Unveiling Istanbul’s City Dynamics: Spatiotemporal Hotspot Analysis of Vegetation, Settlement, and Surface Urban Heat Islands

Abstract

Investigation of cities’ spatiotemporal dynamics, including vegetation and urban areas, is of utmost importance for understanding ecological balance, urban planning, and sustainable development. This study investigated the dynamic interactions between vegetation, settlement patterns, and surface urban heat islands (SUHIs) in Istanbul using spatiotemporal hotspot analysis. Utilizing Landsat satellite imagery, we applied the Getis-Ord Gi* statistic to analyze Land Surface Temperature (LST), Urban Index (UI), and Normalized Difference Vegetation Index (NDVI) across the city. Using satellite images and the Getis-Ord Gi* statistic, this research investigated how vegetation and urbanization impact SUHIs. Based on the main results, mean NDVI, UI, and LST values for 2009 and 2017 were analyzed, revealing significant vegetation loss in 37 of Istanbul’s 39 districts, with substantial urbanization, especially in the north, due to new infrastructure development. On the other hand, hotspot analysis was conducted on normalized NDVI, UI, and LST images by analyzing 977 neighborhoods. Results showed a significant transformation of green areas to non-significant classes in NDVI, high urbanization in UI, and the formation of new hot areas in LST. SUHIs were found to cluster in areas with increasing residential and industrial activities, highlighting the role of urban development on SUHI formation. This research can be applied to any region since it offers crucial perspectives for decision-makers and urban planners aiming to mitigate SUHI effects through targeted greening strategies and sustainable urban development. By integrating environmental metrics into urban planning, this study underscores the need for comprehensive and sustainable approaches to enhance urban resilience, reduce environmental impact, and improve livability in Istanbul.

1. Introduction

Urbanization phenomenon, as a consequence of demographic changes and urban land expansion, leads to extreme land use transformation affecting cities’ physical boundaries [1,2]. Over the last decades, an unparalleled pace of urbanization growth has been observed in many cities worldwide [3,4,5,6]. The urbanization process is an unavoidable reality of economic and social development. However, rapid, irregular, and unplanned urbanization, based on spatial distribution and population, is one of the most severe challenges on Earth. Globally, 55% of people live in urban settings, a percentage that is predicted to increase to 68% by the year 2050 [7]. These predictions indicate a continued trend in urban areas. The fast expansion of urbanization worldwide not only results in alterations to the land cover, but also influences the surface roughness, albedo, and vegetation coverage affecting the circulation of the urban hydrological, ecological, and climate systems [8]. Moreover, urbanization is commonly associated with critical environmental problems such as increasing greenhouse gas emissions [9], environmental pollution [10], natural habitat loss [11], Urban Heat Island (UHI) effect [12,13,14], and Land Surface Temperature (LST) increases [15,16,17,18].

A region that is warmer than its surroundings is called a Heat Island (HI). It can appear in rural and urban environments as a Rural Heat Island (RHI) or a UHI, respectively. A UHI effect has negative influences on both the environment and people since it causes high energy consumption [19], raises the level of air pollution [20], and thus has an impact on public health [21]. Therefore, it is crucial to comprehend and mitigate these effects. The two primary forms of UHIs, which the U.S. Environmental Protection Agency (EPA) has introduced, are atmospheric UHIs (AUHIs) and surface UHIs (SUHIs) [22]. AUHIs are identified by direct measurements of weather stations and mobile traverses, whereas SUHIs are determined by indirect measurements such as Remote Sensing (RS) technologies. thermal infrared (TIR) RS technologies enable efficient assessment of Land Surface Temperature (LST) and Water Surface Temperature (WST) using spaceborne and airborne systems such as satellites or unmanned aerial vehicles. These measurements can help explain the SUHI effect better than traditional methods and offer cost-effective and time-effective solutions at varying scales. Furthermore, the TIR-based LST provides a more accurate understanding of climatic conditions compared with the conventional weather stations.

Examining the geographical distribution of Land Use Land Cover (LULC) and the changes brought about by LULC over time is necessary to fully understand the spatial patterns of SUHIs. These changes affect the LST distribution, thus, to more accurately assess the SUHI effect, it is crucial to examine the LULC change. Considering the literature, numerous researchers have used various LULC extraction techniques to examine the effects of LULC alterations on LST [23,24,25,26,27]. On the other hand, several studies analyzed LST-related indicators such as Normalized Difference Vegetation Index (NDVI), Impervious Surface Area (ISA), Normalized Difference Water Index (NDWI), Fractional Vegetation Cover (FVC), and Normalized Difference Built-up Index (NDBI) since these indicators generally show a strong relationship between LST and SUHI effect [28,29,30,31,32,33,34]. The aforementioned studies generally presented identical results, showing that built-up and bare lands usually retain higher LST values, causing the SUHI effect to grow, and vegetative lands have low LST values, helping to mitigate the SUHI effect. Additionally, the drawbacks of the abovementioned works are (i) using only one image for each year or season and (ii) conducting interannual LST analyses without normalization, which reduces the temporal variability of the atmospheric conditions [35,36]. In this study, both of these issues were considered.

Concerning the literature, various methods such as mean LST [37,38], Discomfort Index (DI) [39,40], and Urban Thermal Field Variance Index (UTFVI) [41,42] are commonly employed to demonstrate the SUHI effect’s spatial extent. Apart from these methods, the Getis-Ord Gi* statistic, a metric that shows each data point’s spatial autocorrelations [43], is one of the most preferred ways of illustrating the spatial distribution of hotspots and coldspots in many different applications involving agriculture [44,45], forest fire hotspots [46], healthcare [47], economic development [48], environmental pollution incidents [49], and SUHIs [50,51,52]. The Getis-Ord Gi* statistic on LST images has been utilized in several investigations to determine hotspots and coldspots. However, the analysis often ignores hot or cold spots in built-up and vegetation areas. Furthermore, the study areas analyzed are mostly limited to city-scale boundaries.

