Sustainable Cities and Society 73 (2021) 103128

Contents lists available at ScienceDirect

Sustainable Cities and Society

journal homepage: www.elsevier.com/locate/scs

Evaluating the role of urban fabric on surface urban heat island: The case
of Istanbul
Deniz Erdem Okumus a, *, Fatih Terzi b
a
Department of Urban and Regional Planning, Yildiz Technical University, Istanbul, Turkey.
b
Department of Urban and Regional Planning, Istanbul Technical University, Istanbul, Turkey.

A R T I C L E I N F O A B S T R A C T

Keywords: Urban heat islands, one of the fundamental anthropogenic impacts on local climates, have been a growing
Land surface temperature concern especially for high-density urban areas such as Istanbul. This paper outlines the use of a supervised
Surface urban heat island machine learning technique to understand the effects of the urban fabric on surface urban heat island (SUHI)
Ridge regression model
formation in Istanbul, and identify effective variables to provide a basis for research and practice focusing on
Urban fabric
Istanbul
SUHI mitigation. An analysis using the Ridge Regression Model found that 71% of land surface temperature
anomalies in Istanbul are linked to building coverage ratio (BCR), surface/volume ratio (SVR), sky-view factor
(SVF), canyon geometry factor (CGF), and vegetation index (NDVI). NDVI and BCR were the urban fabric
components with the highest contribution to SUHI formation, while the effects of SVF and CGF remained rela­
tively low. This research can help planners and designers gauge the contribution of the urban fabric to micro-
climate issues and adapt SUHI mitigation strategies for projects aiming to build climate-sensitive urban envi­
ronments. It also provides insight into and improves knowledge of supervised machine learning approaches to
the urban planning and design disciplines.

1. Introduction Shishegar, 2013; Todhunter, 1990; Yin et al., 2018). Building coverage
ratio and floor area ratio are amongst the frequently used indicators of
The urban heat island (UHI) phenomenon draws the attention of building density (Giridharan, Lau & Ganesan, 2005; G. Guo et al., 2016).
researchers in various disciplines to take part in discussions of mitiga­ Density indicators are generally presented as positively correlated to
tion solutions. Urban planning and design, two intertwined disciplines, LST (Giridharan et al., 2005; G. Guo et al., 2016); however, Stone Jr and
are particularly interested in both the role of the urban fabric on land Rodgers (2001) demonstrated that low-density urban patterns may
surface temperature (LST) variations and surface urban heat island create higher temperatures than high-density development patterns.
(SUHI) mitigation strategies. The urban fabric comprises the spatial Besides that, indicators of geometrical patterns may also have both
configurations and physical forms of two and three-dimensional urban negative and positive correlations with the micro-thermal climates of
components, including urban geometry, building density, and vegeta­ street canyons (Arnfield, 1990; Giridharan et al., 2007; G. Guo et al.,
tion coverage (Boyko & Cooper, 2011; Chokhachian, Perini, Giulini & 2016; Hu et al., 2016; Oke, 1981; Shishegar, 2013; Strømann-Andersen
Auer, 2020; A. Guo et al., 2020). Urban geometry and building density & Sattrup, 2011; Yin et al., 2018). For instance, while Oke (1981) argues
are amongst the major contributors to LST variations in cities, as they that low sky-view increases the SUHI effect because obstructions of the
are strongly related to the redirection of natural ventilation (He, Ding & open sky delay the cooling of the surface, Giridharan et al. (2007)
Prasad, 2020a, 2020b; Merlier, Kuznik, Rusaouën & Salat, 2018), cre­ suggest that the restricted access of solar radiation to urban surfaces in
ation of airflow potential for heat loss (G. Guo, Zhou, Wu, Xiao & Chen, low sky-view environments may have the opposite effect. Entirely
2016; Yin, Yuan, Lu, Huang & Liu, 2018) and regulation of the amount different opinions on the same factor open up the possibility that specific
of solar radiation reflected (EPA, 2008; X. Yang & Li, 2015; Yin et al., characteristics of each urban environment contribute differently to SUHI
2018). The geometrical pattern of an urban area is generally measured formation.
by canyon geometry factor, sky-view factor, surface/volume ratio. SUHI is also closely related to and negatively correlated with vege­
(Giridharan, Ganesan & Lau, 2004, 2007; Hu, White & Ding, 2016; tation coverage (Guha, Govil, Dey & Gill, 2018; Li, Li & Duan, 2016;

* Corresponding author: Deniz Erdem Okumus. Yildiz Technical University, Faculty of Architecture, Barbaros St., 34349 Yildiz-Besiktas-Istanbul.
E-mail addresses: denizer@yildiz.edu.tr, denizerdemokumus@gmail.com (D. Erdem Okumus), terzifati@itu.edu.tr (F. Terzi).

https://doi.org/10.1016/j.scs.2021.103128
Received 26 February 2021; Received in revised form 27 June 2021; Accepted 30 June 2021
Available online 4 July 2021
2210-6707/© 2021 Elsevier Ltd. All rights reserved.
D. Erdem Okumus and F. Terzi Sustainable Cities and Society 73 (2021) 103128

Meadows, Meadows, Randers & Behrens III, 1972; Yin et al., 2018). multicollinearity. In this paper, we used RRM to quantify the effects of
While LST reaches its highest level in built-up areas, it is much lower in the urban fabric on SUHI in Istanbul and identify the importance of the
vegetated lands such as urban parks and forests and other recreational contributing variables. We also applied the Monte Carlo simulation
areas (Yin et al., 2018). According to Li et al. (2016), green coverage of technique to test the accuracy, robustness, and sensitivity of the esti­
over 30% within an urban area is associated with significant decreases in mated model and coefficients. Our research pioneers the use of RRM for
UHI effect. The normalized difference vegetation index, a negative index a clear understanding of the relationship between the urban fabric and
designed to aid in the understanding of surface temperature anomalies, SUHI on the neighbourhood scale.
is used together with LST to measure the effects of urbanisation on RRM estimates unique coefficients that differ from the least-squares
surface temperature; generally, high index numbers result in low LSTs (Witten, Hastie & Tibshirani, 2013), by identifying certain patterns in
(Gallo & Owen, 1999; Lo, Quattrochi & Luvall, 1997; Rao, 1972; Weng, the training dataset, and determining the highly correlated variables.
2009; Yuan & Bauer, 2007). The model is used to enhance the least-squares estimator and improve
Research in this field has quantified the relationships between the the accuracy of the estimation model through the application of bias and
aforementioned urban features and the UHI effect through correlation the reduction of its variance (Hoerl & Kennard, 1970; Lan & Zhan, 2017;
analysis, OLS models, and spatial regression approaches (Buyantuyev & Ma & Cheng, 2016). As a linear transformation method, it has a smaller
Wu, 2010; Chun & Guhathakurta, 2017; Giridharan et al., 2004, 2005, mean square error, i.e. the estimated coefficients of ridge regression are
2007; Weng, Lu & Schubring, 2004; Yin et al., 2018). One study closer to real values (Ebrahimi Khusfi, Roustaei, Ebrahimi Khusfi &
comparing different urban fabrics demonstrated that urban design var­ Naghavi, 2020; Fan et al., 2017; Thompson, Kim, Aloe & Becker, 2017).
iables could explain the UHI intensity between 45% and 96% (Gir­ This study aims to discover the main contributors in the urban fabric
idharan et al., 2004), while another indicated that the reflectivity of to SUHI formation in Istanbul. It focuses on (1) investigating the effects
surface materials, building density, and sky-view factor have a major of Istanbul’s urban design metrics on the formation of SUHI, (2) iden­
impact on UHI around 80% (Giridharan et al., 2005). However, only a tifying the high-importance variables in order to provide a background
few of these studies discussed the validity, robustness, and sensitivity of for mitigation research and practice in Istanbul and (3) discussing
their models, which generally suffer from multicollinearity (Chun & whether the outcomes of the research can open up new possibilities for
Guhathakurta, 2017; Fan, Rey & Myint, 2017; Giridharan et al., 2007; SUHI mitigation through modification of design strategies.
Lan & Zhan, 2017; Myint, Brazel, Okin & Buyantuyev, 2010).
The use of advanced statistical approaches such as supervised ma­ 2. Motivational context for Istanbul
chine learning (SML) algorithms has also been attempted to quantify the
role of spatial structures and urban three-dimensional features on SUHI. Istanbul is one of many high-density megacities facing urban
SML algorithms are quite precise in the identification of interaction warming. Since the 1950s, the city’s spatial development has been
patterns and the indication of highly correlated and determinant vari­ affected by informal housing, uncontrolled urbanisation, and mass
ables. Few studies employ different SML techniques to define urban housing projects (Bolen, 2004; Keles, 1993; Terzi & Bolen, 2012). In the
contributors to the UHI effect (Rhee, Park & Lu, 2014; Sun, Gao, Li, course of the city’s urbanisation process, a sprawling pattern that started
Wang & Liu, 2019; Yoo, 2018). Yoo (2018) conducted macroscale with squatter areas and unplanned areas turned into a high-density
research on planning zones to understand physical and socioeconomic urban pattern, especially after the 1980s (Terzi & Bolen, 2009). In the
vulnerability to UHIs. Rhee et al. (2014) examined the relationship be­ urban densification process, high-density areas, especially the city cen­
tween land cover patterns and surface temperatures based on edge, tres, have become critical areas in terms of the UHI effect.
shape, and aggregation metrics. Sun et al. (2019) investigated the effects Density has continued to increase in the last 20 years with large-scale
of urban form on LST, but their study was limited to high-density urban housing and transportation projects, and urban sprawl towards the pe­
centres and the scale of urban blocks. ripheries has accelerated (Bolen et al., 2007; Terzi & Bolen, 2009, 2012).
Moreover, there is no single choice in SML approaches to clarify the The bidirectional spatial development process—low-density sprawl in
relationship between UHIs and urban features; both conventional peripheries and high-density housing development through regenera­
regression models and the more advanced statistical models that have tion activities in the city centre—has resulted in substantial impacts on
been attempted are limited by issues of multicollinearity. However, the the local microclimate and caused an upward trend in urban warming
Ridge Regression Model (RRM), one of the regularisation methods in and UHI effects (Bektas Balcik, 2014; Ezber, Sen, Kindap & Karaca,
SML techniques, is equipped to overcome the problem of 2007; Kaya, Basar, Karaca & Seker, 2012).

