Urban Heat Island Growth Modeling Using Artificial Neural Networks and Support Vector Regression : A case study of Tehran , Iran

Numerous investigations on Urban Heat Island (UHI) show that land cover change is the main factor of increasing Land Surface Temperature (LST) in urban areas. Therefore, to achieve a model which is able to simulate UHI growth, urban expansion should be concerned first. Considerable researches on urban expansion modeling have been done based on cellular automata. Accordingly the objective of this paper is to implement CA method for trend detection of Tehran UHI spatiotemporal growth based on urban sprawl parameters (such as Distance to nearest road, Digital Elevation Model (DEM), Slope and Aspect ratios). It should be mentioned that UHI growth modeling may have more complexities in comparison with urban expansion, since the amount of each pixel’s temperature should be investigated instead of its state (urban and non-urban areas). The most challenging part of CA model is the definition of Transfer Rules. Here, two methods have used to find appropriate transfer Rules which are Artificial Neural Networks (ANN) and Support Vector Regression (SVR). The reason of choosing these approaches is that artificial neural networks and support vector regression have significant abilities to handle the complications of such a spatial analysis in comparison with other methods like Genetic or Swarm intelligence. In this paper, UHI change trend has discussed between 1984 and 2007. For this purpose, urban sprawl parameters in 1984 have calculated and added to the retrieved LST of this year. In order to achieve LST, Thematic Mapper (TM) and Enhanced Thematic Mapper (ETM+) night-time images have exploited. The reason of implementing night-time images is that UHI phenomenon is more obvious during night hours. After that multilayer feed-forward neural networks and support vector regression have used separately to find the relationship between this data and the retrieved LST in 2007. Since the transfer rules might not be the same in different regions, the satellite image of the city has divided to several parts and for each part a specific CA model has defined. In the training step some pixels have randomly selected to calibrate the neural network and the regression. Then, using the trained neural network and support vector regression, LST in year 2007 has retrieved for all pixels. Results have indicated a great relationship between the simulated LST and the real one which has retrieved from thermal band of satellite image in 2007 (r = 0.843 for ANN method and r = 0.856 for SVR method). Although SVR caused to a better result, this method is much more time consuming than ANN method, especially when the number of training pixels increase.


INTRODUCTION
Difference in temperature between urban and its surrounding rural areas creates Urban Heat Island (UHI).Advent of remote sensing technology makes it possible to study more comprehensively this phenomenon.Lots of factors affect creation of this temperature discrepancy.Chen, et al., (2005) applied artificial neural networks to carry out the correlation between LST and factors like land-cover, population density, energy consumption and water ratio in urban area (Chen, et al., 2005).However, numerous studies on urban climate have shown that the most affective factor on UHI is the land-cover change especially conversion of vegetation and bare soil to concrete, asphalt and other man-made structures (Dousset & Gourmelon, 2003) (Pu, et al., 2006).On the other hand, other human activities especially those cause to burning fossil fuels, which increase the amount of carbon dioxide, may raise temperature in global scale in comparison with small scales (urban area).Therefore, to achieve a model which is able to simulate UHI, urban expansion should be concerned first.Considerable researches on urban expansion modeling have been done based on cellular automata (White, et al., 1997) (Gar-On Yeh & Li, 2002) (Alkheder & Shan, 2005).In CA models there is a grid of cells, each in one of a finite number of states.The goal of CA is modeling how these states of cells change in discrete time steps (Menard & Marceau, 2005).In this study, to find the CA model's transfer rules, Artifitial Neural Networks and Support Vector Regression have been implemented seperately and the ability and acuracy of these methods have been compared.

Study Area
Tehran is located in the North of central part of Iran (52°E, 35°N) with a population over 8 million.It has semi dry and semi cold climate.The elevation difference in the northern and southern parts of the city is about 900 meter which causes significant difference in the city's temperature.Increasing of the population in the last three decades causes to converting lots of surrounding natural land-covers to the urban structures.