The mentioned LST and SUHI studies generally focused on analyzing the LST changes, mainly considering the vegetative and built-up areas or general LULC classes. Even though these studies’ results are essential to understanding the SUHI phenomenon, providing the SUHI effect on an administrative scale is more critical. Only the study conducted by Ranagalage et al. [53] investigated SUHI distribution in Colombo and considered administrative borders. However, they just used two satellite images acquired in 1997 and 2017 in a dry season without normalization of LST. This study aims to extract and analyze the hotspots and coldspots of NDVI, Urban Index (UI), and LST using Landsat satellite imagery for Istanbul city, Türkiye, from 2009 and 2017. The original contributions of this research can be categorized as (i) considering mean LST, UI, and NDVI images calculated from three images for each year to represent the corresponding year accurately, (ii) applying normalization to all mean images to lessen the atmospheric conditions’ temporal unpredictability, and (iii) extracting the hotspots and coldspots of the LST, UI, and NDVI at a neighborhood scale and interpreting the obtained results at a district level, which none of the previous studies considered.

2. Study Area and Data

Istanbul, which is situated in northwest Türkiye between 28°01′–29°55′ E and 40°28′–41°33′ E, has been chosen as the study area (Figure 1). Istanbul, with 39 districts and a population of approximately 16 million, is ranked among the world′s largest 15 megacities. The city, which serves as a bridge between the continents of Asia and Europe, is divided into two parts by the Istanbul Strait, a vital sea transportation route connecting the Black Sea and the Sea of Marmara. Istanbul′s climate is characterized as transitional between that of the Mediterranean and the Black Sea, with moderate summer temperatures and a fair quantity of precipitation throughout the summer months. The city′s closeness to the sea and the presence of a substantial body of water—the Bosphorus Strait—that passes through the city have a major impact on the climate [54].

With its advanced industry, Istanbul, a significant economic and cultural center, meets 20% of industrial employment in Türkiye. Due to the high rate of industrialization, Istanbul has experienced rapid and uncontrolled population growth in the last 65 years. As a result, unplanned urbanization rates and industrial areas have increased in the city. This situation has caused a dramatic change in land cover, environmental problems, and an abnormal increase in surface temperature values [15,55,56]. As a result of all these changes, formation of the UHI is an expected condition. Determining and analyzing these variations using remote sensing and spatial statistical methods is crucial for sustainable urban monitoring and planning in the megacity of Istanbul.

Concerning satellite imagery, two of the most widely utilized satellites for environmental monitoring studies are Landsat 5 and Landsat 8, which are outfitted with Thematic Mapper (TM) and Operational Land Imager (OLI) and Thermal Infrared Sensor (TIRS), respectively. Within this context, Landsat 5 TM images acquired on 17 June 2009, 19 July 2009, and 5 September 2009, and Landsat 8 OLI and TIRS images acquired on 23 June 2017, 25 July 2017, and 11 September 2017 were utilized in this study. We selected 2017 Landsat 8 data based on the period of construction of three mega construction projects in İstanbul, namely, Istanbul Airport (3rd airport in the city), the Yavuz Sultan Selim Bridge (3rd Bosphorus Bridge in the city), and the connection roads (North Marmara Highway). Concerning the 2009 datasets, they were selected because they had cloud-free images for the same months as in 2017. Therefore, the SUHI conditions in the city were investigated before the mega projects and just after the mega projects. The spatial resolution of the Landsat 5 and Landsat 8 bands used to extract LST, NDVI, and UI is 30 m. Thermal bands originally had a spatial resolution of 120 m for Landsat 5 TM and 100 m for Landsat 8 TIRS. However, the USGS provides them at a 30 m resolution after applying cubic convolution resampling. Furthermore, station-based near-surface air temperature and relative humidity measurements, which are collected simultaneously with Landsat acquisitions, are used in the application. The meteorological station data were combined with the mono-window method to extract LST from remotely sensed data.

3. Methodology

In this study, Landsat 5 TM and Landsat 8 OLI/TIRS images were used to calculate the mean and normalized NDVI and UI for the years 2009 and 2017. The mono-window algorithm was used to obtain LST, and the Getis-Ord Gi* statistics, used to identify spatial clusters of high or low values, was applied to normalized LST, NDVI, and UI images to examine the effects of the SUHIs and to determine the effect of settled areas at the administrative border. Figure 2 illustrates the steps of the selected methodology used in the study. Briefly, three sets of Landsat 5 and Landsat 8 images were utilized to extract UI, NDVI, and LST images. Then, mean UI, NDVI, and LST images were generated. After that, vulnerability maps were retrieved, and normalized index images were obtained to apply the Getis-Ord Gi* statistics. Detailed information on the methodological steps is presented in the subtitles below.

3.1. Image Pre-Processing and NDVI Calculation

Solar elevation angle and the distance between the earth and the sun affect the variations in solar illumination; the effects of these two parameters are eliminated with the reflectance conversion [18]. Concerning the reflectance conversion of the Landsat 5 TM data, first, the radiance conversion is applied and then the radiances are converted into the reflectance values using Equation (1) [57]. However, for Landsat 8 data, Digital Number (DN) values can be directly converted into Top of Atmosphere (TOA) reflectance as in Equation (2) [58].

$${\mathsf{\rho }}_{\mathsf{\lambda }}=\frac{\mathsf{\pi }\cdot {\mathrm{L}}_{\mathsf{\lambda }}\cdot {\mathrm{d}}^2}{{\mathrm{ESUN}}_{\mathsf{\lambda }}\cdot {\cos \mathsf{\theta }}_{\mathrm{s}}}$$

(1)

where ${\mathsf{\rho }}_{\mathsf{\lambda }}$ is unitless TOA reflectance, θ_s is the solar zenith angle, ESUN_λ (Watts/(m²∙μm)) refers to the mean solar exo-atmospheric spectral irradiances, and d is the Earth–Sun distance in astronomical units. ESUN_λ values for each band of Landsat 5 can be obtained from the mission’s handbook [57]. On the other hand, the metadata file provides the d and θ_s values.

$${\mathsf{\rho }}_{\mathsf{\lambda }}=\frac{{ \mathrm{M}}_{\mathrm{p}}·{\mathrm{Q}}_{\mathrm{CAL}}+{\mathrm{A}}_{\mathrm{p}}}{{\sin \mathsf{\theta }}_{\mathrm{SE}}}$$

(2)

where ${\mathsf{\theta }}_{\mathrm{SE}}$ is the local sun elevation angle, ${\mathrm{A}}_{\mathrm{p}}$ is the band-specific additive rescaling factor, and ${ \mathrm{M}}_{\mathrm{p}}$ is the band-specific multiplicative rescaling factor, which exists in the metadata file.