Fig. 1. Location of Istanbul in Turkey (a) and the Marmara Region (b). Istanbul’s land-use map shown in (c) was produced from CORINE CLC 2018 open data source.

2
D. Erdem Okumus and F. Terzi Sustainable Cities and Society 73 (2021) 103128

Fig. 2. Location and distribution of sample grids in Istanbul. (a) shows 1949 grid cells that are out of the dominant effects of meteorological and topographical
factors, (b) shows the 60 grid cells used as a sample set in the research.

Istanbul’s multi-fragmented and uncontrolled urbanisation has feedback mechanism and provides multiple benefits to the city as a
formed a heterogeneous spatial structure combining various types of the roadmap for the re-design of urban geometry in terms of SUHI mitiga­
urban fabric. This structure has become a composition of morphologi­ tion. We anticipate that through this research both academic and insti­
cally different sub-regions that incorporates various urban forms, geo­ tutional capacity will be developed to design and build climate-sensitive
metric patterns, and building typologies. A few examples of this urban forms.
heterogeneity include the dense and organic pattern in the city’s his­
torical centre, informal housing in the peripheries, high-rise and high- 3. Materials and methods
density urban blocks in the city centre, mass housing developments,
and single-family housing areas. This spatial structure has also The research in this paper comprises five phases: (1) grid-based
contributed to ever higher urban temperatures; unfortunately, the city sampling design, (2) mapping land surface temperature, (3) analysing
has reached a point where UHI effects have begun to impact human the urban fabric metrics of sample fingerprints, (4) running the Ridge
health and quality of life. Regression Model, and (5) applying Monte Carlo Simulation for
The urban planning and design profession has failed to comprehend sensitivity.
in particular the effects of urban geometry on temperature variations in
Istanbul. Consequently, it has fallen short in generating an adequate 3.1. Study area and sample selection procedure
response and redesigning the principles that direct urban planning and
design practices, especially in urban regeneration projects, the renewal Istanbul (41.0◦ N, 28.9◦ E), located in north-western Turkey, is the
of existing building stock, and the development of mass housing pro­ nation’s largest city, with a population of approximately 16 million
jects. In such projects, policymakers, urban planners, urban designers, people (TSI, 2018) (Fig. 1). The Bosphorus strait divides the city’s
and architects must understand the importance of the synergy between Anatolian and the European sides. The city includes diverse climato­
urban climate and urban design and establish design priorities through logical and topographical sub-regions due to its location in the transition
space-based knowledge. This paper, therefore, suggests the integration zone between the Marmara and the Black Seas. Its rough topography and
of heat island impact assessment to planning and design processes as a natural ventilation potential also affect its climate.

3
D. Erdem Okumus and F. Terzi Sustainable Cities and Society 73 (2021) 103128

The study was conducted through the grid-based sampling method to Table 1
include various forms of the urban fabric. Istanbul was first divided into Mathematical definitions of urban fabric metrics.
500 × 500 m sample grid cells where one cell approximately corre­ Urban Metrics Formulas Explanations
sponds to a neighbourhood, which is the lowest planning unit (21.571
( )
Building ∑ Abi Where Ab[i] is the area of i’th
BCR = ni=1
grids in total). As this study aims to reveal the effects of building and Coverage Ap building polygon. Ap is the plot area
Ratio (grid cell in this study). n is the
green coverage on UHIs, all determinants apart from these factors had to
number of building polygons.
be eliminated. The process of excluding the dominant effects of meteo­ Surface/ ∑n
(
Sbi
)
Where Sb[i] is the total surface area
rological and topographical factors involved 4 stages, resulting in the SVR = ( )/n
Volume Ratio i=1
Vbi of i’th building mesh. Vb[i] is the
reduction of the number of sample grids to 1949—1265 located on the volume of i’th building mesh. n is the
European and 684 on the Anatolian side of the city. The exact stages are number of building meshes.
∑
Sky-view SVF = ( ni=1 (1 − Where β[i] is the angle between the
as follows:
Factor cos2βi (αi /360∘ ))/k canyon floor and the sky vault
radius. α[i] is the radius of the
- Stage1: Removing sample grids containing water masses and no hemisphere of the sky (22.5◦ ). k is
buildings. the number of observation points (
Oke, 1987).
- Stage2: Removing grids located less than 1 km from a large water ( )
Canyon Where h[i] is the height of the i’th
mass. CGF = (
∑n hi
)/k
Geometry i=1
di building mesh. d[i] is the continuous
- Stage3: Removing grids located within forests or valleys. Factor distance from the i’th building
- Stage4: Removing grids on the Bosphorus slopes and those with a polygon to the building polygons. k
slope coverage of more than 25%. is the number of observation points.
Vegetation NIR − RED Where NIR is band 5 and RED is band
NDVI =
Index NIR + RED 4 in the Landsat-8 OLI/TIRS image
Each horizontal axis on the same line of latitude identifies a separate
sample set. The grids shown in red in Fig. 2 constitute the sample set
used in the research, based on their diverse morphological characteris­ measurements due to the small number of monitoring stations in
tics and building typologies. This sample dataset represents the urban Istanbul. LST anomaly (LSTa) values were therefore used for validation
fabric of each topological sub-region in the city. Out of 60 sample grids, during the statistical analyses.
30 are located on the European and 30 on the Anatolian side (Fig. A.1).
The grids, which extend on each side between the Bosphorus and the
inner-city, are coded A1-A30 on the Anatolian side and E1-E30 on the 3.3. Urban fabric metrics
European side (Fig. 2).
These grids are referred to in this paper as the urban fingerprints of Urban geometry, building density, and vegetation patterns, the key
Istanbul (Fig. A.1). We believe that every city has its own fingerprints components of the urban fabric, have certain impacts on LST variations
that set it apart from others and will therefore demonstrate different and SUHI intensity. Spatial configurations and geometrical patterns of
results in the analysis of SUHI contributing factors. Throughout this the buildings control the solar radiation access to urban areas, the
paper, the terms sample grids and fingerprints will be used release capacity of the heat stored and ventilation conditions for heat
interchangeably. loss. Green coverage decreases the SUHI intensity through significant
cooling services of shade effect and evapotranspiration.
3.2. Land surface temperature Buildings tend to store the heat generated by the absorption of
shortwave radiation during the day and release the heat via long-wave
SUHI measurements were obtained from LST map produced from the radiation after sunset (Oke, 1981, 1982). The difference between
remotely-sensed thermal data of Istanbul collected by NASA. It is well- absorbed and emitted heat forms the residual heat that contributes to the
known that SUHI effects peak in the summer, when temperatures and SUHI intensity (Oke, 1981). The amount of the residual heat based on
solar radiation are the highest in climatic regions similar to Turkey the energy exchange between urban surfaces varies on the design vari­
(Arnfield, 2003; EPA, 2008; Jain & Sarkar, 2017; Oke, 1982; Roth, Oke ables linked to the density and continuity of building masses, building
& Emery, 1989). Since Turkey’s highest annual temperature in 2017 was footprints, heights, volumes, distances between buildings, and quality of
recorded in July (TSMS, 2018), Landsat-8 OLI (operational land open spaces (Oke, 1981, 1988; Todhunter, 1990; Voogt & Oke, 1997).
imager)/TIRS (thermal infrared sensor) multispectral images from July Particularly, in high-rise and compact urban development patterns, the
25, 2017 with 0% cloud cover were used for LST calculation and map­ emission of the longwave radiation is restricted by the spatial configu­
ping. The satellite recorded images of Istanbul for 1 min between rations of buildings and residual heat restrained in the urban area gen­
08:43–08:45. Even though Landsat-8 OLI/TIRS satellite images contain erates warmer urban environments with a higher SUHI effect. Therefore,
two thermal bands (Band 10–11), Band-10 (10,60–11,19 µ) was used as we established our model with five urban fabric metrics including the
a single spectral band (Guha et al., 2018) in this study due to concerns design variables abovementioned in order to both explore the major
about the calibration uncertainty of Band-11 (11,50–12,51 µ) for contributors of the urban fabric on SUHI and develop an effective
quantitative analyses and the calculation of LST values (USGS, 2019). approach towards cooling Istanbul based on the re-design of the
In order to make a high-quality LST estimation based on thermal neighbourhoods: Building Coverage Ratio (BCR), Surface/Volume Ratio
images, the following steps were taken to process images through (SVR), Sky-view Factor (SVF), Canyon Geometry Factor (CGF), and
geographical information technologies: (1) Conversion of the digital Normalized Difference Vegetation Index (NDVI).
numbers of pixels to Top of Atmosphere reflectance using the calibration The geometric indicators were also used as a proxy for vertical and
data of Band 10 – thermal band (Barsi et al., 2014). (2) Transformation horizontal building density, which plays an important role in the plan­
of the band data to the brightness temperature of the atmosphere from ning system and dominates the re-design of the neighbourhoods’ two
spectral brightness using the thermal constants in the metadata file and three-dimensional configurations, while the vegetation index rep­
(USGS, 2013; Xu & Chen, 2004). (3) Calculation of the Normalized resented the density and the quality of green coverage in the
Difference Vegetation Index, Proportion of Vegetation, and Land Surface neighbourhoods.
Emissivity (Jiménez-Muñoz et al., 2009; Jiménez-Muñoz, Sobrino, Gil­ Urban fabric metrics were measured and modelled using a three-
lespie, Sabol & Gustafson, 2006; Sobrino, Jiménez-Muñoz & Paolini, dimensional digital building database (composed of vector data gath­
2004; Weng et al., 2004). (4) Calculation and validation of LST. The ered in 2017 by the Istanbul Metropolitan Municipality) and Landsat-8
ground surface temperature values could not be confirmed by in-situ raster images processed through geographical information systems

