Introduction

Urban areas are comprised of different types of land-cover each with their own thermal properties. The difference between temperatures of various land-covers forms the Urban Heat Island (UHI) (Rinner and Hussain 2011). The main focus is particularly on the temperature of urban areas being higher compared to the surrounding rural areas. Numerous studies on urban climate have shown that the most effective factor on UHI is the land-cover change especially vegetation and bare soil conversion to concrete, asphalt and other man-made structures (Rinner and Hussain 2011).

In global scale, human activities especially those related to burning fossil fuels increase the amount of carbon dioxide which in turn may increase the temperature. However, in local scale of an urban area, the main factors leading to temperature increase due to the changes in land-cover are: (1) physical characteristic changes of the surface (e.g. albedo and thermal capacity), (2) the decrease of surface moisture available for evapotranspiration and (3) change of the radiative fluxes and near surface heat flow caused by geometric features of the city surface (Dousset and Gourmelon 2003). Each city has its own particular environment and specific land-cover types making UHI react differently in each individual city. Depending on the materials forming the urban core and its surroundings, the influence and magnitude of UHI differ during the day. Many studies indicate that UHI intensity during night hours is more than that of during day-time (Wilsona et al. 2003; Imhoff et al. 2010). Seasons also may impact the UHI as Liu and Zhang indicated that UHI is more intense in mid latitude regions in summer. They also claimed that long days in summer in tropical climates may increase the UHI intensity (Liu and Zhang 2011). According to Rasul et al. (2015), apart from temperate and sub-tropical climates in which differences between the temperature of cities and their surrounding areas can manifest itself as the Urban Heat Island, in more arid areas, these differences can cause Urban Cool Islands (Liu and Zhang 2011). According to Rasul et al. (2015).

The temperature rise in an urban area leads to some negative impacts on local weather. UHI causes variation in the local wind pattern and has an impact on the development of clouds and the amount of evaporation (Liu and Zhang 2011). High temperature can cause some illnesses such as heat exhaustion, fainting and heat rash (Rinner and Hussain 2011). UHI has also some economic effects by increasing the energy demand for cooling in hot seasons (Hu and Gensuo 2010). Therefore, understanding the origins and the consequences of this phenomenon, and finding ways of preventing the intensity and expansion of UHI is an important priority for urban planners and city managers.

Depending on the method of measuring temperature, there are three types of UHI (Fabrizi et al. 2010). The first one is Canopy Layer Heat Island (CLHI). Canopy is the closest layer of air to the surface in cities, extending upwards to approximately the mean building height. The second type is Boundary Layer Heat Island (BLHI) which is above the Canopy layer and may reach to 1 km in thickness. The third type of UHI is Surface Heat Island (SHI) or Surface Urban Heat Island (SUHI). Meteorological stations can provide precise measurements for analyzing urban heat islands. However, the use of accurate data acquired by thermometers in ground stations has some spatial limitations especially because of relatively insufficient number of stations, their distribution and inhomogeneity (Hamdi 2010). But the advent of remote sensing technology has made it possible to study LST with relatively high spatial resolution. It is also capable of showing the relationship between SUHI and land-cover changes. In this study, satellite images have been used for LST estimation and SUHI assessment. It should be mentioned that the results of analyzing SUHI may be different than CLHI and BLHI.

Monitoring land-cover changes as an important step in SUHI studies can benefit from using a variety of sensors with different resolutions and capabilities in urban temperature surveying. For instance, Dousset and Gourmelon combined the land cover information obtained from SPOT HRV satellite images with the LST retrieved from AVHRR to establish a relationship between different land covers and LST in urban areas (Dousset and Gourmelon 2003). Pu et al. worked with ASTER and MODIS images to measure LST in Yokohama City. They classified the region into six classes (i.e. water, forest, grass, agriculture, bare soil and built up) using visible and NIR channels of ASTER and discussed the relationship between NDVI and LST. They also investigated the differences between MODIS and ASTER images for studying SUHI (Pu et al. 2006). Zhu et al. used aerial hyperspectral image of OMIS1 which has 128 bands in the range of 0.4–12.5 µm in order to find the relationship between five different land cover classes and the LST. They defined a special index for each class based on their spectral signatures and created an optimum regression model between LST and spectral index of each class (Zhu et al. 2006).