The Normalized Difference Vegetation Index (NDVI) ranges from −1 to +1. A higher positive value indicates the presence of vegetation in the remotely sensed data. A lower positive score, on the other hand, denotes built-up or barren soils. Water bodies are indicated by negative values. The typical range for green and healthy vegetation in NDVI is between 0.2 and 0.8. The band math between the reflectance of the near-infrared (NIR) band and the red (R) band is utilized to compute the NDVI as in Equation (3) [59]. The NIR band is defined as band 4 and the Red band is defined as band 3 by Landsat 5 TM, whereas the NIR band is defined as band 5 and the Red band is defined as band 4 by Landsat 8 OLI/TIRS.

$$\mathrm{NDVI}=\frac{{\mathsf{\rho }}_{\mathrm{NIR}}-{\mathsf{\rho }}_{\mathrm{red}}}{{\mathsf{\rho }}_{\mathrm{NIR}}+{\mathsf{\rho }}_{\mathrm{red}}}$$

(3)

where ${\mathsf{\rho }}_{\mathrm{NIR}}$ is the NIR reflectance band and ${\mathsf{\rho }}_{\mathrm{red}}$ refers to the Red reflectance band.

3.2. Urban Index (UI) Calculation

The Urban Index (UI), as proposed by Kawamura et al. [60] in 1996, utilizes the relationship between the reflectance of urban areas in the near-infrared (NIR) and shortwave infrared 2 (SWIR2) regions. With increasing urban density, the UI value also increases, making it a helpful tool in identifying residential areas, as demonstrated by Gautam et al. in 2018 [61]. The UI is determined by analyzing the electromagnetic spectrum’s near-infrared (NIR) and shortwave infrared 2 (SWIR2) regions of Landsat images (Equation (4)). The NIR band is defined as band 4 and the SWIR2 band is defined as band 7 by Landsat 5 TM, whereas the NIR band is defined as band 5 and the SWIR2 band is defined as band 7 by Landsat 8 OLI/TIRS.

$$\mathrm{UI}=\frac{{\mathsf{\rho }}_{\mathrm{SWIR}2}-{\mathsf{\rho }}_{\mathrm{NIR}}}{{\mathsf{\rho }}_{\mathrm{SWIR}2}+{\mathsf{\rho }}_{\mathrm{NIR}}}$$

(4)

3.3. Calculation of LST Using the Mono-Window Algorithm

Mono-window Algorithm (MWA), a widely used method in remote sensing, was employed to obtain LST data. This algorithm uses the thermal infrared band of the satellite sensor to estimate the land surface temperature. Qin et al. [62] introduced the MWA to derive LST retrieval initially from Landsat TM data. Land Surface Emissivity (LSE), effective mean atmospheric temperature, and atmospheric transmittance are the three main crucial factors for MWA-based LST retrieval. LST values can be obtained from MWA using Equation (5).

$$\begin{array}{l}{\mathrm{LST}}_{\mathrm{MWA}}=\left\{\mathrm{a}\cdot \left(1-\mathrm{C}-\mathrm{D}\right)+\left[\mathrm{b}\cdot \left(1-\mathrm{C}-\mathrm{D}\right)+\mathrm{C}+\mathrm{D}\right]\cdot \mathrm{T}-\mathrm{D}\cdot {\mathrm{T}}_{\mathrm{a}}\right\}÷\mathrm{C}\\ \mathrm{C}=\mathsf{\epsilon }\times \mathsf{\tau },\mathrm{D}=(1-\mathsf{\tau })[1+(1-\mathsf{\epsilon })\times \mathsf{\tau }],\mathrm{a}=-67.355351,\mathrm{b}=0.458606\end{array}$$

(5)

where T (K) represents the brightness temperature, C and D are the algorithm parameters, T_a (K) refers to the effective mean atmospheric temperature, ε represents LSE, τ is the atmospheric transmittance, and a and b are the algorithm constants. The theoretical estimations of the above-mentioned parameters are presented in the following Sections.

3.3.1. Brightness Temperature (T) Calculation

In order to calculate the brightness temperature from Landsat data, radiance conversion is first applied to the Digital Number (DN) values of the related TIR band. The radiance conversion of any Landsat 5 TM band is performed using Equation (6) [57], whereas Equation (7) is applied for the radiance conversion of Landsat 8 data [58].

$$\mathrm{L}\mathsf{\lambda }=\left[\frac{{\mathrm{L}}_{\mathrm{MAX}\mathsf{\lambda }}-{\mathrm{L}}_{\mathrm{MIN}\mathsf{\lambda }}}{{\mathrm{QCAL}}_{\mathrm{MAX}}-{\mathrm{QCAL}}_{\mathrm{MIN}}}\right]\times \left[{\mathrm{Q}}_{\mathrm{CAL}}-{\mathrm{QCAL}}_{\mathrm{MIN}}\right]+{\mathrm{L}}_{\mathrm{MIN}\mathsf{\lambda }}$$

(6)

where L_λ (Watts/(m²∙srad∙μm)) is TOA spectral radiance, L_MAXλ (Watts/(m²∙srad∙μm)) is the spectral radiance scaled to QCAL_MAX, L_MINλ (Watts/(m²∙srad∙μm)) is the spectral radiance scaled to QCAL_MIN, QCAL_MAX is the maximum quantized calibrated pixel value in DN, QCAL_MIN is the minimum quantized calibrated pixel value in DN, and Q_CAL is the quantized calibrated pixel value in DN. The metadata file of the relevant Landsat TM data contains the information for all variables in Equation (6).

