Volume 6, Issue 6, June – 2021 International Journal of Innovative Science and Research Technology

ISSN No:-2456-2165

Urban Heat Island Prediction Using ANN

Gitika Kaur1, Garima Sharma2, Dr. Meenu Vijarania3, Swati Gupta4
1
Amity University Haryana ,122413, India

Abstract:- A urban heat island (UHI) is an urban region

or territory that is fundamentally hotter than its
encompassing rural zones because of man practices. The
UHI can be explained as either the surface temperature
difference or air temperature difference between the
urban and the rural regions. The project is based on the
study of “Urban Heat Island prediction using ANN”
grounded on the values of various test results done on an
individual technique. The technique used for this project
is the Artificial Neural Network (ANN) algorithm and
Time Series analysis to correctly predict a region as
urban heat island. The data of six weather stations (WS)
was collected each having 12929 instances to determine
whether a particular region forms UHI or not. The ANN Figure 1: An illustration of an urban heat island.
model was implemented on all the six weather stations
and after examining the ANN model, the two weather 1.2 Why does this happen?
stations WS4 and WS6 proved out to be the best in terms The primary reason of the UHI impact is from the
of correlation between dependent and independent change of land surfaces. There are a few reasons for an UHI;
variable that was evaluated using MAE and R2 score. for instance, dark surfaces retain essentially progressively
The two weather stations WS6 and WS4 having R2 score sunlight-based radiation, which makes urban groupings of
or accuracy as 79.6 and 79.3 respectively was further streets and structures heat extra than rural and provincial
chosen for time series analysis. In time series analysis, we territories throughout the day.
just have one variable i.e., time. We can analysis the time
series data in order to extract meaningful insights and Materials regularly utilized in urban regions for
other features. Time series is a set of observation taken asphalt and rooftops, for example, cement and black-top,
at specified times at equal interval. have fundamentally extraordinary thermal mass properties
and surface radiative properties than the encompassing rural
Keywords:- Urban Heat Island Prediction, ANN, Time regions.
Series Analysis.
1.3 Machine Learning and UHI
I. INTRODUCTION
Machine learning (ML) is the process of making our
Temperatures in urban zones have for quite a while machine or model learn the correlations between various
been concentrated as a limit of metropolitan morphology, input and output features and based on what the model
land use and anthropogenic activities than just focus learns it is capable of making predictions on a new data. [2]
meteorological boundaries. The wonder of urban heat island
is currently an ongoing issue of worry in numerous urban
areas of created as well as creating countries of the world.

An urban heat island (UHI) is a metropolitan region or

territory that is fundamentally hotter than its encompassing
rural zones because of man practices. The word heat Figure 2: Machine Learning
island is used to reflect any area that is comparatively hotter
than the surrounding, but generally describe human- ML has seen a tremendous growth in predictions of
disturbed areas. The temperature difference is generally UHI because of its high accuracy and large datasets.
bigger around evening time than throughout the day, and is Number of machines learning algorithms are being used by
most obvious when breeze [1] researchers in the prediction of UHI and estimating the
important parameters causing UHI.