Data Sets
TM and ETM+ on board Landsat4 and Landsat5 have spatial resolution of 30 meters in thermal band respectively.These sensors images are capable to be classified to main land-cover types and retrieve LST.Another remotely sensed data implemented here is SRTM 1 digital elevation model which provides other spatial information like slope and aspect ratios in addition to altitude as urban sprawl parameters.Aside from satellite images, auxiliary data such as road layers has been used for developing transportation network.

LST Retrieval
Each of the LANDSAT sensors has unique calibration coefficients.Digital numbers are converted to top-ofatmospheric TOA radiance ( ) using a linear equation (Weng, et al., 2004): G and b can be obtained from the header file of LANDSAT images.Then TOA radiance is converted to at-satellite brightness temperature with the following invers Planck's equation (Almeida, et al., 2008): Where coefficients and are shown in table1.But Difference in blackbody and land surface emissivity and also atmospheric effects make discrepancy between brightness temperature and land surface temperature.The usual method for driving land surface emissivity is using a vegetation index like NDVI.(Sobrino et al) proposed a method for retrieving emissivity from NDVI as follows (Sobrino, et al., 2004): According to the calculated emissivity, LST is retrieved from brightness temperature (Hu & Jia, 2009).
To perform atmospheric correction in sensors like ETM+ and TM which have only one thermal band, the primary method is to use radiative transfer model like LOWTRAN, MODTRAN and WINDOW, coupled with knowledge about atmospheric 1 Shuttle Radar Topography Mission structures like pressure, temperature and humidity profile (Kerr, et al., 2005).The existence of enough metrological stations that produce atmospheric information and the synchronism of this data with satellite imaging time are two major problems that can impact atmospheric correction methods using these radiative transfer codes.But in local studies like UHI monitoring there are some alternatives to do atmospheric correction.Xu, et al., (2009) apply a normalization technique and rescaling LST between 0 and 1 as follows: Where is the normalized for , is the value for pixel , and and are the minimum and maximum values of the image (Xu, et al., 2009).Another approach is to use Difference Land Surface Temperature ( ) instead of LST which can achieve as follows: Where is the mean temperature of a selected area in southwest of Tehran in which landcover types have not changed during last decades.The advantage of using or instead of is that some sources of systematic errors like radiometric calibration errors and atmospheric effects are partially removed in the differencing or normalization procedure (Fabrizi, et al., 2010) (Xu, et al., 2009).

Neural Networks for UHI Modelling
UHI growth may have more complexities in comparison with urban expansion (since the amount of each pixel's temperature should be investigated instead of its state).Artificial intelligence and especially neural networks have significant abilities to handle the complications of such a spatial analysis (Gar-On Yeh & Li, 2002).Although neural network has more flexibilities to solve this simulation (in comparison with other artificial intelligence methods like Genetic or Swarm intelligence), as a black box it has some disadvantages.As an illustration, the main problem is that neural network does not determine the role of each site attribute in land-cover change.
Here, the multilayer perceptron (MLP) neural network has been used to model how UHI has changed between 1984 and 2007 considering urban expansion.Therefore, defining suitable site attributes is crucial in proposed method.In this study, five spatial variables have been defined as urban sprawl parameters, which include: 1. Neighborhood development level which determines the number of urban pixels in a predefined neighborhood size.This parameter can derive from classified images.The Satellite image has been classified to the classes of urban area, bare soil and vegetation using maximum likelihood algorithm.Defining the size of the neighborhood is based on image spatial resolution and the difference between simulation time steps.

Distance to nearest road as an important urban sprawl
parameter that can greatly determine the direction of urban expansion.

Digital Elevation Model (DEM). This parameter affects
both urban expansion directions and LST 4. Slope which is derived from DEM and has influence on urban expansion and LST.Zhao, et al., (2007) discussed the relationship of north and south slope on LST.