$${\mathrm{L}}_{\mathsf{\lambda }}={ \mathrm{M}}_{\mathrm{L}}·{\mathrm{Q}}_{\mathrm{CAL}}+{\mathrm{A}}_{\mathrm{L}}$$

(7)

where ${\mathrm{A}}_{\mathrm{L}}$ is the band-specific additive rescaling factor and ${\mathrm{M}}_{\mathrm{L}}$ is the band-specific multiplicative rescaling factor. As in Equation (6), the metadata file for the relevant Landsat 8 data includes the values of these two variables. After the radiance conversion of the corresponding Landsat data, the brightness temperature calculation is conducted using Equation (8) [57,58].

$$\mathrm{T}=\frac{{\mathrm{K}}_2}{\ln \left(\frac{{\mathrm{K}}_1}{{\mathrm{L}}_{\mathsf{\lambda }}}+1\right)}$$

(8)

The calibration constants K₁ (Watts/(m²∙srad∙μm)) and K₂ (Kelvin) are 607.76 (Watts/(m²∙srad∙μm)) and 1260.56 K for Landsat 5 band 6, and 774.89 (Watts/(m²∙srad∙μm)) and 1321.08 K for Landsat 8 band 10, respectively [57,58].

3.3.2. Estimation of the Effective Mean Atmospheric Temperature (T_a)

As a crucial MWA parameter, the practical formulas for estimating the effective mean atmospheric temperature (T_a) from near-surface air temperature (T_o) are shown in

Table 1. Practical formulas for estimating effective mean atmospheric temperature (T_a) from near-surface air temperature (T_o).

Region	Mean Atmospheric Temperature (Ta) in Kelvin
USA 1976 Region	Ta = 25.940 + 0.8805 × To
Tropical Region	Ta = 17.977 + 0.9172 × To
Mid-latitude Summer Region	Ta = 16.011 + 0.9262 × To
Mid-latitude Winter Region	Ta = 19.270 + 0.9112 × To

[62]. The mid-latitude summer area was taken into account in the computation of this work.

3.3.3. Atmospheric Transmittance (τ) Estimation

The atmospheric transmittance is another substantial variable of the MWA, which can be estimated depending on water vapor content (w). Meteorological stations measure the water vapor content directly; it can also be calculated from relative humidity and near-surface air temperature [30], as we considered in this study. Qin et al. [62] suggested simple equations for Landsat TM data to determine atmospheric transmittance using water vapor content (

Table 2. Simple equations for estimating atmospheric transmittance via water vapor.

Profiles	Water Vapor (w) (g/cm2)	Transmittance Estimation Equation (τ)	Squared Correlation	Standard Error
High Air Temperature	0.4–1.6	0.974290–0.08007 × w	0.996	0.00237
	1.6–3.0	1.031412–0.11536 × w	0.998	0.00254
Low Air Temperature	0.4–1.6	0.982007–0.09611 × w	0.996	0.00334
	1.6–3.0	1.053710–0.14142 × w	0.999	0.00238

).

3.3.4. Calculation of Land Surface Emissivity (LSE, ε)

LSE is not only vital for MWA, but it is also vital for all other methods to acquire satellite-based LST accurately. Roughness, structure, water content, and chemical composition are among the important parameters that control the emissivity of a surface [63,64]. So far, researchers have developed many methods for obtaining LSE from satellite images [63,65,66,67,68,69,70]. The Normalized Difference Vegetation Index (NDVI)-based threshold method is an easy and alternative way for satellite-based LSE retrieval. Different NDVI-based LSE models that provide satisfactory results have been introduced in the literature [69,70,71,72]. Zhang et al. [72] proposed LSE estimation from the NDVI thresholds as in

Table 3. Estimation of LSE depending on NDVI thresholds.

NDVI	LSE (ε)
NDVI < −0.185	0.995
−0.185 ≤ NDVI < 0.157	0.970
0.157 ≤ NDVI ≤ 0.727	1.0094 + 0.047 × ln(NDVI)
NDVI > 0.727	0.990

, and in this study, we used this NDVI-based threshold method for LSE estimation.

3.3.5. NDVI, UI, and LST Standardization

The topographic and climatic variances introduce bias; therefore, it is inappropriate to compare seasonal and annual variations for different years without standardization [73,74,75]. The standardization, enabling the creation of vulnerability maps, was done using the equations in

Table 4. Standardization equations.

Category	Category Interval
Very Low	DN ≤ DNmean − 1.5 × σ
Low	DNmean − 1.5 × σ < DN ≤ DNmean − σ
Mean	DNmean − σ < DN ≤ DNmean + σ
High	DNmean + σ < DN ≤ DNmean + 1.5 × σ
Very High	DN > DNmean + 1.5 × σ

. The calculated NDVI, UI, and LST values were standardized for 2009 and 2017 to make the changes proportional to each other before change detection.

3.3.6. Pearson Correlation Calculation

In this study, the response of the LST to the change in the NDVI and UI was estimated based on Pearson correlation analysis to quantify the strength of the association between the variables. The range of values for the Pearson correlation coefficient is (−1,1), where −1 and +1 indicate a completely negative and positive correlation, respectively, while the 0 value stands for no correlation.