The main objective of this study is to find the variations of SUHI for cities which are surrounded by different land-cover types. In order to achieve this, land surface temperature of Paris, Riyadh and Manama (taken during different times of the day) are analyzed using AVHRR satellite images. High temporal resolution of AVHRR images helps to measure LST in different hours. By comparing the resulting information, it is possible to assess how SUHI changes during the day in different cities with regards to their climate and land-covers. Also, to assess the seasonal changes of SUHI, LST maps of Paris, Riyadh and Manama are examined in summer and winter. In addition to land cover types, other parameters such as wind condition, altitude and size of the city, density of population, compactness and height of buildings as well as their shapes and materials may affect UHI. Particularly wind velocity has a greater impact on air temperature than LST. However, this study is focused on the role of land cover types as the main feature in forming UHI and SUHI in different regions.

Dataset and Study Area

AVHRR

NOAA program has at least two satellites in orbit at all times, with one satellite crossing the equator in early morning and early evening and the other one crossing the equator in the afternoon and late evening (Robel 2009). This high temporal resolution of AVHRR makes it possible to study SUHI variations during the day. The use of band 4 (11 µm) and band 5 (12 µm) with the sensor’s spatial resolution of 1.1 km enables us to perform precise atmospheric corrections. There are several methods for retrieving land surface emissivity and atmospheric correction using AVHRR images. Although sensors like AVHRR and MODIS have a coarse resolution in their thermal bands, their high imaging frequency helps studying patterns of thermal anomalies related to surface properties (Pu et al. 2006). MODIS and AVHRR have almost the same spatial resolution in their thermal bands. They both collect data in adequate number of thermal wavelengths which makes it possible to retrieve LST precisely. But the temporal resolution of AVHRR is even better than MODIS. As mentioned above, with at least two satellites in orbit at all times, there are at least four AVHRR images collected in different times of the day. But MODIS instruments on boards Terra and Aqua image the entire earth every 1–2 days.

TM, ETM + and OLI

TM, ETM + , and OLI sensors on Landsat5, Landsat7 and Landsat8 respectively, have been quite popular in environmental studies, such as research on coastal waters, crop identification, vegetation index calculation, detection of clouds, ice, and snow, and also calculations for LST retrievals (Sobrino et al. 2004). TM has six spectral bands in visible, near, and shortwave infrared as well as one single band in thermal infrared region. ETM + offers several improvements on TM, i.e. the addition of panchromatic band, improved spatial resolution of the thermal band and also reduced noise (Masek et al. 2001). However, on 31 May 2003, the ETM + Scan Line Corrector (SLC) failed, causing scan to scan gaps which are mostly noticeable along the edges of the scene. The Landsat8 satellite is equipped with two earth observing sensors: the Operational Land Imager (OLI) and the Thermal Infrared Sensor (TIRS). OLI measures in visible, near infrared, and shortwave infrared wavelengths, while TIRS measures land surface temperatures in two thermal bands (Garner 2016). Although the spatial resolution of Landsat images is adequate for many remote sensing studies, the low temporal resolution limits the instrument’s capability of monitoring fast changing phenomena such as temperature. This problem has made the data unsuitable for the purpose of LST retrievals in this study. However, these sensors have been valuable in finding the relation between land-cover types and LST, through classification of the study area into the main classes of vegetation, urban, bare soil, and water body.

Study Areas

The main goal of this research is to map the SUHI in cities with different land-cover types. The cities chosen as case studies are Paris, Riyadh and Manama, all located in different climates and surrounded by different land cover types. Comparing the changes in land surface temperature between these cities will help to understand the behavior of SUHI phenomenon in urban areas with different land-cover types.

Paris (2.20°E, 48.50°N) has a typical western European oceanic climate which is affected by the North Atlantic current, resulting in a generally mild and wet weather. The Seine River runs through the city and the dominant type of land cover is vegetation. Paris has a dense urban area and a population of about 9 millions.