IJISRT21JUN413 www.ijisrt.com 252

Volume 6, Issue 6, June – 2021 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
1.3 Deep Learning (DL) II. RELATED WORK
Deep learning is an artificial intelligence that imitate
the functions of the human mind in processing data and A lot of work has been done on this particular topic of
pattern creation. DL is a subset of machine learning which is urban heat island prediction. This section survey some of the
based on ANN (artificial neural networks). So, deep related work already done in this field.
learning is a kind of mimic of human brain.
In recent years the UHI phenomenon has become one
The whole concept of deep learning is to mimic how of the most growing environmental problems in the urban
human brain works. It involves an input, hidden and output areas due to the land cover, urbanization and population
layer. We use activation functions in neural networks. density [4]. UHI can be assessed by applying thermal
Usually, rectifier function is used in the hidden layers and remote sensing to perform land cover classifications and
for output layers we use sigmoid.[3] thermal behavior of various urban surfaces which varies in
response to the surface energy balance [5]. Besides that,
We have parameters like batch size and epochs which UHI detection can also be performed using fixed-station or
can be changed so as to get much higher accuracies. meteorological station and car traverse measurement to
detect UHI intensity and the air temperature in the regions
Deep learning algorithms are: [6].
 Artificial Neural Networks (ANN) – used for
regression and classification. Several research studies have been performed on the
 Convolutional Neural Networks (CNN) – used for UHI phenomenon. The UHI intensity differs between
dealing with images. seasonal climate and tropical cities. In the seasonal climate,
 Recurrent Neural Networks (RNN) – applicable in time the UHI are strongest in the summer or winter season.
series analysis. According to [7,8], the UHI intensity is positively correlated
with solar radiation and relative humidity during summer
1.5 Artificial Neural Network while on the other hand it is negatively correlated with wind
The whole perception of ANN is to mimic how human speed and relative humidity during winter.
brain works. ANN involves an input layer, hidden layer(s),
and an output layer. Each of these layers have a number of The urban heat island (UHI) phenomenon serves as a
neurons and synapses which are responsible for propagating trap for atmospheric pollutants, deteriorates the quality of
input data through our network. In the ANN we have the life and has a socio-economic impact in the urbanized areas
unit of calculation which are called neurons. These neurons [9,10]. Important research has been accomplished over the
are attached by the weight values called as synapses. This last hundred years to quantify its impact on the urban
means given a number, a neuron will perform some climate [11,12]. Various heat island studies have been
calculations with the help of function known as activation performed in Europe during the last 15 years [13]. Urban
function. heat island and increased urban temperatures [14,15,16],
exacerbate the cooling load of buildings, increase the peak
electricity demand for cooling and decrease the efficiency of
air conditioners [17].

Moreover, the urban agglomeration has a negative

impact on the cooling effectiveness of natural and night
ventilation [18]. [19], was the first to relate UHI intensity to
meteorological elements such as cloudiness, wind speed,
temperature, and absolute humidity using a multiple linear
regression method. He showed that cloudiness and wind
speed parameters are negatively correlated with the UHI
Figure 3: Artificial Neural Network intensity in Uppsala, Sweden, and that the total variance
explained by the regression model is larger in the nighttime
In the project, the Feed Forward ANN has been than in the daytime.
implemented. In this, the flow of information is
unidirectional. A unit sends data to other unit from which it Consequently, the prediction of the urban heat island
does not get any data. There is no feedback allowed. They behavior has gained a significant attention. Although a
have fixed number of inputs. number of modelling approaches for urban heat island do
exist [20], the complexity of the phenomenon, the bulk of
This paper is ordered as follows: section 2 related urban details required to attain an accurate urban model and
work reviews about the great work done in the area of urban the increased cost and computational time of analytical
heat island detection model, section 3 deals with methods modelling approaches has led to the exploration of other
and materials used, section 4 carries out the actual prediction methods. Artificial neural networks (ANNs) have
implementation work done followed by section 5 contains been used in a number of prediction studies that involves
results and analysis of the model, section 6 concludes with atmospheric time series data. [21], predicted daily maximum
some directions for future work. ozone levels in Texas metropolitan areas with a standard

IJISRT21JUN413 www.ijisrt.com 253

Volume 6, Issue 6, June – 2021 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
three-layer ANN model with nine inputs and four hidden Both of these techniques helped in getting a much
nodes and found it to be superior to statistical methods higher accuracy over the test data.