4
D. Erdem Okumus and F. Terzi Sustainable Cities and Society 73 (2021) 103128

Fig. 3. Plots of ridge trace (a) and variance inflation factors (b) showing the appropriate k-ridge parameter value which controls the biasness in RRM to solve the
multicollinearity. The smallest k-value at 0.1486 indicates the smallest bias in the estimation model and after which the coefficients remain constant.

(Table 1). transformation (NDVIlog) for NDVI to reduce data variability and make
The Building Coverage Ratio, the horizontal building density, ranges the data more consistent with the normal distribution.
between 0 and 1. A BCR value of 1 indicates complete building coverage
of the area of a parcel. For the purposes of this study, BCR values were 3.5. Estimation model: Ridge regression and L2 regularisation
obtained by calculating the proportion of the total base area of the
buildings in the sample grid to the grid area. The Surface/Volume Ratio The Ridge Regression Model, a regularized regression technique, was
of the buildings determine the energy performance of an area. SVR in adopted to quantify the effects of the urban fabric on SUHI. This pre­
this study was calculated using the ratio of the total surface area of each diction model used the LST anomaly value as the dependant variable and
building to its volume, excluding the adjacent facades. BCR, SVR, SVF, CGF, and NDVIlog as explanatory variables. The hy­
The Sky-view Factor indicates the visible part of the sky from a pothesis suggests that these explanatory variables have a major impact
certain vantage point, quantifying the openness of the sample grid. Its on the spatial variability of the LSTa in the sample grids.
values range from 0 to 1. To calculate SVF, observation points were LSTa = β0 + β1*BCR + β2*SVR + β3*SVF + β4*CGF + β5*NDVIlog + ε
established at 15 m intervals on the ground level, and the target point (where β0 is a constant value and ε is an error term of the regression
cloud was generated at intervals of 22.5◦ for each observation point. model)
Target point cloud formed a sky hemisphere to calculate the sky visi­ Also known as Tikhonov regularisation, RRM is a biased estimation
bility. Then, sightlines were constructed from each observation point method and parametric machine-learning algorithm. The model avoids
(origin) to target points in its specific target point cloud. After the data overfitting, solves multicollinearity problems, and provides better
calculation of each observation point’s sky-view factor based on the long term predictions through regularisation, i.e. introducing a small
ratio of the number of sightlines that coincided with buildings to the amount of bias (with the ridge parameter) to the model, which results in
total number of sightlines, the spatial interpolation method was a significant drop in variance. The ridge parameter (k-parameter ranges
employed to generalize the factor values to the grid area (Okumus, between 0 and 1), which controls the biasness authorized in the estimate
2017). (if k = 0; then the estimation is unbiased), was calculated using the ridge
The Canyon Geometry Factor or canyon depth of an urban area is trace method and variance inflation factors (VIF) graph. Ridge trace is a
determined by the height of and distance between buildings in the graphical method used to determine the variability and the edge of
streets, which affect the amount of solar radiation that reaches urban stability of β regression coefficients. The plot of the ridge trace indicates
surfaces. It was measured at observation points placed between build­ the smallest value for the k-parameter in which the ridge regression
ings at 15 m intervals. The specific target point cloud of each observa­ coefficients as a function of the k-parameter remain consistent. More­
tion point was created at 12◦ intervals, and only the target points with a over, the plot of the variance inflation factors indicates the point at
height value of zero were selected to calculate the distance between which multicollinearity problems can be avoided. The value (k =
buildings. The sightlines representing the distance between buildings 0.1486), where the variance inflation factor approached 1, was chosen
were constructed by combining observation and target points at ground as the k-ridge parameter (Fig. 3).
level. The canyon depth value was acquired by measuring the propor­ The regression model quantitatively shows the main components of
tion of the height of the buildings to the length of the sightline. The the urban fabric (urban form, geometry, and vegetation cover) that play
spatial interpolation method was applied to generalize the factor values an important role in the formation of SUHI effects in Istanbul. It provides
to the grid area. A canyon geometry factor of plus/minus 1 represents a the local coefficient values (β) of the variables that will form the basis of
uniform street canyon; values below 0,5 represent a shallow street UHI studies in Istanbul. This outcome also suggests that physical in­
canyon, and a CGF equal to or higher than 2 characterizes a deep street terventions in urban form have great potential for mitigation efforts.
canyon (Shishegar, 2013).
Normalized Difference Vegetation Index helps to assess the density of
3.6. Model accuracy
live green vegetation in the sample grids. The NDVI value was obtained
through Landsat-8 OLI/TIRS images recorded on 25th July 2017 using
The performance of the ridge regression estimation was evaluated
image processing techniques in geographical information systems (Yuan
through the calculation of various accuracy metrics, including Adjusted
& Bauer, 2007). NDVI map values range from 1 to − 1. Positive values
R-squared (R2), the Durbin-Watson statistic, Standard Error of the esti­
refer to a particular amount of vegetation cover.
mate, and Mean Absolute Error (MAE). These metrics indicate the ac­
curacy, validity, and quality of the estimator. Adjusted R-squared shows
3.4. Normality tests the data points fitting the line of the regression function and indicates
what percentage of variation can be explained by the independent
Statistical and graphical tests of normality evaluate the assumption variables having a high effect on the dependant variable. A higher value
of a null hypothesis and reveal whether the dataset is close to the normal of adjusted R-squared points to a model better at estimating LSTa. The
distribution. We used the Shapiro Wilk test for independent variables Durbin-Watson statistic examines the residuals to identify any serial
and examined skewness-kurtosis scores statistically, and assessed his­ correlation; the P-value is expected to be greater than 0.05 to indicate
tograms and Q-Q plots graphically. We applied logarithmic there is no autocorrelation in the residuals. The lower Standard Error of