3.3.7. Hotspot and Coldspot Retrieval (Getis–Ord Gi*)

In the last stage, the Getis-Ord Gi* statistics, a spatial statistical method used to identify spatial clusters of high or low values, was applied to normalize LST, NDVI, and UI images [43]. This method helps to identify areas where the surface temperature is significantly higher or lower than the average, indicating the presence of SUHIs or other temperature anomalies. Hotspots refer to areas with high feature values, while coldspots refer to areas with low feature values. The Getis-Ord Gi* statistic is calculated using Equations (9)–(11).

$$G_i^{*}=\frac{{\sum }_{j=1}^n{\omega }_{i,j}{x}_j-\overline{x}{\sum }_{j=1}^n{\omega }_{i,j}}{S\sqrt{\frac{\left[n\sum _{j=1}^n{\omega }_{i,j}^2-{\left(\sum _{j=1}^n{\omega }_{i,j}\right)}^2\right]}{n-1}}}$$

(9)

$$\overline{x}=\frac{{\sum }_{j=1}^n{x}_j}{n}$$

(10)

$$S=\sqrt{\frac{{\sum }_{j=1}^n{x_j}^2}{n}-{\overline{x}}^2}$$

(11)

where n is the total number of features, w_i,j is defined by the spatial weight between features i and j, and x_j is defined by the feature attribute value of j.

4. Results

Based on the applied methodology, NDVI, UI, and LST analyses are presented under separate subtitles below. In addition, vulnerability maps and hotspot–coldspot analyses were examined to offer insightful information on sustainable urban management and planning in Istanbul.

4.1. NDVI Results

Both NDVI and UI results for Istanbul, when examined using district boundaries, revealed intriguing patterns. In NDVI images (Figure 3), the darkest color mainly shows the artificial surfaces, water bodies, and cultivated agricultural fields and the brightest color shows the forest and green areas. The districts with the highest NDVI values for 2009 were Şile (0.62) and Beykoz (0.59), while the lowest values were found in Bayrampaşa (0.11) and Bağcılar (0.11). Similarly, the highest NDVI values for 2017 were observed in Şile (0.43) and Çatalca (0.41), with the lowest values in Bayrampaşa (0.10) and Bağcılar (0.09). Notably, a decrease of 0.12 or more was observed in NDVI values in Çekmeköy, Eyüp, Beykoz, Şile, Sarıyer, and Çatalca districts from 2009 to 2017. Between 2009 and 2017, while Avcılar and Büyükçekmece districts had an increase in mean NDVI values, all other 37 districts in Istanbul experienced a drastic decrease. These findings offer valuable insights into the changing and transforming landscape of Istanbul, and its implications for sustainable urban planning and management.

By standardizing the mean NDVI images, NDVI vulnerability maps were produced as in Figure 4. Between 2009 and 2017, the area with a low NDVI and high NDVI category decreased by 13,648.3 ha and 71,256.3 ha, respectively, based on the categories of the NDVI vulnerability maps (Figure 5). Figure 6 represents the change map of the corresponding NDVI vulnerability categories to better understand the spatial variation between 2009 and 2017. It is clear from Figure 6 that the Northern part of Istanbul was exposed to drastic NDVI changes from the high category to the low and medium categories between 2009 and 2017.

4.2. UI Results

Based on the resultant UI images, the darkest color shows the forest and green areas and the brightest color shows artificial surfaces, bare lands, and other land covers without vegetation cover (Figure 7). Upon examining the mean UI images within the district administrative borders of 2009, it was discovered that Güngören (0.02) and Gaziosmapaşa (0.01) had the highest UI values, while Beykoz (−0.51) and Şile (−0.53) had the lowest. When comparing the 2017 UI values with the ones in 2009, Güngören (−0.04) and Gaziomanpaşa (−0.05) had the highest values, while the lowest were observed in Şile (−0.40) and Çatalca (−0.37) districts. Beykoz is a district through which the Yavuz Sultan Selim (3rd Bridge of İstanbul) connection roads pass. Unfortunately, the construction of these roads has led to the destruction of green areas in Beykoz. As a result, there has been an increase in artificial surfaces (roads) and a decrease in green areas within the district [76].

An increase of 0.09 and above was observed in districts such as Arnavutköy, Beykoz, Çekmeköy, Çatalca, Eyüp, Sarıyer, and Şile, where green areas are more dense. In particular, Beykoz and Sarıyer districts host the Northern Forest in Istanbul. Furthermore, increases were observed in districts such as Avcılar, Eyüp, Kartal, Küçükçekmece, Pendik, Sancaktepe, and Sultangazi.

Considering the results obtained at a district level, it was noticed that there were consistent outcomes between UI and NDVI values. In districts where NDVI values, used to determine green areas, were high, UI values, used to determine residential areas, were low; on the other hand, in districts where NDVI values were low, UI values were high.

Figure 8 illustrates the UI vulnerability maps of 2009 and 2017. The results show an increase of 48,481.8 ha in the high and very high classes from 2009 to 2017, while a decrease of 36,528.1 ha was observed in the low class (Figure 9). Figure 10 illustrates the change map of the corresponding UI vulnerability categories to better understand the spatial variation between 2009 and 2017. Figure 10 also highlights the decrease in the low classes and an increase in the high classes.

4.3. LST Results

Within the scope of this study, LST values were retrieved from Landsat 5 TM and Landsat 8 OLI and TIR images acquired in 2009 (June, July, and September) and 2017 (June, July, and September). On the other hand, the Mono-Window algorithm was used to calculate LST values and

Table 5. Detailed information on the parameters needed to calculate LST for the corresponding dates.

Data Type	Parameter	Objective	2009			2017
			17 June	19 July	5 September	23 June	25 July	11 September
Calculated	Near-Surface Air Temperature (T0)	Calculation of Mean Atmospheric Temperature and Land Surface Temperature	299.20 K	303.10 K	302.78 K	302.16 K	306.65 K	304.06 K
	Mean Atmospheric Temperature (Ta)	Calculation of Land Surface Temperature	293.13 K	296.74 K	296.43 K	295.88 K	300.03 K	297.64 K
	Relative humidity (RH)	Calculation of Atmospheric Water Vapor	0.4793	0.5098	0.3330	0.3917	0.3451	0.4571
	Atmospheric Water Vapor (wi)	Calculation of Atmospheric Transmittance (τ)	1.46	1.85	1.25	1.40	1.53	1.75
	Land Surface Emissivity ($\mathsf{\epsilon }$)	Calculation of Land Surface Temperature	NDVI threshold	NDVI threshold	NDVI threshold	NDVI threshold	NDVI threshold	NDVI threshold
Constant	a b	Calculation of Land Surface Temperature	A = −67.3553 B = 0.4586
	K1 K2	Calculation of Brightness Temperature	K1 L5 = 607.76 (Watts/(m2∙srad∙μm)), K1 L8 = 774.89 (Watts/(m2∙srad∙μm)) K2 L5 = 1260.56 K, K2 L8 = 1321.08 K

shows all the detailed information needed to calculate LST values for İstanbul.

