poland). The aim of SAVIAH study was to develop and validate the methods for analysis of the relationship between air pollution and health on a small scale. After this research, several studies used the application of this model for modeling of urban air quality. The purpose of this present research is to forecast the concentrations of NO2, PM10 and CO in Tehran city using the land use regression in 2010. The independent variables such as land area, road network and meteorological variables have been used for modeling of these pollutants. Although 16 cases of air quality monitoring stations (AQMs) are located in Tehran city, limitations of monitored concentration of pollutants are different because of changes in traffic, land use, elevation and surrounding environment of the air quality monitoring stations.
 
 
Methodology
The areas of land-use and length of urban roads around the 16 AQMs have been measured by GIS techniques with input variables in land-use regression (LUR) models to explain pollution concentrations over space and time. These variables and other meteorological variables (surface and upper) have been calculated and used as explanatory factors. Pollution concentrations monitored at each AQM have also been used as the dependent variables. The areas of the five land uses including residential, commercial, industrial, transformational, and vegetative regions have been calculated using Tehran land use map obtained from Tehran municipality. These explanatory variables have been measured in 4 buffers and 16 sectors, and the wind-direction (WD) frequencies used to weight urban road length (WURL) and land uses (WLU). The meteorological factors generating chemical and physical reactions contribute to creation, destruction, and dispersion of the pollutants. Hourly measured values of temperature, humidity, and wind speed are seasonally summarized and included in the panel regression models to investigate these impacts. Eleven circular buffers from 500 to 2000 meters and sixteen sectors have been delineated around each AQM. URLs for entire transportation links have been calculated and then apportioned to each buffer and sector. WD frequencies have also been used to calculate WD-weighted URL (WURL). The same process has been applied to the five land uses (WLU). A panel data set has been created by the pooling of time-series and cross-sectional observations. It is also called as pooled dataset, time-series cross-sectional dataset, or longitudinal dataset. Regression models based on such data are called panel data regression models. Traffic flows are a key determinant of the concentrations of directly emitted and secondary pollutants. Since concentrations and traffic flows vary over space and time, it is proposed here to measure the spatiotemporal variations of the dependent and independent variables across geographical locations (AQMs) and hours of the day in a given region and period (season). As a reasonable proxy for traffic emissions, URL has been calculated for each buffer (ring and sector). Pollution concentrations display important differences between the four seasons. In order to compare the difference impacts of the explanatory variables on pollution concentrations across the four seasons, hourly concentrations are averaged over each season. This can generate four seasonal hourly panel data sets, each with 384 observations (16 AQMs × 24 hours). The four regression models have been formulated and their estimates are compared. Wind-direction-weighted URLs and land-use variables are recomputed for each season. Then, the best-radius-buffer for a variable is used. The proposed panel regression model is expressed as:
 
Where the indices and variables are defined as follows:
p: Pollutant (PM10, NO2, CO)
i: Cross-sectional observation (1 → 16 AQMs)
t: Time-series observation (1 → 24 hours)
C: Pollution concentration
X: Explanatory variables (for URL, four land uses, and four meteorological factors)
Uit: Error term for AQM i and hour t.
 
Results and discussion
The results of this research show differences among dependent and independent variables for each pollutant. Major urban road's length in four seasons has a positive impact on concentration of three pollutants but the impacts of land uses and meteorological factors are different in seasons. For example, in case of CO, the area of residential land use has a positive impact on the concentration in four seasons that is stronger in the winter. The green space areas have negative impact on concentration of CO that is impressive in summer and spring.
Influence of meteorological factors on the concentration of CO is negative for wind speed and positive for the upper air index (shelter) in the four seasons. Humidity impacts on CO concentration are positive in summer and negative in other seasons. In case of PM10, the industrial land use areas have positive and other land uses not efficient impact on the concentration in the four seasons. Impact of meteorological factors on concentration values of PM10 is negative in winter and positive in other seasons. Wind speed has negative impacts on the values in summer and spring and positive impacts in autumn and winter. In case of NO2, land use areas such as residential, commercial, industrial, and roads have positive impacts on NO2 concentrations. Among the meteorological factors, wind speed has unexpected impact on the concentration of NO2. The impact of this variable in four seasons is positive because of the chemical reactions among NO, O3 and NO2 that is prepared in low wind speed. Evaluation of model validity shows that there are more accurate predictions of CO and NO2 than PM10, particularly in spring and winter.

Conclusion
Application of land use regression model for Tehran city shows the high accuracy of the model for predictions of three pollutants in four seasons. The special features of this model are simplicity and not requiring complex data that enables its use in specific conditions.