III. METHOD AND MATERIALS IV. IMPLEMENTATION

Based on the previous work done by all the This section describes the dataset used for the project,
researchers, this research-based project will provide with data preprocessing phase and a brief of all the classification
ANN algorithm and time series analysis for the prediction of algorithms used.
UHI.
4.1 Data Preprocessing
The dataset is collected from different areas of This phase involved dealing with missing values,
Rajasthan and the dataset consist of four input parameters Dummy variables, standardization, etc. Work done on the
and an output or dependent parameter. Dataset we choose is above-mentioned dataset:
different from the existing datasets as in the existing ones  Checked for any missing values.
we have more parameters on the basis of which the  Dummy variables were created to deal with categorical
calculation was performed whereas we took the 4 major values.
parameters for our calculations. The input parameters are the  StandardScaler object was applied on the dataset.
maximum temperature, minimum temperature, wind
velocity, relative humidity and the dependent that is the 4.2 Build an ANN Model
output parameter is the solar radiations. The dataset totally In the project, the ANN model comprises of input
comprises of 6 different areas having all the input layer, 3 hidden layer, and 1 output layer. The total quantity
parameters in each of the workstation. The data consist of of nodes in the hidden layers was set to (40,40,30)
12929 instances. The collected data is a day wise collection respectively and epoch was set to 30. The activation
which starts from 01-01-1979 to 31-07-2014. function for the hidden layer was set as a rectifier function
(relu) and sigmoid function for the output layer.
Inside the training and testing directory data belonging
to each class is present. Training and testing images are The ANN model was used for data classification using
present in separate directories inside the main directory. The keras. Compiling the model is done by Adam optimizer and
research work used libraries part of Anaconda4 framework. loss is “categorical_crossentropy.

The architecture of our model is:

model = Sequential ()
model.add(Dense(40, input_shape=(4,), activation = 'relu'))
model.add(Dense(40, activation='relu'))
model.add(Dense(30, activation='relu'))
model.add(Dense(1,))
model.compile(loss='mean_absolute_error',
optimizer='adam', metrics=['mean_absolute_error'])
history = model.fit(X_train, y_train, epochs = 20, verbose =
1)

Figure 4: Flow chart of the model

GridSearchCV – This is a technique used for finding out

the best values of hyper-parameters. The GridSearchCV has
been used to find out the best number of hidden layers and
their respective number of nodes.

K-Fold Cross Validation – This is a model evaluation

technique. This consists of dividing the training dataset into Figure 5: ANN model
“K” folds and then training the model on “K-1” folds and
testing is done on the remaining one-fold. 10-Fold Cross The implementation steps are as follows:
Validation was used to find out the optimal values of hyper- a. Data collection and importing of dataset. The dataset
parameters in each of the classification algorithms totally comprises of 5 different parameters in each of the
implemented. workstation. Each work station consists of 12999

IJISRT21JUN413 www.ijisrt.com 254

Volume 6, Issue 6, June – 2021 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
instances. Table 1. Model Evaluation Results
b. The input parameters are Maximum temperature, Weather MSE Train Accuracy Test Accuracy
minimum temperature, wind velocity, relative humidity. Stations (R2score) (R2 Score)
c. Output parameters is the solar radiations. 1 1.8570 0.782 0.761
d. Pre-processing the dataset, dealing with missing values, 2 1.8380 0.797 0.773
Dummy variables, standardization, normalization, etc. 3 1.8210 0.789 0.768
e. Forming training and test set, using the train_test_split() 4 1.7687 0.772 0.796
method to split training and testing data 5 1.7964 0.808 0.778
f. Building an ANN model, for each Work station. We 6 1.8120 0.808 0.793
Build 6 different ANN models.
g. The ANN model has been implemented on the real time Table 1 shows the values of MAE and R2 score
data of 6 WS which consists of the inputs (independent calculated for each weather stations.
variables) as Maximum Temperature, Minimum
Temperature, Relative Humidity and Wind velocity. The 5.2 Effect of Solar Radiation and wind velocity on UHI
output (dependent variable) is Solar Radiation. Many researches have shown that the direct solar
h. Compiling the model is done by Adam optimizer and radiation has a positive effect on UHI and the wind velocity
loss is “categorical_crossentropy”. The activation has a negative outcome on the UHI. It means that if solar
function for the hidden layer was set as a rectifier radiation increases then UHI may form on that region and
function (relu) and sigmoid function for the output similarly, if the wind velocity decreases then the UHI may
layer. not form on that region.
i. Subsequent to building the model design, we train the
model using model.fit(). The model performs better 5.3. Time Series Analysis Results
After 20 epochs. The time series analysis plot of solar radiation v/s time
j. GridSearchCV and K-Fold Cross Validation helped in and wind velocity v/s time has been plotted for WS4 and
getting a much higher accuracy over the test data. WS6. After visualizing the plots, it has been found that the
k. After implementing the ANN model for all the 6 WS, it plot corresponding to the WS6 i.e (Wind v/s time) shows
has been found that the WS4 and WS6 shows the best that wind velocity decreases over the time and (Solar
correlation between the solar radiation and the input radiation v/s time) shows that solar radiation significantly
parameters based on the MAE and R2 score which are increases over the time.
used as model evaluation tools. So, we can say that the region corresponding to the
l. Applying the time series analysis on WS6 and WS4 in WS6 forms the UHI over the time.
order to find out among then which workstation is more
likely to form an Urban heat island (UHI). 5.5 Graphical Representation of Results A.
m. Many statistics techniques like Rolling statistics and Plot of WS 6
Dickey Fuller test have been applied on WS 6 and WS4
on parameters wind velocity and solar radiations, in
order to extract meaning full insights from the data.
n. Visualizing the time series plot. Autocorrelation &
partial correlation graphs of solar radiation and
Autocorrelation & partial correlation graphs of wind
velocity for WS6 as well as WS4 is plotted.