Aspect.
This parameter which is also derived from DEM has unique effect on LST especially in rough areas because of the sun shadow.It should be noted that in urban expansion studies, it is not essential to use all the parameters of slope, aspect and elevation together, but as each of them has its unique impacts on temperature, we consider all of them here.Figure 1 shows the structure of proposed neural network for UHI simulation.In the training phase, a number of random pixels have been chosen.Site attributes for the selected pixels have been calculated and given to the neural network as its inputs.All site attributes have been scaled into the range of [0, 1] to make sure that all parameters are equally affective in neural network model (Sobrino, et al., 2004).The output of neural network, in the training step, is LST in year 2007 for the selected pixels, which have been obtained from Landsat7 ETM+ night-time image.Therefore the neural network has six neurons in its input layer and one neuron as output layer.There is no specific way to determine number of hidden layers and their neurons.However, there are some suggestions that can help to design neural network hidden layers.For instance in Kolmogorov's theorem a continuous function Ø:X n → R c can be represented by a neural network with n neurons in its input layer, (2n+1) neurons in its single hidden layer and c neurons which forms the output layer (Almeida, et al., 2008).Gar-On Yeh, et al., (2002) indicated that a network with (2n/3) neurons in one hidden layer can generate results of almost the same accuracy level, compared with (2n+1) neurons in the hidden layer but consumes much less time for training (Gar-On Yeh & Li, 2002).In this paper multiple neural networks with different number of hidden layers and neurons were tested.There wasn't significant difference between them.The proposed neural network has two hidden layers with eight and four neurons in its first and second ones respectively.It should be noted that large number of hidden layers and neurons can extremely increase the training time.

Support Vector Regression for UHI Modelling
Support Vector Regression (SVR) is a new supervised learning method which emerged in late 1970s (Vapnik, 1979).SVR allows computing a powerful nonparametric approximation of the relationship between site attributes and UHI change.This technique is also used in many remote sensing applications like LST and SST retrieval (Moser, et al., 2009), leaf area index estimation and other biophysical parameter estimation from multispectral satellite images (Durbha, et al., 2007).The most advantage of SVR, in comparison with ANN, is its ability to find global optimized result instead of local optimized one (Tso & Mather, 2009).Here we use Gaussian radial basis function as the SVR Kernel:

IMPLEMENTATION
Satellite images can explicitly display how Tehran has been expanded since 1984.Figure 2a  After training step, the fixed neural network and trained support vector regression, are able to simulate UHI.To do this, site attributes for all of the pixels have been given to them and the simulated UHI in 2007 has been computed.Figure 3a

CONCLUSION
Satellite thermal images have been used for simulation of Tehran UHI growth.As land-cover types extremely affect UHI, the developed simulation model is based on urban expansion.Artificial neural networks and support vector regression were used separately to find the relationship between LST in 1984 and 2007 by considering urban sprawl parameters.Great compatibility of ANN and SVR techniques, cause to strong correlation between real and simulated LSTs (r ≈ 0.844 for ANN and r ≈ 0.856 for SVR).However, as a disadvantage of neural networks and support vector regression, proposed models couldn't explicitly define the role of each site attribute on UHI growth.Simulation with different number of training pixels indicated that when less training pixels are used, SVR has a better estimation than ANN.This is because SVR method is more capable in finding the global optimized results.Additionally, different sizes of windows have been implemented for determination of neighborhood development level.The best result was achieved for the 41×41 pixel size window (20 radius neighborhood).
Figure 3. (a) in 2007 derived from ETM+ nighttime image, and (b) simulated computed from proposed model

Table 1 . coefficients of inverse of Plank's equation for Landsat sensors
Table1indicates that the best radius for computing neighborhood is 20 pixels.However, as it was mentioned before, the proper neighborhood size depends on data spatial resolution and difference between simulation time steps.Using the trained neural network it is feasible to simulate UHI at next time step (2030).In this case, neighborhood development level and distance to nearest road should be recomputed for year 2007, and of input layer should be derived from ETM+ 2007 thermal band.Other site attributes in input layer are invariant.