5
D. Erdem Okumus and F. Terzi Sustainable Cities and Society 73 (2021) 103128

Fig. 4. LST distribution (a) and map of the SUHI effect (b) on 25th July 2017 in Istanbul. LST anomalies above the average (34.73 ◦ C) indicate the SUHI effect at
various levels.

the estimate, which explains uncertainty in the regression model by 3.7. Sensitivity analysis: Monte Carlo simulation
measuring the distance between points from the regression line, results
in estimation points closer to the regression line and the better estima­ We used Monte Carlo simulation-based sensitivity analysis to esti­
tion of LSTa. MAE measures the error of the estimation and the differ­ mate the distribution of random variables and understand the model’s
ence between the observed and predicted data values of the regression stability and robustness and the sensitivity of the model output (LSTa) to
model. Lower MAE values indicate a more accurate and valid estimation variations in the independent variables. A series of design scenarios
model. (10.000, the number of design samples) of multiple quantitative com­
binations of the independent variables was generated randomly by the
simulation technique as a test dataset. For each variable in the simula­
tion, the measures of central tendency, variability, and shape were

6
D. Erdem Okumus and F. Terzi Sustainable Cities and Society 73 (2021) 103128

Fig. 5. SUHI profiles (a) and quantitative patterns of urban metrics (b), (c), (d), (e), (f) in urban fingerprints. Vertical axes demonstrate the measured values of
variables in each grid.

of the independent variables by evaluating the level of impact on SUHI

Table 2
formation of the five variables. The Monte Carlo simulation is common
Descriptive statistics of the variables for the 60 sample grids.
in machine learning because it provides the resampling for estimating
N Min. Max. Mean Std. Dev. the accuracy of a model on a limited dataset (Murphy, 2012).
LSTa 60 − 4.18 2.27 0.17 1.42
BCR 60 0.09 0.58 0.33 0.12
SVR 60 0.54 0.82 0.67 0.06 4. Results
SVF 60 0.36 0.97 0.67 0.14
CGF 60 0.85 2.26 1.43 0.31
4.1. SUHI characteristics and spatial patterns of urban fingerprints
NDVI* 60 0.07 0.28 0.14 0.04
NDVIlog 60 − 1.13 − 0.55 − 0.89 0.13
*
Istanbul follows a linear spatial development pattern spread over the
NDVI was not used in the statistical analysis.
east-west axis and covering the south part of the city. The distribution of
LST on 25th July 2017 was directly linked to the level and direction of
analysed. Standardized skewness and standardized kurtosis were eval­ urbanisation in Istanbul. The observed surface temperatures were higher
uated to assess whether the test dataset originated from a normal dis­ in the south than in the north, where natural elements (i.e. water basins,
tribution. The Sensitivity Tornado Plot also tested the model coefficients forests, and agricultural areas) dominate (Fig. 4). However, there was a

Fig. 6. The scatter plot matrix visualizing bivariate relationships between pairs of variables in the model.

7
D. Erdem Okumus and F. Terzi Sustainable Cities and Society 73 (2021) 103128

Table 3
Normality Test results of independent variables used in the model (p-values>0.05 indicate the variables close to normal distribution).
BCR SVR SVF CGF NDVIlog
Stat. P-value Stat. P-value Stat. P-value Stat. P-value Stat. P-value
Shapiro-Wilk W 0.99 0.69 0.98 0.60 0.98 0.57 0.98 0.60 0.98 0.38
Skewness Zscore 0.07 0.94 0.72 0.47 0.50 0.62 0.58 0.56 0.92 0.36
Kurtosis Zscore − 0.76 0.45 − 0.42 0.67 − 0.92 0.36 − 0.41 0.68 − 0.52 0.60

tendency towards higher surface temperatures in the north as well. As NDVI had an apparently negative impact on SUHI (Fig. 5, Fig. 6). The
Istanbul becomes more populated, the city sprawls further to the north, Anatolian side (maxndvi(A04)= 0.28; minndvi(A24)= 0.08; meanndvi= 0.15)
and mega-projects such as the new transnational highway and Istanbul possesses more vegetation cover than the European (maxndvi(E29)= 0.21;
Airport have invaded the large open spaces on the coast of the Black Sea. minndvi(E19)= 0.07; meanndvi= 0.13). Even though the NDVI value
Istanbul’s surface temperatures ranged from 22.03 ◦ C to 47.76 ◦ C, reached up to 0.28 on the Anatolian side, the amount of green coverage
and the temperature difference between urban and rural areas reached was very low in most urban fingerprints (Table 2).
4.29 ◦ C (The average temperature was 33.70 ◦ C in urban and 29.41 ◦ C in There were other positive relationships between CGF-BCR and NDVI-
rural areas). Although the mean temperature value for the total area of SVF and negative relationships between BCR-SVF, BCR-NDVI, and SVF-
urban and rural zones was 30.40 ◦ C, the average temperature in the CGF pairs, regarding the matrix plot of the urban metrics (Fig. 6).
1949 sample grid was estimated at 34.73 ◦ C. The anomaly values of each
sample grid were thus calculated from an average temperature of
4.2. Model estimation and validation
34.73 ◦ C.
The SUHI effect peaked in the central areas, which are characterized
Normality tests indicated that the distribution of the dataset is close
by high building and population density, with surface temperatures
to the normal distribution for all independent variables. Since the
reaching almost 5 ◦ C above the average. Approximately 25% of the total
smallest P-value amongst the tests performed is greater than 0.05, the
city area displayed SUHI effects, with 21% of the Anatolian side and
study can assert that the variables come from a normal distribution, with
26% of the European side showing SUHI influence. Areas of dense ur­
95% confidence (Table 3).
banization near the wetlands located in the south of the city also pre­
The Ridge Regression Model demonstrates that the urban metrics
sented high surface temperatures. Even though wetlands within cities
were responsible for 71% (R2= 0.736484, Adj. R2= 0.712085) of the
have the capacity to balance the urban temperatures, the dense urban
LSTa. Although R2 statistic explains 74% of the variability in LSTa, the
fabric in these areas has reduced the chance of a balancing effect.
adjusted R2 statistic, which is more acceptable for evaluating models
However, despite the high building density in the southern Bosphorus
with different values in their independent variables, was calculated at
region, the coastal temperatures were lower because of the natural
71%. This estimation indicates that the urban fabric has a significant
ventilation potential of the strait (Fig. 4).
impact on the formation of SUHI effects in Istanbul. In this case, the
Descriptive statistics for the 60 sample grids show that LSTa reached
fitted regression model is as follows:
2.27 ◦ C above the average on 25th July in 2017, indicating a serious
LSTa = − 6.74137 + 2.32493*BCR + 1.86304*SVR - 0.84395*SVF -
SUHI effect in summer (Table 2). SUHI profiles of the sample urban
0.543001*CGF - 7.05941*NDVILog10 + ε
fingerprints demonstrate that although the average anomaly is higher on
Durbin-Watson statistic (Pvalue= 0.0673, greater than 0.05) did not
the European side (maxa(E19)= 1.99 ◦ C; mina(E26)=− 1.49 ◦ C; meana=
show any serial autocorrelation in the residuals at a 95% confidence
0.42 ◦ C) than on the Anatolian side (maxa(A30)= 2.27 ◦ C; mina
interval. The standard error of the estimate explaining the uncertainty in
(A07)=− 4.18 C; meana=− 0.08 C), the Anatolian side has much higher-
◦ ◦
the regression model indicated a standard deviation of 0.645 for the
degree anomalies in certain grids (Fig. 5).
residuals. MAE, the average value of the residuals, was calculated as
The maximum building coverage ratio was 0.58, i.e. there are fin­
0.413.
gerprints where the total building floor area covers more than half of the
grid area amongst the sample grids (Table 2). Considering the graphical
view of the quantitative spatial patterns in Fig. 5 and Fig. 6, LST 4.3. Effect of urban fabric on SUHI in Istanbul
anomalies seem to be positively correlated with BCR values. The average
values of BCR were similar both on the European (maxbcr(E12)= 0.58; The impact level of controllable urban design factors on the neigh­
minbcr(E29)= 0.09; meanbcr= 0.34) and Anatolian sides (maxbcr(A18)= bourhood scale was found to be around 71% in Istanbul in the sum­
0.48; minbcr(A07)= 0.09; meanbcr= 0.32). mertime period. According to the coefficient values of the variables, the
Moreover, the quantitative spatial patterns of the sample grids imply most important contributor of the model was NDVI (β= − 7.05941)
a negative correlation between SVF-LSTa and CGF-LSTa pairs (Fig. 5, (which was entered into the model as log-transformed), which corre­
Fig. 6). Sky-view factor extends to 0.97 at its highest level, revealing the sponds to the vegetation cover in a neighbourhood. Although the
presence of a sample grid that has nearly full open sky and a very low observed values of vegetation cover in the highly urbanized sample grids
building density on the European side (Table 2). Both the European side were very low, it emerged as the variable with the greatest amount of
(maxsvf(E29)= 0.97; minsvf(E13)= 0.36; meansvf= 0.71) (maxcgf(E27)= influence on LST variation and SUHI formation. Apart from its strong
2.26; mincgf(E29)= 0.85; meancgf= 1.53) and Anatolian side (maxsvf correlation with LST, the coefficient of NDVIlog in the model was
(A04)= 0.93; minsvf(A18)= 0.51; meansvf= 0.69) (maxcgf(A16)= 1.73; mincgf considerably negative in summertime (Guha et al., 2018; Macarof &
(A07)= 0.88; meancgf= 1.33) of the city possess approximately similar Statescu, 2017; Peng, Jia, Liu, Li & Wu, 2018). Since a high NDVI creates
average values of SVF and CGF. The highest value for canyon geometry a cooling effect in urban spaces, LST is expected to decrease as NDVI
factor indicates that the high-density urban fabric on the European side increases; the RRM model confirms this assumption. As far as can be
can be characterized in many cases as deep street canyons (Fig. 5). inferred from satellite images, this cooling effect is mainly attributable
to the shadow effect of vegetation structures. Ground-level green spaces