Riyadh (46.73°E, 24.62°N) is located in the center of Arabian Peninsula and is the capital and largest city of Saudi Arabia. Riyadh has a hot desert climate especially during the summer months. The city receives very little rainfall throughout the year and is largely surrounded by desert, which makes it an ideal place for dust storms.

Manama (50.34°E, 26.12°N) is the capital and largest city of Bahrain, an island surrounded by the Persian Gulf. It has an arid climate with very high temperatures in summer. Table 1 shows the annual climate variables of Paris, Riyadh and Manama (World Climate & Temperature 2016).

Table 1 Annual climate variables of Paris, Riyadh and Manama
Full size table

Methodology

Preprocessing

During each scan line, in addition to targets on earth, AVHRR also views the cold deep space behind it as well as its own internal blackbody target. Using these two extra views, calibration coefficients \( S_{i} \) nd \( I_{i} \) re calculated for each band. The top-of-atmosphere (TOA) radiance can then be determined by the AVHRR sensor as a linear function of the input digital numbers (ranging from 0 to 1023) as follows:

$$ E_{i} = S_{i} C + I_{i} $$
(1)

where \( E_{i} \) is the TOA radiance value in mW/(m2-sr-cm−1), \( C \) is the input pixel value, and \( S_{i} \) and \( I_{i} \) are the scaled slope and intercept values respectively for band i (Trishchenko and Li 2001).

The next step in preprocessing of the data is cloud masking. Cloudy pixels have an immense effect on the results, and therefore it is vital to find an accurate method to locate those cloudy pixels. Based on the characteristics of satellite images, different methods are used for cloud detection. For example, supervised and unsupervised classification are accepted approaches when working with Landsat or ASTER images, but for AVHRR data, the most common methods of image cloud detection are the decision-tree threshold-based algorithms (Ghosh et al. 2012). These algorithms are able to find even the pixels that are partially cloudy. They can also detect the different types of clouds (Ghosh et al. 2012; Poli et al. 2010). In this study, three properties associated with clouds are used to detect clear pixels. The first property is the extreme reflection of clouds. If a pixel’s albedo derived from channel 2 exceeds a threshold of 25%, it will be regarded as a cloud pixel. This condition is only valid for daytime images. The second condition is based on very low temperatures of clouds. Any pixel with a brightness temperature (calculated from channel 4) less than a specified threshold will not be considered as a clear pixel. Depending on the season and hour of imaging, this threshold may vary. The final property is the very high variance of clouds. In a window of 9 pixels, if the difference between maximum and minimum of temperature is more than 3°, the center pixel will be considered as cloudy. If at least one of the above conditions is found to be true, the pixel will be flagged as a cloud pixel.

The final step in preprocessing is the geometric correction of AVHRR and Landsat images for each case study. The result is images that are transformed to the same UTM projection system and WGS84 datum. Since Landsat images have a higher spatial resolution than AVHRR, and in order to avoid losing information, the AVHRR images have been resized to 30 meters using the Nearest Neighbor resampling method.

LST Retrieval

The radiance which was calculated during the preprocessing of AVHRR data is now converted to brightness temperature using the inverse of Planck’s equation:

$$ T = \frac{{C_{2} v}}{{In\left( {\frac{{C_{1} v^{3} }}{E} + 1} \right)}} $$
(2)

where T is the temperature (K) for radiance value E, v is the central wave number of each channel (cm−1), \( C_{1} = 1.1910659 \times 10^{ - 5} \) mW/(m2 sr cm−1) and \( C_{2} = 1.438833 \times 10^{ - 5} \) cm K (Kidwell 2008).

Atmospheric attenuation due to absorption and emission leads to reduced correlation between brightness temperature and land surface temperature. Depending on the amount of water vapor and other gases such as CO2, the level of absorption may change (Poxo Vhquex et al.1997). The main approach to atmospheric correction is to use a radiative transfer model which needs accurate data on the characteristics of atmospheric structure. However, because of spatial and temporal variability of atmospheric conditions, lack of such information may introduce large errors (Kerr, et al. 2005). An alternative for radiative transfer models is to use the Split-Window technique. Sensors like AVHRR which have at least two thermal bands make it possible to perform atmospheric correction by estimating differential absorption of their channels. The main equation for this method is:

$$ T = a_{0} + \sum a_{i} T_{i} $$
(3)

where \( T \) is the LST and \( T_{i} \) are the brightness temperatures of two close thermal bands (Poxo Vhquex et al. 1997). Based on the Eq. (3), several algorithms have been developed for atmospheric correction using the Split-Window technique (Prata and Platt 1991; Price 1984) (Caselles et al. 1997) (Sobrino and Raissouni 2000). Each of these methods may depend on spectral emissivity, view angel, water vapor content or a combination of these factors (Kerr et al. 2005). In this study, the global algorithm for atmospheric correction proposed by Sobrino and Raissouni (Sobrino and Raissouni 2000) is used:

$$ T_{s} = T_{4} + c_{1} \left( {T_{4} - T_{5} } \right) + c_{2} \left( {T_{4} - T_{5} } \right)^{2} + c_{0} + \left( {c_{3} + c_{4} W} \right)\left( {1 - \varepsilon } \right) + \left( {c_{5} + c_{6} W} \right)\Delta \varepsilon $$
(4)

where \( T_{s} \) is the surface temperature, \( T_{4} \) and \( T_{5} \) are the brightness temperatures retrieved from AVHRR band 4 and 5 respectively, \( \varepsilon_{4} \) and \( \varepsilon_{5} \) are emissivities estimated for band 4 and 5, and \( W \) is the total amount of the columnar atmospheric water vapor (g/cm2). \( W \) can be retrieved from local meteorological stations or from satellite images. For the calculation in this study, the Total Precipitable Water Product from the moderate-resolution imaging spectroradiometer (MODIS) is used, which combines the results from both the NIR (\( 1 \times 1 \) km) and the IR (\( 5 \times 5 \) km) algorithms (Hubanks 2015). The mean effective emissivity is \( \varepsilon = \left( {\varepsilon_{4} + \varepsilon_{5} } \right)/2 \) and the emissivity difference is \( \Delta \varepsilon = \left( {\varepsilon_{4} - \varepsilon_{5} } \right) \). Algorithm coefficients (\( c_{0} \)–\( c_{6} \)) can be obtained using a radiative code program. Lahraoua et al. (2013) determined these coefficients for the AVHRR sensor on board NOAA-7, 9, 11, 12, 14, 15, 16, 17, 18, and 19 satellites using MODTRAN 4.0 radiative code simulations. The total error when using this method is about 1.3 (K) (Lahraoua et al. 2013).

The Split-Window method is mainly used in Sea Surface Temperature (SST) calculations, but for LST retrieval, it brings more uncertainties (Kerr et al. 2005) (Lahraoua et al. 2013) due to variations in land surface emissivity. This requires the surface emissivity to be included in the Split-Window equation. There are two main methods for deriving surface emissivity from space (Kerr, et al. 2005). The first one, which needs more than one thermal band to function, calculates emissivity by solving the radiometric equation at the surface in the thermal infrared region. An example of this approach was proposed by Becker and Li (1990a). They used the properties of AVHRR channel 3 (3.7 μm) in day/night image pairs to measure emissivity. Day-time channel 3 radiance is a combination of emitted radiance by the surface and reflected radiance due to sun illumination. The night-time radiance, however, is solely the emitted radiance by the surface. Therefore, the use of day-time and night-time images makes it possible to estimate the contribution of the sun’s reflected radiance and hence the emissivity (Becker and Li 1990).The second method is based on a relationship between emissivity and near infrared bands (0.5–1.1 µm) (Sobrino et al. 2008) (Valor and Caselles 1996). This approach is almost empirical and its coefficients are surface dependent (Kerr et al. 2005). The method used in this study was proposed by Cihlar et al. (1997) based on the relationship between NDVI and emissivity for AVHRR data (Cihlar et al. 1997):

$$ \varepsilon_{4} = 0.9897 + 0.029{ \ln }\left( {\text{NDVI}} \right) $$
$$ \varepsilon_{4} - \varepsilon_{5} = 0.01019 + 0.1344{ \ln }\left( {\text{NDVI}} \right) $$
(5)

where \( \varepsilon_{4} \) and \( \varepsilon_{5} \) are emissivities estimated for bands 4 and 5 respectively.