According to the LST maps, the highest and the lowest LST values for 2009 were determined as 323.36 K and 292.70 K, respectively, and for 2017, 324.88 K and 295.50 K (Figure 11). The LST difference between the hottest and coldest area is calculated to be approximately 29 K. When the LST values were analyzed at a district level, it was found that Şile (301.47 K), Adalar (301.51 K), and Beykoz (302.07 K) had the lowest LSTs in 2009. Meanwhile, Gaziosmanpaşa (309.09 K), Bağcılar (309.02 K), and Sultanbeyli (308.97 K) had the highest LSTs. Similarly, in the year 2017, Şile (302.67 K), Beykoz (303.27 K), and Adalar (303.32 K) had the lowest LSTs. On the other hand, Bayrampaşa (312.56 K), Gaziosmanpaşa (312.37 K), and Bağcılar (312.02 K) had the highest LST values. Furthermore, when the mean LST values were analyzed based on the district administrative boundaries, it was observed that there was an increase in all districts, except Silivri, between 2009 and 2017. In particular, there was a minimum increase of 2.5 K in the districts of Bağcılar, Bahçelievler, Bakırköy, Bayrampaşa, Beyoğlu, Esenyurt, Fatih, Güngören, Gaziosmanpaşa, Kağıthane, Kartal, Şişli, and Zeytinburnu.

Moreover, it was found that districts with high LST values in both years corresponded to artificial surfaces based on the field surveys. Conversely, the areas with the lowest LST values were in regions where forest and green areas were concentrated. In addition, the mean LST values in the regions corresponding to forest and green areas increased from 302.40 K in 2009 to 303.56 K in 2017. It is worth noting that building materials used in urban areas can retain more heat than surfaces in rural areas, causing heat increases. It has been previously stated that building materials used in urban areas can retain heat more than surfaces in rural areas and then spread to the environment, causing heat increases. The mean LST values in the regions corresponding to settlements increased from 308.18 K in 2009 to 311.09 K in 2017.

Figure 12 illustrates the LST vulnerability maps of 2009 and 2017. From 2009 to 2017, the very high LST class increased by 14,442.2 ha while the very low LST class decreased by 11,424.9 ha (Figure 13). Figure 14 illustrates the change map of the corresponding LST vulnerability categories to better understand the spatial variations between 2009 and 2017. Figure 14 shows that in the agricultural zones (western part of the study area), the change is from high and medium to low, while in the urbanized areas (the central south part of the study area), the change is from almost all classes to high and very high ones.

4.4. Relationship between LST, NDVI, and UI

Regression analysis and Pearson correlation analysis were carried out to examine the effects of green areas and artificial surfaces on LST. Scatter graphs were created by selecting random points corresponding to the same pixel in each image to determine the statistical relationship (Figure 15). The images used in this analysis were normalized NDVI, UI, and LST images. The Pearson correlation coefficients for the year 2009 were calculated to be −0.80 for NDVI–LST and 0.85 for UI–LST. Similarly, for 2017, they were calculated to be −0.82 for NDVI–LST and 0.86 for UI–LST. The correlation results showed that while LST has a positive correlation with UI, it has a negative correlation with NDVI. The slope of regression lines in Figure 15 represents the change in LST as NDVI and UI changes. Based on this information, Figure 15a,c reveal that the normalized LST decreases by approximately 0.7861 and 0.5768 units for every 1 unit rise in normalized NDVI for 2009 and 2017, respectively. On the other hand, Figure 15b,d highlights that the normalized LST increases by approximately 0.7157 and 0.5533 units for every 1 unit rise in normalized UI for 2009 and 2017, respectively.

4.5. Hotspot and Coldspot (Getis-Ord Gi*) Analysis

The Getis–Ord Gi* approach was used to perform a hotspot analysis of the spatial distribution of NDVI, UI, and LST over Istanbul city. In order to determine hotspots and coldspots, this method takes into account the values of nearby characteristics in the city landscape based on NDVI, UI, and LST values. In general, clusters of features with high values are called hotspots, whereas features with low values are called coldspots. Seven classifications were identified from the research on the city landscape, namely, extremely cold, cold, cool, not significant, warm, hot, and very hot.

In the application, the raster data were first converted to vector format to statistically map hot and cold areas, with a mean pixel value assigned for each administrative border (neighborhood border). A total of 977 neighborhood regions in Istanbul were analyzed for hot areas. An incremental spatial autocorrelation tool utilizing the Global Moran’s I statistic was used to determine the appropriate analysis distance before applying hot/cold spot area analysis. This tool measures the density of spatial clustering for a series of increasing distances, with the clustering density shown by the z-score. The z-score typically increases as the distance increases, indicating that clustering is concentrated. However, the z-score usually peaks at a certain distance. In this study region, there were multiple distances where the z-score peaked. Therefore, a distance was chosen based on the clusters of administrative units adjacent to each other.

The hot and cold areas were clustered according to Z-scores: hot area (99% significant), hot area (95% significant), hot area (90% significant), not statistically significant, cold area (90% significant), cold area (95% significant), and cold area (99% significant).

4.5.1. NDVI Getis-Ord Gi* Analysis Results

Upon examination of the change between years, the results showed that the hotspot areas (areas with significant and high NDVI values) are clustered in the northern part of Istanbul, where there are dense forests. On the other hand, coldspot areas are clustered in the central-southern part of Istanbul, where most residential areas are located (as shown in Figure 16). From 2009 to 2017, it was noted that 9021.7 ha of the area changed from hot areas to non-significant areas, while 22,588.0 ha of the area changed from cold areas to non-significant areas. In the 8 years, it was also observed that 23,998.0 ha of cold areas and 4161.0 ha of hot areas were formed from non-significant areas.