8
D. Erdem Okumus and F. Terzi Sustainable Cities and Society 73 (2021) 103128

Table 4 prevent urban cooling by changing wind profiles, had a significant effect
Descriptive statistics of the simulated test dataset (10,000 samples) by Monte on UHI formation (G. Guo et al., 2016; X. Yang & Li, 2015).
Carlo Simulation. Values of standardized skewness and standardized kurtosis Two other urban metrics related to building geometry, SVF (β=
outside the range of − 2 to +2 indicate significant departures from normality. − 0.84395) and CGF (β= − 0.543001), had relatively less impact on
N Average Std. Dev. Stnd. Skewness Stnd. Kurtosis LSTa. SVF and CG are components related to building height and vertical
LSTa 10,000 0.16 1.03 0.09 − 0.69
BCR 10,000 0.33 1.12 − 0.57 − 0.07
SVR 10,000 0.67 0.06 1.58 − 0.45
SVF 10,000 0.70 0.14 − 0.006 0.10
CGF 10,000 1.43 0.31 − 0.46 1.64
NDVIlog 10,000 − 0.88 0.13 0.65 − 0.97

also contribute to the cooling effect by creating permeable surfaces,

increasing water storage capacity and the amount of evapotranspiration
in the urban space (Santamouris, Ding & Osmond, 2019).
The second most influential variable was BCR (β= 2.32493), used to
understand the importance of horizontal building density for SUHI. The
results of the model point to building density as a highly determinant
factor of LST variations amongst the five independent variables, which is
Fig. 9. Sensitivity Tornado Plot displaying the importance of the independent
consistent with the literature (Yin et al., 2018). The coefficient of BCR
variables in the model. The vegetation index is at the top as the highest
showed a positive correlation with LST, which tended to increase in
contributor to the SUHI in Istanbul.
urban fabrics with high building coverage. SVR (β= 1.86304), which
refers directly to building geometry and is mostly related to energy
consumption in a neighbourhood, was the third-ranked effective vari­
able in the model results. Its coefficient value was also positive, similar
to BCR. Indeed, large-volume buildings, which reduce wind speed and

Fig. 7. Histograms of the simulated dataset regarding the variables in the model. Graphs indicate the data is normally distributed.

Fig. 8. Box and Whisker Plots of the simulated dataset by Monte Carlo Simulations showing the interquartile range and outliers of the data associated with
the variables.

9
D. Erdem Okumus and F. Terzi Sustainable Cities and Society 73 (2021) 103128

building density, and the correlation between LST and the vertical urban design parameters to UHI formation is mostly above 50% (Gir­
density was lower than that between LST and horizontal density (G. Guo idharan et al., 2004, 2005; Sun et al., 2019; Yoo, 2018). These results
et al., 2016). Indeed, coefficients for SVF and CGF demonstrated a strongly support the concept of SUHI mitigation capacity through
negative impact on LST; apparently, an increase in vertical density on a modifications to and re-design of the urban fabric.
neighbourhood scale might result in a decrease in LST. The local characteristics of a city and the time of the year when the
surface temperature was measured should also be considered when
4.4. Model sensitivity and simulation discussing the mitigation potential of urban planning and design. This
study is based on the LST values produced from the thermal images
Model sensitivity, which was tested through the Monte Carlo simu­ saved in the summer period (the warmest time of the year). Under these
lation, indicated that the distribution of the simulated dataset (test data) circumstances, the components of Istanbul’s urban fabric that contrib­
overlaps with the training dataset gathered from the 60 sample grids. uted to SUHI formation most are NDVI and BCR, while the impacts of
The standardized skewness and standardized kurtosis values demon­ SVF and CGF remained relatively low.
strated that the test data follows a normal distribution pattern (Table 4). Vegetation coverage, which is negatively correlated with SUHI var­
The normality of a simulated test dataset is a key component in evalu­ iations, has great potential for the mitigation of surface temperatures. As
ating the robustness of a parametric estimation model. is argued in the literature, increasing urban vegetation density may
Moreover, the histogram and box plots showed the spread of the decrease urban temperatures by around 0.2–4.7 ◦ C through the increase
simulated test dataset with the range, mean, median and average values. in shadow surfaces and evapotranspiration in the urban environment
While the box length pointed to the variability of the dataset, there were (Kikegawa, Genchi, Kondo & Hanaki, 2006; Kleerekoper, Van Esch &
a few outliers in the simulated dataset for each variable (Fig. 7, Fig. 8). Salcedo, 2012; McPherson, 1994; Rizwan, Dennis & Chunho, 2008; J.
Considering, in particular, the averages of the training dataset (gathered Yang et al., 2021). Based on the local coefficient of NDVI in Istanbul,
from 60 sample grids) and the test dataset, the values of the variables even an only 0.1-unit1 increase in the density of vegetative surfaces in a
were calculated as almost equal, which indicates that the simulated test neighbourhood would provide at least a 0.7 ◦ C reduction in surface
dataset behaved similarly to the training dataset. temperature anomalies. Most studies have indicated that the effects of
The tornado diagram also verified the RRM results regarding the NDVI are not dependant on temporal patterns and that the regression
level of significance of the independent variables at a 95% confidence function is not sensitive to seasonal changes (Chen, Li, Huang & Xu,
interval (Fig. 9). According to the diagram, NDVI was the highest 2013; Sun et al., 2019); however, Peng et al. (2018) argue that NDVI
contributor to the model, while SVF and CGF were the lowest. Obser­ could have a positive impact on LST in the wintertime of subtropical
vations demonstrate that LST anomalies were more sensitive to NDVI climate regions similar to Turkey (Peng et al., 2018).
and BCR than to other features related to the urban fabric, while SVF and In terms of geometrical pattern, indicators of building density at the
CGF were relatively insignificant, with the model appearing rather horizontal level (BCR) suggest a higher potential for SUHI reduction
insensitive to their variations. than do those of vertical density (SVF, CGF) on the neighbourhood scale.
Contrary to the common notion that SVF has a particularly significant
5. Discussion impact (Chun & Guldmann, 2014; Hu et al., 2016) on SUHI formation, it
seems to have had a relatively smaller influence in Istanbul.
Istanbul has experienced continuous urban regeneration and re- The local coefficients in Istanbul demonstrated that a 0.1-unit in­
development as a result of the rapid urbanisation triggered by the crease in BCR1 resulted in a 0.23-unit increase in LSTa. In other words, a
migration it has received since the 1950s. This ongoing process has been 0.1-unit decrease in horizontal density created a 0.23 ◦ C reduction in the
characterized until the recent past mainly by unplanned and uncon­ SUHI effect. On the other hand, SVF, a vertical density indicator nega­
trolled urban development; in the present day, it is mostly dominated by tively correlated with LST, provided for an approximately 0.1 ◦ C
very large mass housing projects. Regardless of its form, however, urban reduction in LST with a 0.1-unit increase in index value1. The findings
development has shaped the built environment. Due to the unplanned suggest that lower levels of sky visibility produce higher SUHI effects in
and uncontrolled development experienced in the past, factors such as urban environments, which is in line with the findings of Oke (1981).
density, building type, street width and form, all effective in shaping the However, some researchers argue that the SUHI effect tends to increase
built environment, were determined principally by economic and so­ with high SVF values due to the high exposure of the surfaces to solar
ciological concerns, without much (if any) regard for micro-climatic radiation (Arnfield, 1990; Giridharan et al., 2007; Strømann-Andersen &
factors. However, even as the struggle to cope with the negative ef­ Sattrup, 2011). Guo et al. (2016) assert that moderate SVF values
fects of climate change on cities continues on a macro-scale, the public generate the lowest temperatures, while the highest and the lowest SVF
health, social, economic, and environmental benefits of combating the values generate the highest surface temperatures.
effects of UHIs on a micro-scale are coming into view. This research was Previous research has tended to associate low building density with
conducted in order to contribute to such a requirement and has resulted high rise buildings (i.e. horizontally low, vertically high density),
in promising findings. arguing that the cooling effects of extensive vegetated surfaces and the
shadows of the buildings decrease LST (G. Guo et al., 2016; X. Yang & Li,
5.1. The SUHI mitigation potential of urban planning and design 2015). However, horizontally low-density urban areas could also have
an inverse effect if the open spaces are covered by non-reflective ma­
This study demonstrates that in Istanbul, the temperature difference terials (Stone Jr & Rodgers, 2001; Zhou, Huang & Cadenasso, 2011).
between urban and rural areas exceeded 4 ◦ C in July 2017. It suggests Urban planning professionals should therefore attempt to find a balance
that the urban fabric highly impacts the formation of SUHI, which makes in climate-sensitive neighbourhood design between the open space
it necessary to rethink the potential of urban planning and design, as a compositions and horizontal-vertical building density.
reorganisation of spatial configurations in Istanbul could result in a
~70% decrease in SUHI. 5.2. Twofold approach for cooling Istanbul
The impact of urban fabric components on over 70% on the forma­
tion of SUHI demonstrates the greater influence of controllable and We suggest a twofold SUHI mitigation approach for cooling Istanbul
modifiable urban design parameters than that of uncontrollable (i.e.
meteorological and topographic) features on the neighbourhood scale.
Indeed, studies conducted with different parameter combinations and 1
Unit increments are evaluated as 0.1 since the maximum index value equals
varying regression methods have also shown that the contribution of 1.