V. RESULTS AND ANALYSIS

5.1 ANN Results:

The aim of the ANN model is to find the WS which
shows the best correlation between the solar radiation and Figure 6: Graphical representation of wind velocity dataset.
the input parameters. And resulted weather stations has been
used for time series analysis to predict the UHI.

After implementing the ANN model for all the 6 WS,

it has been found that the WS4 and WS6 shows the best
correlation between the solar radiation and the input
parameters based on the MAE and R2 score which are used
as model evaluation tools.

Figure 7: Graphical representation of solar radiation

dataset.

IJISRT21JUN413 www.ijisrt.com 255

Volume 6, Issue 6, June – 2021 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165

Figure 11: Graphical representation of solar radiation

dataset.

Figure 8: Autocorrelation and partial correlation graphs of

solar radiation.

Figure 12: Autocorrelation and partial correlation graphs of

solar radiation.

Figure 9: Autocorrelation and partial correlation graphs of

wind velocity.

B. Plot of WS4

Figure 13: Autocorrelation and partial correlation graphs of

wind velocity

As we can see the autocorrelation graphs for both the

work stations (WS) 4 and 6, In work station 4 there is a
decrease in solar radiation with increase in lags whereas in
WS 6 solar radiation the decreasing factor is not to that
extend. So, by this we can predict that WS6 has more
chances of being an urban heat island.
Figure 10: Graphical representation of wind velocity
dataset.