A closer look at the changes at a district level shows that in Başakşehir district, where there has been an increase in residential areas over the years, 3984.5 ha have changed from 90% confidence interval cold area class to 99% confidence interval cold area class, and 17,021.0 ha have changed from non-significant areas to cold areas. Moreover, it was found that 821.9 ha in Beylikdüzü district, 473.9 ha in Üsküdar, 204.63 ha in Ümraniye, and 543.8 ha in Beşiktaş have transformed from non-significant areas into cold areas, creating new cold areas. In the Arnavutköy district, where the 3rd airport (The İstanbul Airport) was built, it was determined that 6862.5 ha of land had turned from hot areas to non-significant areas.

4.5.2. UI Getis-Ord Gi* Analysis Results

When comparing changes over the years, it can be seen that the hottest areas are situated in Istanbul’s central-southern region. In contrast, cold areas are concentrated in the northern part (Figure 17). From 2009 to 2017, 15.5 ha of cold areas turned into non-significant areas, while 1779 ha of hot areas transformed into non-significant areas. Over the same eight-year period, 3.2 ha of hot areas and 19.0 ha of cold areas changed from non-significant areas. At a district level, new hot areas were formed on 4214.8 ha in the Başakşehir district, where new residential areas have been growing over the years, and 1765.8 ha in the Büyükçekmece district. Additionally, 10,042.8 ha of land in the Arnavutköy district changed from cold to non-significant.

4.5.3. LST Getis-Ord Gi* Analysis Results

In general, the suburban region near water and vegetation held the majority of coldspots, whereas the core urban area regularly included hotspots (Figure 18). The hotspot and coldspot clustering tended to increase during the eight-year period. Forest areas are the coldest regions—the map for 2009 shows that the uncultivated agricultural areas on the European Side also have high temperatures. The previous sections explained this situation as the reflected temperature of the soil in the thermal zone changes with the water content. In addition, it is seen that the clustering in the cold areas in the region where Istanbul Airport is built has turned into a non-significant distribution due to the effect of warming.

When change over the years was examined statistically, it was seen that the cold area, corresponding to an area of 3148 ha, and the hot area, corresponding to an area of 71,955 ha, turned into a non-significant cluster. It was also observed that the non-significant regions turned into new cold areas of 49,624 ha and new hot areas of 10,032 ha. Concerning the district-based change detection analysis, 1789.116 ha in Bakırköy district, 479.83 ha in Avcılar, 426.32 ha in Küçükçekmece, 881.99 ha in Kadıköy, 1686.14 ha in Maltepe, and 1463.56 ha in Kartal are the regions where new hot areas formed. In Arnavutköy district, it was observed that an area of 3183.36 ha turned from a cold area into a non-significant cluster. According to the classification results, cold areas occur in Sarıyer, Beykoz, Şile, and Çatalca districts, although they also include residential areas. This is because green and forest areas are very dense in these districts. Hot areas are seen in Küçükçekmece, Bahçelievler, Güngören, Bayrampaşa, Beyoğlu, and Fatih districts on the European side, where settlements are always dense and known to increase over time, and in Ümraniye and Sancaktepe districts on the Anatolian side, in both years.

5. Discussion

Istanbul, Türkiye’s significant economic, cultural, and educational center, has undergone rapid and unplanned urbanization over the years due to excessive population growth and migration from rural areas. During this rapid urbanization process, and with unplanned LULC changes, the majority of naturally vegetated regions have been replaced by impervious surfaces such as roads and bridge (the Yavuz Sultan Selim Bridge and North Marmara Highway and its connections), buildings (new settlement areas), and industrial areas (the İstanbul Airport) [15,56,77]. This change in land use has caused an increase in LST and, therefore, in the SUHI effect. Khorrami and Gunduz [78] stated that the densely populated areas of Istanbul’s core districts are where the heat island phenomenon is concentrated, according to the SUHI maps for the spring and summer of 2017–2018. From 2009 to 2017, LST in Istanbul continued to increase. Based on Getis-Ord Gi* figures, the hot/coldspots in the city center and its suburbs continued to expand, while the hotspots in the southeast and southwest continued to stretch. These results show that the LST has deteriorated and become more spatially heterogeneous as a result of increased urbanization [79]. Algancı [77] analyzed LULC changes due to the rapid urbanization rate in Istanbul. In the study, it was stated that a 16,509.7 ha area was transformed into settlement or mega construction projects between the years 2013 and 2017, mainly due to mega construction project-oriented changes. These outputs support the land cover change-based SUHI increase in İstanbul.

It is vital to understand that the LST values represent the skin temperature of the surface, not the air. This means that there could be significant temperature differences between these two variables depending on the land cover types. When looking at the LST images, the green areas mainly represent the coldspots with the lowest LST values. These green areas comprise 54.41% of the total area of Istanbul province. Residential and industrial buildings are the areas with the highest values. However, this study found that non-cultivated agricultural lands were also at high temperatures in the LST temperature maps of 2009. The emission of soil surfaces in the thermal infrared region is determined by their water content. Surfaces with a higher water content will cool down due to evaporation, thus reducing their radiation rates. Dry soils emit more radiation as they absorb solar energy and get warmer. Yang et al. [80] specifically selected dry seasons in their study and found that soil surfaces had almost the same LST values as residential areas. In this study, LST results showed that residential areas had the highest LST values, while the lowest LST values were observed in the regions with concentrated green areas. In the regions with green areas, the mean LST values increased from 302.40 K in 2009 to 303.56 K in 2017. In contrast, the mean LST values in regions corresponding to settlements increased from 308.18 K in 2009 to 311.09 K in 2017. Based on the results, it was determined that the mean LST value has increased for all districts between 2009 and 2017. For instance, Bektaş Balçik’s (2014) study [15] in Istanbul showed that high LST values were observed in districts such as Tuzla, Kartal, Kağıthane, Zeytinburnu, and Esenler, where industrial areas are concentrated. The study also found that for the year 2009, the LST values for Atatürk Airport and Sabiha Gökçen Airport and the surrounding areas were higher than 312 K and 314 K, respectively. In 2017, Istanbul Airport and its surrounding area, initially a rural region in 2009, transformed into an impervious surface after construction. Despite having a slightly high LST in the 2017 LST image, hotspot analysis showed that this area shifted from coldspot to not significant. This is due to the influence of prevailing winds and the sea. The transformation of the rural area into an impervious surface did not result in a hotspot designation for the LST value, as expected, due to the cooling effect of the wind and sea. Additionally, our research showed that urban areas’ usage of impervious surfaces and construction materials retains heat more than rural regions, which raises the mean LST of Bağcılar, Bayrampaşa, Gaziosmanpaşa, Esenler, and Esenyurt districts significantly. These results are consistent with the outputs obtained within the scope of the study.