10
D. Erdem Okumus and F. Terzi Sustainable Cities and Society 73 (2021) 103128

and encouraging urban planning and design professionals to build Regression Model, helped to overcome the multicollinearity problem in
climate-sensitive neighbourhoods: minor improvements and structural the model by increasing its biasness and decreasing its variance. The
interventions. Minor improvements are interventions that can be easily model thus enabled the researchers to recognize the patterns in the
integrated into neighbourhoods without changing the main structural training dataset and make more accurate predictions. The conclusions
and land-use decisions and basic design principles. The high NDVI effect and implications are as follows:
in Istanbul is the clearest indicator that a significant SUHI reduction can
be achieved by increasing the amount of green space and vegetative (1) Urban fabric metrics are highly correlated with LST patterns,
surfaces in the urban fabric. Green elements such as trees, bushes, bio­ potentially accounting for more than 70% of SUHI formation.
swales, surface greening, and permeable pavement are highly effective NDVI and BCR in particular were the main contributors to LST
in creating cool places in public spaces (Rosenzweig, Solecki & Slosberg, variations out of the five metrics of the urban fabric. While NDVI
2006). Moreover, the encouragement of green and permeable surfaces had a strong negative correlation with LSTa, BCR had a positive
on private lands, roofs, and facades may strengthen cooling service, impact on it. Urban planners and designers should consider
especially in high-density urban areas (Kleerekoper et al., 2012). The increasing green spaces and vegetation coverage in urban areas to
mitigation of surface temperatures may reach 3–4 ◦ C through an in­ decrease the SUHI effect. In a situation where providing or
crease of up to 50% in the amount of green coverage in Istanbul’s enlarging green areas may be found unfeasible due to the scarcity
neighbourhoods by such kind of relatively low-cost minor imple­ of available land in high-density urban areas, it is still possible to
mentations. The presence of larger green surfaces and greater tree improve the landscape patterns of existing green spaces. More­
coverage provides a cooling effect through increased shaded spaces, over, the use of green roofs or more vegetative surfaces on streets
greater ground storage of rainfall, and higher evapotranspiration might be practically implemented as minor improvements. In
capacity. addition, urban planning and design professionals should
Structural interventions, despite the requirement of a great deal of consider decreasing urban densities based on building coverage
effort and funding, might be implemented during the projects’ planning and controlling the spatial configurations of the urban fabric
and design stage to achieve further reduction of SUHI in Istanbul. through structural interventions. Especially in high-density urban
Through a decrease of 40% in BCR, it is possible to achieve a 1 ◦ C centres, reducing density might also be unfeasible due to the
mitigation in temperatures and mitigation level can be raised to 1.5 ◦ C ownership rights and economic concerns; however, Istanbul’s
by a 60% reduction in building coverage. However, in doing so, it is recent urban transformation potential provides a unique oppor­
important to ensure a balance between the parametric values of tunity to perform structural modifications in built-form to
building-based variables. If BCR reduction leads to a decrease in SVF and decrease the SUHI effect.
CGF, indirect secondary effects may result in a re-increase in SUHIs. (2) The model and outcomes can help assess the SUHI impact of an
Alternative design combinations can be produced practically by urban design and arrive at better solutions for SUHI mitigation.
running our regression equation and local coefficient values. For SUHI impact assessment is a powerful tool that aids in monitoring
instance, in a neighbourhood with a 2.5–3 ◦ C LSTa, the SUHI effect can the potential impacts of modifications to the existing urban fabric
be minimized through a 50% increase in vegetation coverage as a minor and overhauling the design strategies that influence the current
improvement, or through a 30% rise in vegetation and a 30% reduction urban form of Istanbul before they are implemented. It is a pro­
in BCR. The parametric values of the variables in the planning and cess based on estimation and prevention that should develop in
design projects will aid in the recognition of what level of SUHI reduc­ line with the decision-making phase. Accordingly, the model used
tion might be achieved in neighbourhoods in Istanbul. in this paper has established a roadmap for mitigation strategies
through the modification of variables by measuring the SUHI
effect of existing urban fabrics, creating room for redesign
5.3 Technical limitation and future studies
through the evaluation of the SUHI impacts of design principles
before implementation.
The main limitation of this study was the effects of anisotropy, or the
(3) This study encourages designers, planners and urban pro­
directional variations in LST measurements by satellite images due to
fessionals to consider the impacts of SUHI before making de­
the position and the reflection angle of the sun, the view direction of the
cisions. Integrating SUHI impact assessment into planning and
sensor, and the shaded and invisible parts of the three-dimensional
design processes will pave the way for better urban design al­
urban space (Voogt & Oke, 1997, 1998). This limitation was unavoid­
ternatives for SUHI mitigation. As feedback and decision-support
able because of the lack of alternative imagery on the same date for
mechanism, it will help planning and design authorities to un­
Istanbul.
derstand the effects of design decisions on local micro-climates.
For future research, analysis of the relationship between the urban
fabric and LST variations in different grid resolutions with machine
Declaration of Competing Interest
learning algorithms could lend a greater understanding of scale effects
for different land uses and planning zones. We also would like to
The authors declare that they have no known competing financial
emphasize that anisotropy effects should be considered in future studies.
interests or personal relationships that could have appeared to influence
the work reported in this paper.
6. Conclusion
Acknowledgements
This study evaluates the impact of urban fabric on the SUHI effect
and provides a clear understanding of the effect’s main contributors so This paper was produced from the PhD dissertation conducted by the
as to enable practitioners to make more appropriate decisions in plan­ corresponding author under the supervision of the second author. This
ning and design implementations in Istanbul. Its comprehensive meth­ work was supported by Scientific Research Projects Department of
odology involved processing thermal images recorded by remote sensing Istanbul Technical University, Istanbul, Turkey [grant number 42088].
technologies and modelling the urban fabric metrics through
geographical information systems and statistical analyses based on
machine learning techniques. One specific SML approach, the Ridge

11
D. Erdem Okumus and F. Terzi Sustainable Cities and Society 73 (2021) 103128

Appendices

Fig. A.1.