Implementation and Analysis

Diurnal Changes of SUHI in Cities with Different Types of Land-Cover

In order to assess diurnal variations of SUHI in different cities, land surface temperature maps of Paris, Riyadh, and Manama are generated using AVHRR images at night and day time. The land-cover map of each city is obtained from one of the TM, ETM + or OLI images using Maximum Likelihood classification. Visible and near infrared wavelengths (R, G, B, NIR bands) of each sensor are used in classification process. The next step examines the role of each major land-cover type, namely, urban, bare soil, vegetation, and water body. Table 2 shows the overall accuracy and kappa coefficients of the classification for each city.

Table 2 Overall accuracy and Kappa coefficients for land-cover classifications
Full size table

Figure 1a shows the Landsat TM bands 4-3-2 false color combination of Paris. Figure 1b is the land-cover classification of Fig. 1a into urban, vegetation, bare soil, and water body classes. Figures 1c and d are the day-time and night-time LSTs respectively retrieved from AVHRR images.

Fig. 1
figure 1

Diurnal changes of SUHI in Paris, a Landsat5_TM_432 false color combination image of 6/4/2010, b classification of image (a), c LST (K) retrieved from AVHRR, Time (yyyy ddd hh:mm): 2011 154 10:52, d LST (K) retrieved from AVHRR, Time (yyyy ddd hh:mm): 2011 152 20:11

Full size image

As Fig. 1 shows, the urban areas in Paris are warmer than the surrounding vegetation in both day and night, but the SUHI is more significant at night-time.

Figures 2a and b show the false color combination and classification of ETM + satellite image for Riyadh on 10 June 2012. In order to correct the effect of ETM + SLC off, three other ETM + images on June and July 2012 were used to fill the gaps; Although still some scan gaps remained in the upper-left corner of the image.

Fig. 2
figure 2

Diurnal changes of SUHI in Riyadh, a Landsat7_ETM + _432 false color combination image of July 2012, b classification of image (a), c LST (K) retrieved from AVHRR, Time (yyyy ddd hh:mm): 2012 219 12:47, d LST (K) retrieved from AVHRR, Time (yyyy ddd hh:mm): 2012 218 01:31

Full size image

As seen from Fig. 2, Riyadh is completely surrounded by desert. Figures 2c and d are the day-time and night-time land surface temperature retrieved from AVHRR data. Similar to Paris, the urban areas in Riyadh are warmer than the surrounding regions at night, but during the day, the city is cooler than its surroundings.

Figure 3 shows the diurnal change of SUHI in Manama. Unlike Riyadh, in this city, urban area is warmer than its surrounded area during the day hours but not at night hours.

Fig. 3
figure 3

Diurnal changes of SUHI in Manama, a Landsat8_OLI_543 false color combination image of 17/09/2016, b classification of image (a), c LST (K) retrieved from AVHRR, Time (yyyy ddd hh:mm): 2016 261 14:37, d LST (K) retrieved from AVHRR, Time (yyyy ddd hh:mm): 2016 257 02:14

Full size image

Another point that should be considered is that the regions around the urban area of Manama which are classified as bare soil, are cooler than urban area in Fig. 3d. So if Manama was completely surrounded by bare soil, its SUHI would unfold at night hours, instead of day-time. Same result is obtained for Riyadh, where the small vegetation area in the south of the city is always cooler than the urban area. So if Riyadh was completely surrounded by vegetation, its SUHI would be apparent at both day and night hours. All the above figures point to the conclusion that the main parameter affecting the state of an SUHI is the land-cover types around the city. Depending on the type of the main materials covering the city, the surface temperature in urban areas may be warmer than their surroundings in specific times of the day.

Seasonal Changes of SUHI

To examine the seasonal variations of SUHI, LST maps of Riyadh, Paris, and Manama are compared in summer and winter using AVHRR data. Figure 4a shows the mean LST map of Riyadh comprised of 11 LST maps retrieved in a warm month (June, 2011). Figure 4b shows the mean LST map of Riyadh comprised of 13 LST maps retrieved in a cold month (January, 2011). It should be noted that in all of the images, Riyadh is located near the center of the image scene. Also, only non-cloudy and non-dusty pixels of the night-time images were selected.