NDVI and UI were calculated from three images in 2009 and three images in 2017. These indices were associated with LST and used in hotspot analyses. The images were averaged over the years and standardized to eliminate atmospheric and temporal effects. Upon examining the obtained spectral index results, it was found that there were consistent results between UI and NDVI values. According to the literature, in areas where NDVI values used to determine green areas are high, UI values used to determine artificial surfaces are low [81]. Conversely, UI values are high in areas where NDVI values are low. Concerning the correlation analysis between NDVI, UI index images, and LST images, the correlation coefficients were consistent with those in the study conducted by Ranagalage et al. [53]. They found the coefficient of determination (R²) between NDVI–LST obtained from Landsat missions from two different years to be 0.77 and 0.88, respectively, while they were 0.80 and 0.91 between NDBI–LST.

Surface temperature mitigation can be greatly aided by vegetation cover, which has a negative correlation with SUHI variations. Over the past eight years, the negative correlation between LST and NDVI has increased slightly, and the positive correlation between LST and UI has increased slightly as well. This shows that while vegetation within the study area contributes to a decrease in the effect of LST, the residential area strengthens the effect of LST and SUHI. When the NDVI values between two years were analyzed based on the NDVI values in districts of Istanbul for two selected years, it was determined that the NDVI values decreased for all districts. This is a very effective reason for the high LST values. Kuru [82] investigated the relationship between urban morphology metrics and LST at four selected subprovinces in İstanbul using Landsat OLI and TIR images. To determine whether urban metrics are more connected than others in certain neighborhood sizes, thirty-eight urban metrics were examined. Their findings demonstrated that, with reference to the size of the urban unit, the urban morphological metrics exhibited differing degrees of strong connections with LST values. These results are consistent with the outputs obtained within the scope of the study.

Overall, the results of this study can serve as a foundation for decision-makers to strive toward more efficient urban planning for Istanbul and reduce the negative impacts of SUHI on the environment and human population. As a limitation, this study used three image sets of Landsat 5 and 8 from 2009 and 2017, respectively, which cannot represent the whole year. However, Istanbul is covered by seas and has a transient climate between the Black Sea and the Mediterranean region. Thus, it is almost impossible to find cloud-free images for the whole study area for the same months in different years. Despite this limitation, this study will be a basis for sustainable cities and will shed light on future research on city dynamics, including vegetation and urban areas, and SUHI analysis. On the other hand, the strengths of this research include: (i) using mean LST, UI, and NDVI images derived from three images per year to accurately represent each year; (ii) normalizing all mean images to reduce the temporal variability in atmospheric conditions; and (iii) identifying hotspots and coldspots of LST, UI, and NDVI at a neighborhood scale, which previous studies have not addressed. Furthermore, the design of the research can be improved or directly applied to any city to analyze its spatiotemporal dynamics.

6. Conclusions

This study presents NDVI, UI, and LST analysis to examine spatiotemporal changes in vegetation, urbanization, and SUHIs in Istanbul, and to investigate the effect of the vegetation and urbanization of SUHI effect. The spatiotemporal changes were analyzed using mean satellite images and the Getis-Ord Gi* statistic. In the first stage, mean NDVI, UI, and LST analyses were carried out separately, and vulnerability maps for each variable were generated for 2009 and 2017. Except for two districts, Avcılar and Büyükçekmece, mean NDVI values decreased drastically in all other 37 districts in Istanbul between 2009 and 2017. Considering the UI results, many districts were faced with high urbanization rates from 2009 to 2017, especially the northern part of İstanbul where Istanbul airport and new road structures are constructed with the Yavuz Sultan Selim Bridge.

In the second stage of the study, hotspot analysis using the Getis-Ord Gi* statistical method was performed on normalized NDVI, UI, and LST images to analyze the effects of SUHIs using spatial statistical methods. In order to statistically map hot- and coldspot areas, firstly, the raster data were converted to vectors, then the average pixel value was assigned for each administrative border (neighborhood border), and hotspot area analysis was performed for a total of 977 administrative regions. According to the hotspot area analysis performed on NDVI, it was observed that 9021.7 ha of area transformed from the “hot areas” class representing green areas in NDVI images to the “not significant” class, while 23,998 ha of area turned into hot areas from not significant clustering. According to the hotspot area analysis carried out on UI, it was seen that 15,455 ha of the area changed from the “cold areas” class to the “not significant” class, and a new “hot area” class was formed in an area of 3202 ha. In the hotspot area analysis on LST, 3148 ha of the area belonged to the “cold areas” class. While it changed from the “cold areas” class to the “not significant” class, it was observed that a new “hot area” class was formed over an area of 10,032 ha. The results of hot area analyses revealed that SUHIs are clustered in areas where residential areas are increasing in Istanbul. Industrial areas, buildings, roads, and other construction activities in these regions contribute to the formation of the SUHI phenomenon in Istanbul. For future studies, long-term time-series analysis of the spectral indices and LST will be conducted to determine the changing trends, and mitigation strategies will be applied and their effects investigated in another study.