Fig. A.1. Urban fingerprints of 60 sample grid cells in Istanbul.

12
D. Erdem Okumus and F. Terzi Sustainable Cities and Society 73 (2021) 103128

References Kaya, S., Basar, U. G., Karaca, M., & Seker, D. Z. (2012). Assessment of urban heat islands
using remotely sensed data. Ekoloji, 21(84), 107–113.
Keles, R. (1993). Kentleşme politikası. İmge Kitabevi.
Arnfield, A. J. (1990). Street design and urban canyon solar access. Energy and Buildings,
Kikegawa, Y., Genchi, Y., Kondo, H., & Hanaki, K. (2006). Impacts of city-block-scale
14(2), 117–131.
countermeasures against urban heat-island phenomena upon a building’s energy-
Arnfield, A. J. (2003). Two decades of urban climate research: A review of turbulence,
consumption for air-conditioning. Applied Energy, 83(6), 649–668.
exchanges of energy and water, and the urban heat island. International Journal of
Kleerekoper, L., Van Esch, M., & Salcedo, T. B. (2012). How to make a city climate-proof,
Climatology, 23(1), 1–26.
addressing the urban heat island effect. Resources, Conservation and Recycling, 64,
Barsi, J., Schott, J., Hook, S., Raqueno, N., Markham, B., & Radocinski, R. (2014).
30–38.
Landsat-8 thermal infrared sensor (TIRS) vicarious radiometric calibration. Remote
Lan, Y., & Zhan, Q. (2017). How do urban buildings impact summer air temperature? The
Sensing, 6(11), 11607–11626.
effects of building configurations in space and time. Building and Environment, 125,
Bektas Balcik, F. (2014). Determining the impact of urban components on land surface
88–98.
temperature of Istanbul by using remote sensing indices. Environmental Monitoring
Li, S. T., Li, J., & Duan, P. (2016). Relationship between urban heat island effect and land
and Assessment, 186(2), 859–872. https://doi.org/10.1007/s10661-013-3427-5
use in Taiyuan City, China. In , 63. MATEC Web of Conferences (p. 4024). EDP
Bolen, F. (2004). Housing policy and housing systems in Turkey. A| Z ITU Journal of the
Sciences.
Faculty of Architecture, 1(1), 14–31.
Lo, C. P., Quattrochi, D. A., & Luvall, J. C. (1997). Application of high-resolution thermal
Bolen, F., Türkoğlu, H., Ergun, N., Yirmibeşoğlu, F., Terzi, F., Kaya, S., et al. (2007).
infrared remote sensing and GIS to assess the urban heat island effect. International
Quality of residential environment in a city facing unsustainable growth problems:
Journal of Remote Sensing, 18(2), 287–304.
Istanbul. In Joint Congress of the European Regional Science Association and the
Ma, J., & Cheng, J. C. P. (2016). Estimation of the building energy use intensity in the
Association de Science Régionale de Langue Française, Paris, August/September.
urban scale by integrating GIS and big data technology. Applied Energy, 183,
Boyko, C. T., & Cooper, R. (2011). Clarifying and re-conceptualising density. Progress in
182–192.
Planning, 76(1), 1–61.
Macarof, P., & Statescu, F. (2017). Comparasion of ndbi and ndvi as indicators of surface
Buyantuyev, A., & Wu, J. (2010). Urban heat islands and landscape heterogeneity:
urban heat island effect in landsat 8 imagery: A case study of iasi. Present
Linking spatiotemporal variations in surface temperatures to land-cover and
Environment and Sustainable Development, 11(2), 141–150.
socioeconomic patterns. Landscape Ecology, 25(1), 17–33.
McPherson, E. G. (1994). Cooling urban heat islands with sustainable landscapes. In The
Chen, L., Li, M., Huang, F., & Xu, S. (2013). Relationships of LST to NDBI and NDVI in
ecological city: Preserving and restoring urban biodiversity, 151-171 pp. 151–171).
Wuhan City based on Landsat ETM+ image. In , 2. 2013 6th International Congress on
Amherst, MA: University of Massachusetts Press.
Image and Signal Processing (CISP) (pp. 840–845). IEEE.
Meadows, D. H., Meadows, D. L., Randers, J., & Behrens III, W. W. (1972). The limits to
Chokhachian, A., Perini, K., Giulini, S., & Auer, T. (2020). Urban performance and
growth. in a report for the club of rome’s project on the predicament of mankind. New
density: Generative study on interdependencies of urban form and environmental
York: Universe Book.
measures. Sustainable Cities and Society, 53, Article 101952.
Merlier, L., Kuznik, F., Rusaouën, G., & Salat, S. (2018). Derivation of generic typologies
Chun, B., & Guhathakurta, S. (2017). Daytime and nighttime urban heat islands
for microscale urban airflow studies. Sustainable Cities and Society, 36, 71–80.
statistical models for Atlanta. Environment and Planning B: Urban Analytics and City
Murphy, K. P. (2012). Machine learning: A probabilistic perspective. MIT press.
Science, 44(2), 308–327.
Myint, S. W., Brazel, A., Okin, G., & Buyantuyev, A. (2010). Combined effects of
Chun, B., & Guldmann, J.-. M. (2014). Spatial statistical analysis and simulation of the
impervious surface and vegetation cover on air temperature variations in a rapidly
urban heat island in high-density central cities. Landscape and Urban Planning, 125,
expanding desert city. GIScience & Remote Sensing, 47(3), 301–320.
76–88.
Oke, T. (1981). Canyon geometry and the nocturnal urban heat island: Comparison of
Ebrahimi Khusfi, Z., Roustaei, F., Ebrahimi Khusfi, M., & Naghavi, S. (2020).
scale model and field observations. Journal of Climatology, 1(3), 237–254. https://
Investigation of the relationship between dust storm index, climatic parameters, and
doi.org/10.1002/joc.3370010304
normalized difference vegetation index using the ridge regression method in arid
Oke, T. (1982). The energetic basis of the urban heat island. Quarterly Journal of the Royal
regions of Central Iran. Arid Land Research and Management, 34(3), 239–263.
Meteorological Society, 108(455), 1–24.
EPA. (2008). Reducing Urban Heat Islands: Compendium of Strategies Urban Heat Island
Oke, T. (1987). Boundary layer climates. Routledge.
Basics.
Oke, T. (1988). Street design and urban canopy layer climate. Energy and Buildings, 11
Ezber, Y., Sen, O. L., Kindap, T., & Karaca, M. (2007). Climatic effects of urbanization in
(1–3), 103–113. https://doi.org/10.1016/0378-7788(88)90026-6
Istanbul: A statistical and modeling analysis. International Journal of Climatology: A
Okumus, G. (2017). A gis based model for measurement and evaluation of urban
Journal of the Royal Meteorological Society, 27(5), 667–679.
sustainability. Istanbul Technical University.
Fan, C., Rey, S. J., & Myint, S. W. (2017). Spatially filtered ridge regression (SFRR): A
Peng, J., Jia, J., Liu, Y., Li, H., & Wu, J. (2018). Seasonal contrast of the dominant factors
regression framework to understanding impacts of land cover patterns on urban
for spatial distribution of land surface temperature in urban areas. Remote Sensing of
climate. Transactions in GIS, 21(5), 862–879.
Environment, 215, 255–267.
Gallo, K. P., & Owen, T. W. (1999). Satellite-based adjustments for the urban heat island
Rao, P. K. (1972). Remote sensing of urban" heat islands" from an environmental
temperature bias. Journal of Applied Meteorology, 38(6), 806–813.
satellite. Bulletin of the American Meteorological Society, 53(7), 647–648.
Giridharan, R., Ganesan, S., & Lau, S. S. Y. (2004). Daytime urban heat island effect in
Rhee, J., Park, S., & Lu, Z. (2014). Relationship between land cover patterns and surface
high-rise and high-density residential developments in Hong Kong. Energy and
temperature in urban areas. GIScience & Remote Sensing, 51(5), 521–536.
Buildings, 36(6), 525–534.
Rizwan, A. M., Dennis, L. Y. C., & Chunho, L. I. U. (2008). A review on the generation,
Giridharan, R., Lau, S. S. Y., & Ganesan, S. (2005). Nocturnal heat island effect in urban
determination and mitigation of Urban Heat Island. Journal of Environmental
residential developments of Hong Kong. Energy and Buildings, 37(9), 964–971.
Sciences, 20(1), 120–128.
Giridharan, R., Lau, S. S. Y., Ganesan, S., & Givoni, B. (2007). Urban design factors
Rosenzweig, C., Solecki, W., & Slosberg, R. (2006). Mitigating New York City’s heat
influencing heat island intensity in high-rise high-density environments of Hong
island with urban forestry, living roofs, and light surfaces. A Report to the New York
Kong. Building and Environment, 42(10), 3669–3684.
State Energy Research and Development Authority.
Guha, S., Govil, H., Dey, A., & Gill, N. (2018). Analytical study of land surface
Roth, M., Oke, T. R., & Emery, W. J. (1989). Satellite-derived urban heat islands from
temperature with NDVI and NDBI using Landsat 8 OLI and TIRS data in Florence and
three coastal cities and the utilization of such data in urban climatology. International
Naples city, Italy. European Journal of Remote Sensing, 51(1), 667–678.
Journal of Remote Sensing, 10(11), 1699–1720.
Guo, A., Yang, J., Xiao, X., Xia, J., Jin, C., & Li, X. (2020). Influences of urban spatial
Santamouris, M., Ding, L., & Osmond, P. (2019). Urban Heat Island Mitigation.
form on urban heat island effects at the community level in China. Sustainable Cities
Decarbonising the built environment (pp. 337–355). Springer.
and Society, 53, Article 101972.
Shishegar, N. (2013). Street Design and Urban Microclimate: Analyzing the Effects of
Guo, G., Zhou, X., Wu, Z., Xiao, R., & Chen, Y. (2016). Characterizing the impact of urban
Street Geometryand Orientation on Airflow and Solar Access in Urban Canyons.
morphology heterogeneity on land surface temperature in Guangzhou, China.
Journal of Clean Energy Technologies, 1(1).
Environmental Modelling & Software, 84, 427–439.
Sobrino, J. A., Jiménez-Muñoz, J. C., & Paolini, L. (2004). Land surface temperature
He, B.-. J., Ding, L., & Prasad, D. (2020a). Relationships among local-scale urban
retrieval from LANDSAT TM 5. Remote Sensing of Environment, 90(4), 434–440.
morphology, urban ventilation, urban heat island and outdoor thermal comfort
Stone Jr,, B., & Rodgers, M. O. (2001). Urban form and thermal efficiency: How the
under sea breeze influence. Sustainable Cities and Society, 60, Article 102289.
design of cities influences the urban heat island effect. American Planning Association.
He, B.-. J., Ding, L., & Prasad, D. (2020b). Urban ventilation and its potential for local
Journal of the American Planning Association, 67(2), 186.
warming mitigation: A field experiment in an open low-rise gridiron precinct.
Strømann-Andersen, J., & Sattrup, P. A. (2011). The urban canyon and building energy
Sustainable Cities and Society, 55, Article 102028.
use: Urban density versus daylight and passive solar gains. Energy and Buildings, 43
Hoerl, A. E., & Kennard, R. W. (1970). Ridge regression: Biased estimation for
(8).
nonorthogonal problems. Technometrics, 12(1), 55–67.
Sun, Y., Gao, C., Li, J., Wang, R., & Liu, J. (2019). Quantifying the effects of urban form
Hu, Y., White, M., & Ding, W. (2016). An urban form experiment on urban heat island
on land surface temperature in subtropical high-density urban areas using machine
effect in high density area. Procedia Engineering, 169, 166–174.
learning. Remote Sensing, 11(8), 959.
Jain, G., & Sarkar, S. (2017). Urban Heat Island : Causes, Effects & Mitigating Strategies. (5).
Terzi, F., & Bolen, F. (2009). Urban sprawl measurement of Istanbul. European Planning
Jiménez-Muñoz, J., Sobrino, J., Gillespie, A., Sabol, D., & Gustafson, W. (2006).
Studies, 17(10), 1559–1570.
Improved land surface emissivities over agricultural areas using ASTER NDVI.
Terzi, F., & Bolen, F. (2012). The potential effects of spatial strategies on urban sprawl in
Remote Sensing of Environment, 103(4), 474–487.
Istanbul. Urban Studies, 49(6), 1229–1250.
Jiménez-Muñoz, J., Sobrino, J., Plaza, A., Guanter, L., Moreno, J., & Martínez, P. (2009).
Thompson, C. G., Kim, R. S., Aloe, A. M., & Becker, B. J. (2017). Extracting the variance
Comparison between fractional vegetation cover retrievals from vegetation indices
inflation factor and other multicollinearity diagnostics from typical regression
and spectral mixture analysis: Case study of PROBA/CHRIS data over an agricultural
results. Basic and Applied Social Psychology, 39(2), 81–90.
area. Sensors, 9(2), 768–793.