Fig. 4
figure 4

SUHI of Riyadh in a warm and a cold season: a the mean image of 11 LST maps (K) in June 2011 and b the mean image of 13 LST maps (K) in January 2012, all retrieved from AVHRR night-time data

Full size image

As Fig. 4 shows, there is no significant difference in the shape or the extent of the SUHI in Riyadh during winter and summer. The average LST discrepancy between the urban class and its surrounding bare soil class is about 3.7 (K) in summer and 3.5 (K) in winter.

Similar results were obtained when assessing the heat islands of Manama and Paris. As Table 3 shows, the differences between the LST of urban areas and that of surrounding classes (bare soil, vegetation and water body classes) in summer and winter are insignificant. Also according to Table 3, summer temperature difference for Riyadh is higher than that of the winter but for Paris and Manama it is lower. So it can be concluded that the intensity of SUHI does not systematically change during the seasons.

Table 3 SUHI intensity of warm and cold seasons
Full size table

Relationship Between SUHI and NDVI: a Case Study of Paris

As indicated in previous sections, the temperature of the vegetation class is lower than urban areas. Figure 5a and b show the negative relationship between LST and NDVI for Paris during day and night.

Fig. 5
figure 5

Scatter plot of LST and NDVI retrieved from a day-time and b night-time AVHRR satellite images of Paris

Full size image

According to Fig. 5, as the NDVI increases, in both day-time and night-time, the LST decreases. But the slope of the fitted line in the day-time image is almost twice that of the night-time. This is because the temperature anamoly during the day-time is more that of during night hours. It can be seen in Figs. 1, 2 and 3, that the temperature range of day time images are more than nigh time ones. Also, as shown in Figs. 1, 2 and 3, in the areas where dense vegetation is present, the temperature during the day is less than other areas with sparse vegetation.

Conclusion

In this study, surface temperatures in three different cities were assessed and linked to land-cover using a combination of AVHRR and Landsat imagery data. The results showed that depending on the land-cover types that surround a city, the appearance of its surface urban heat island varies in different times of the day. The main objective of this study is to show that SUHI state might be completely different in cities which have different land-cover classes. Sometimes SUHI appears during the day, sometimes at night and sometimes it can be seen throughout the day. In Paris, which is mainly surrounded by vegetation, the urban areas are always warmer than the outer regions. In Manama, which is enclosed by water, the urban areas are cooler than their surrounding water body at night, and the SUHI phenomenon only appears during the day. On the contrary, for a desert city like Riyadh, the SUHI only appears during night hours.

Table 4 indicates the major land-cover around Paris, Riyadh, and Manama along with the time of forming SUHI in each city. These results lead to the conclusion that thermal inertia, which is different for each land-cover type, is the main factor that specifies when SUHI will appear. For example, thermal inertia of water body is higher than urban structures; therefore, as the sun rises, the temperature of water takes more time to increase compared to urban areas. Also, after the sunset, the water temperature decreases more slowly than that of urban structures. This leads to SUHI appearing only during the day in Manama. On the other hand, thermal inertia of bare soil is lower than urban class. In the morning as the sun starts to shine, temperature of bare soil increases more quickly than urban structures. And as the sun sets in the evening, the soil losses it’s temperature faster. As a result, SUHI in Riyadh appears only at night. Eventually by assessing the temperature of different classes of each city in different hours, it could be realized that if the main land-cover class surrounded the city was changed, SUHI of that city would be completely different. For instance the main class around Manama is water body and its SUHI appears at day time. But if Manama was surrounded by bare soil, then its SUHI would be unfolded at night hours. As well if Manama was surrounded by vegetation, its SUHI would be evident at all day hours.

Table 4 Major land-cover and the appearance of SUHI in Paris, Riyadh and Manama
Full size table

Furthermore, in order to examine the seasonal changes in SUHI, the LST of Paris, Riyadh, and Manama in warm and cold months were assessed. The results, however, did not reveal any significant differences in the SUHI intensity during different seasons. The assessment also showed that in four classes of urban area, vegetation, bare soil, and water body, the temperature discrepancy is less during night-time than that of day-time.