IJISRT21JUN413 www.ijisrt.com 256

Volume 6, Issue 6, June – 2021 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
On the other hand, when it comes to wind velocity, we [6]. Kolokotsa D., Psomas A. Karapidakis E., (2009),
can see in the auto correlation graphs that it is high WS 4 Urban heat island in southern Europe: The case study
than in WS 6. Thus, by all the graphs we can conclude that of Hania, Crete, Solar Energy, 83, 1871-1883.
WS 6 can be predicted to be an urban heat island. [7]. Liu W., Ji C., Zhong J., Jiang X., Zheng Z., (2007),
Temporal characteristics of the Beijing urban heat
VI. CONCLUSION [8]. island, Theoretical and Applied Climatology, 87, 213-
221.
In conclusion, this research-based project dealt with [9]. Papanastasiou D.K., Kittas C., (2012), Maximum
ANN algorithm and time series analysis for the prediction of urban heat island intensity in a medium-sized coastal
UHI and comparing with other traditional methods. K-Fold Mediterranean city, Theoretical and Applied
Cross Validation and GridSearchCV were also applied so as Climatology, 107, 407-416.
to increase the accuracies of the models. [10]. Santamouris, M., Paraponiaris, K., & Mihalakakou, G.
(2007). Estimating the ecological footprint of the heat
After examining the ANN model for 6WS, the two island effect over Athens, Greece. Climate Change, 80,
weather stations WS4 and WS6 proved out to be the best in 265–276.
terms of correlation between dependent and independent [11]. Santamouris, M., Pavlou, K., Synnefa, A., Niachou,
variable that was evaluated using MAE and R2, R2 score or K., & Kolokotsa, D. (2007). Recent progress on
accuracy as 79.6 and 79.3 respectively was further chosen passive cooling techniques advanced technological
for time series analysis. It has been concluded from the time developments to improve survivability levels in low-
series analysis plot that WS6 forms the UHI in comparison income households. Energy and Buildings, 39, 859–
to WS4. 866
[12]. Akbari, H., Konopacki, S., & Pomerantz, M. (1999).
During the development process, we studied carefully Cooling energy savings potential of reflective roofs for
and understanding the criteria for making it more effective residential and commercial buildings in the United
with a high accuracy percentage. We also realized the States. Energy, 24, 391–407.
importance of maintaining a token margin for error. [13]. Mihalakakou, P., Flokas, H., Santamouris, M., &
Helmis, C. (2000). Application of neural networks to
There are many places we can improve. The following the simulation of the heat island over Athens Greece
are the features that we are about to implement in near using synoptic types as a predictor. Journal of Applied
future. Meteorology, 41, 519–527.
 Forecast the solar radiation or the wind velocity for the [14]. Santamouris, M.(2007). Heatisland research in Europe:
coming years. The state ofthe art. Advances in Building Energy
 A technique to forecast in how many years an area will Research, 1, 123–150.
become UHI. [15]. Livada, I., Santamouris, M., & Assimakopoulos, M. N.
 Improve the efficiency of this algorithm. (2007). On the variability of summer air temperature
during the last 28 years inAthens. Journal
VII. ACKNOWLEDGMENTS ofGeophysical Research, 112, D12103.
[16]. Livada, I., Santamouris, M., Niachou, K.,
The authors would like to thank Google for quick and Papanikolaou, N., & Mihalakakou, G. (2002).
efficient results on selected keywords, Google Determination of places in the great Athens area where
Collaboratory. Sincere thanks to mentors for guiding well the heat island effect is observed. Theoretical and
and helping from design and development of deep learning Applied Climatology, 71(3–4), 219–230.
model. [17]. Mihalakakou, M., Santamouris, N., & Papanikolaou,
C. (2004). Cartalis simulation of the urban heat island
REFERENCES phenomenon in Mediterranean climates. Pure and
Applied Geophysics, 161, 429–451.
[1]. https://www.nationalgeographic.org/encyclopedia/urba [18]. Cartalis, C., Synodinou,A., Proedrou, M.,
n-heat-island/ Tsangrassoulis,A., & Santamouris, M.(2001).
[2]. https://www.ibm.com/cloud/learn/machine-learning Modifications in Energy Demand in urban areas as a
[3]. https://www.investopedia.com/terms/d/deep- result of climate changes: An assessment for the
learning.asp southeast Mediterranean region. Energy Conversion
[4]. Sarkar H., (2004), Study of Landcover and and Management, 42(14), 1647–1656.
Population Density Influences on Urban Heat Island in [19]. Geros, V., Santamouris, M., Karatasou, S.,
Tropical Cities by Using Remote Sensing and GIS: A Tsangrassoulis, A., & Papanikolaou, N. (2005). On the
Methodological Consideration, 3rd FIG Regional cooling potential of night ventilation techniques in the
Conf., Jakarta, Indonesia. urban environment. Energy and Buildings, 37(3), 243–
[5]. Chudnovsky A., Ben-Dor E., Saaroni H., (2004), 257.
Diurnal thermal behavior of selected urban objects [20]. Sundborg, A., 1950: Local climatological studies of
using remote sensing measurement, Energy and the temperature conditions in an urban area. Tellus, 2,
Buildings, 36, 1063-1074. 222–232.

IJISRT21JUN413 www.ijisrt.com 257

Volume 6, Issue 6, June – 2021 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
[21]. Mirzaei, P. A., & Haghighat, F. (2010). Approaches to
study urban heat island—Abilities and limitations.
Building and Environment, 45(10), 2192–2201.
[22]. Yi, J., & Prybutok, V. R. (1996). A neural network
model forecasting for prediction of daily maximum
ozone concentration in an industrialized urban area.
Environmental Pollution, 92(3), 349–357.

IJISRT21JUN413 www.ijisrt.com 258