13
D. Erdem Okumus and F. Terzi Sustainable Cities and Society 73 (2021) 103128

Todhunter, P. E. (1990). Microclimatic variations attributable to urban-canyon Witten, G.J.D., Hastie, T., & Tibshirani, R. (2013). An introduction to statistical learning.
asymmetry and orientation. Physical Geography, 11(2), 131–141. Xu, H., & Chen, B. (2004). Remote sensing of the urban heat island and its changes in
TSMS. (2018). Turkish State Meteorological Service. Retrieved March 6, 2020, from Xiamen City of SE. Journal of Environmental Sciences, 16(2), 276–281.
General Statistical Data of Istanbul website: Https://www.mgm.gov.tr/veridegerle Yang, J., Ren, J., Sun, D., Xiao, X., Xia, J. C., Jin, C., et al. (2021). Understanding land
ndirme/il-ve-ilceler-istatistik.aspx?k=A&m=ISTANBUL. surface temperature impact factors based on local climate zones. Sustainable Cities
USGS. (2013). No Title. Retrieved June 1, 2018, from http://landsat.usgs.gov/Landsat and Society, 69, Article 102818.
8_Using_Product.php. Yang, X., & Li, Y. (2015). The impact of building density and building height
Population of Istanbul. Turkish Statistical Institute website. Retrieved June 28, 2019, heterogeneity on average urban albedo and street surface temperature. Building and
from https://biruni.tuik.gov.tr/medas/?kn=95&locale=tr. Environment, 90, 146–156.
USGS. (2019). Landsat 8 data users handbook: Version 5. Nasa. 0 November 2019 8 Yin, C., Yuan, M., Lu, Y., Huang, Y., & Liu, Y. (2018). Effects of urban form on the urban
(November), 97 Retrieved from https://landsat.usgs.gov/documents/Landsat8Data heat island effect based on spatial regression model. Science of The Total Environment,
UsersHandbook.pdf. 634, 696–704.
Voogt, J. A., & Oke, T. R. (1997). Complete urban surface temperatures. Journal of Yoo, S. (2018). Investigating important urban characteristics in the formation of urban
Applied Meteorology, 36(9), 1117–1132. heat islands: A machine learning approach. Journal of Big Data, 5(1), 2.
Voogt, J. A., & Oke, T. R. (1998). Effects of urban surface geometry on remotely-sensed Yuan, F., & Bauer, M. E. (2007). Comparison of impervious surface area and normalized
surface temperature. International Journal of Remote Sensing, 19(5), 895–920. difference vegetation index as indicators of surface urban heat island effects in
Weng, Q. (2009). Thermal infrared remote sensing for urban climate and environmental Landsat imagery. Remote Sensing of Environment, 106(3), 375–386.
studies: Methods, applications, and trends. ISPRS Journal of Photogrammetry and Zhou, W., Huang, G., & Cadenasso, M. L. (2011). Does spatial configuration matter?
Remote Sensing, 64(4), 335–344. Understanding the effects of land cover pattern on land surface temperature in urban
Weng, Q., Lu, D., & Schubring, J. (2004). Estimation of land surface landscapes. Landscape and Urban Planning, 102(1), 54–63.
temperature–vegetation abundance relationship for urban heat island studies.
Remote Sensing of Environment, 89(4), 467–483.

14