**3.1. Test area and data**

The test area is in the city of Heraklion, Greece, and it covers an area of about 90 km<sup>2</sup> . It is a typical Mediterranean city, characterized by mixed land-use patterns that include residential, commercial and industrial areas, transportation networks and rural areas. **Figure 4** shows the Urban Atlas land use map of Heraklion [19]. Apart from the Heraklion city core, the rest of the study area is featuring mixed urban and agricultural land cover pattern, mainly olive trees and vineyards.

Fourteen cloud-free Landsat 8 Surface Reflectance Climate Data Record (CDR) covering the test area corresponding to 1 year (April 1, 2013–April 30, 2014) were used in this study. The Landsat 8 CDR is a higher level data product of surface reflectance information for six bands,

Earth Observation for Urban Climate Monitoring: Surface Cover and Land Surface Temperature http://dx.doi.org/10.5772/intechopen.71986 135

**Figure 4.** Urban Atlas land use polygons of the test site, the city of Heraklion, Greece.

generated from Landsat Ecosystem Disturbance Adaptive Processing System (LEDAPS) distributed by the U.S. Geological Survey [20].

Daily MODIS Level 1B (MOD021) data from both Terra and Aqua satellites for a period between April 1, 2013 and April 30, 2013 were acquired. The daily MODIS water vapor product (MOD05) was also used to provide ancillary atmospheric information on water vapor and cloud cover.

#### **3.2. Urban surface cover mapping**

The Sentinel-2 mission provides continuity to services relying on multispectral high-spatialresolution optical observations (like Landsat and SPOT satellites): it carries a Multispectral Instrument (MSI) covering the electromagnetic spectrum from the visible to the shortwave infrared with a pixel resolution from 10 to 60 m. Two satellites in orbit will provide data at a 5 days interval at the equator. Sentinel-2 combines a large swath, frequent revisit, and systematic acquisition of all land surfaces at high-spatial resolution and with a large number of spectral bands [17]. The pair of Sentinel-2 satellites routinely delivers high-resolution optical

Sentinel-3 mission represents the continuity of the ENVISAT sensors, i.e., MERIS (MEdium Resolution Imaging Spectrometer) and Advanced Along Track Scanning Radiometer (AATSR). In particular, the Sea and Land Surface Temperature Radiometer (SLSTR) will provide TIR data at 1 km resolution with daily revisit at the equator [18]. Among the Sentinel-3 mission objectives is to monitor the land surface temperature with high-end accuracy and

This chapter examines the exploitation of EO data for monitoring the urban climate, with particular focus on satellite data. The mapping of the urban surface and its characteristics, using spectral unmixing is examined in the first part. A method adjusted for urban studies is proposed which accounts for the non-linear mixture of spectral radiances in the urban canyon. The second part focuses on thermal EO data. To capture the intra-urban variations of temperature, EO data of high spatial and temporal resolution are necessary, but no current of forthcoming satellite provides them. Moreover, only one from the series of Sentinel satellites carries a thermal sensor of low spatial resolution. To overcome the limitations of the resolution trade-off, a synergistic methodology between high resolution optical and low resolution thermal satellite measurements with ultimate goal daily local-scale land surface temperature

The test area is in the city of Heraklion, Greece, and it covers an area of about 90 km<sup>2</sup>

typical Mediterranean city, characterized by mixed land-use patterns that include residential, commercial and industrial areas, transportation networks and rural areas. **Figure 4** shows the Urban Atlas land use map of Heraklion [19]. Apart from the Heraklion city core, the rest of the study area is featuring mixed urban and agricultural land cover pattern, mainly olive

Fourteen cloud-free Landsat 8 Surface Reflectance Climate Data Record (CDR) covering the test area corresponding to 1 year (April 1, 2013–April 30, 2014) were used in this study. The Landsat 8 CDR is a higher level data product of surface reflectance information for six bands,

. It is a

images globally, providing enhanced continuity of SPOT and Landsat type data.

**3. Application for surface cover and land surface temperature**

reliability in support of climate monitoring.

134 Multi-purposeful Application of Geospatial Data

estimates.

**3.1. Test area and data**

trees and vineyards.

The sub-pixel land cover information was estimated using spectral unmixing with a neural network, trained using endmember spectra collected from the image and synthetic spectra [21]. The methodology applied to produce the synthetic spectra and to estimate the surface cover fractions is briefly described below.

The urban surface is assumed to be composed of four land cover types: built-up surface, vegetation, non-urban bare surfaces, and water bodies. Using Landsat imagery, it is not easy to discriminate between different materials, but rather surface cover types, due to the medium spatial and low spectral resolution [22]. Thus, a redundant two-level hierarchy is assumed, as shown in **Figure 5**, including the four main surface cover types in the first level and a more detailed one in the second level, serving the endmembers collection.

where *ρ<sup>i</sup>*

produced.

model [25]:

where the ∑*<sup>j</sup>*=1

is the observed spectrum of pixel *i*, *ρ<sup>j</sup>*

**Three endmember Two endmember**

BU + VE + NB BU + BU BU + BU + VE BU + VE BU + BU + NB BU + NB BU + BU + BU VE + NB

(*i*) represent the areal fraction of the endmembers *ρ<sup>j</sup>*

in **Table 1** and assuming combinations of different mixture levels *aj*

*j*=1 *M aj*

linear interactions in the urban structure. The abundance coefficients *aj*

accounts for the linear mixing, while ∑*<sup>j</sup>*=1

model representing the spectrum *ρ<sup>i</sup>*

BU, Built-up; VE, Vegetation; NB, Non-urban bare.

**Table 1.** Mixture models assumed for spectral synthesis.

*ρ<sup>i</sup>* = ∑

models in **Table 1**, with randomly generated *b*

of values in the interval [0,1], matching the coefficients *aj*

*<sup>M</sup> aj* (*i*) <sup>∙</sup> *<sup>ρ</sup><sup>j</sup>*

fraction of the endmember *ρ<sup>j</sup>*

for the Landsat scenes.

*<sup>j</sup>*, and *M* is the number of endmembers in the mixture model. The abundance coefficients *aj*

Earth Observation for Urban Climate Monitoring: Surface Cover and Land Surface Temperature

that the surface corresponding to a pixel is flat and uniformly irradiated. This simple

rather popular [23]. Using the endmember spectra for each cover type, for all the models

Although many studies assume linear mixing effects, it has been known for some time that non-linear spectral mixing effects can be a crucial component in many real-world scenarios, including the urban scenes [24]. The buildings and the street canyons create a complicated 3D structure in the cities at a meter scale, which induces multiple scattering of light between surfaces. Recently, a physics-based approach of a spectral mixture model suitable for urban scenes was presented and validated against simulated urban scenes using a quadratic mixing

> (*i*) ∙ *ρ<sup>j</sup>* + ∑ *j*=1 *M* ∑ *l*=*j M bj*,*l*

> > *j*,*l*

A three-layer feed-forward neural network with a sigmoid activation function was used in this study. The input layer has six neurons, one for each Landsat band and four output neurons, one for each of the land cover types of the first level (**Figure 5**) and a hidden layer with 12 neurons. The number of hidden neurons is set to 12, because the network performance would not improve much by adding more neurons. The network is trained using a Levenberg-Marquardt backpropagation algorithm and endmember and synthetic spectra from the developed spectral library. The network training data output is set to a combination

Once trained, the network is then applied to estimate the surface cover fractions individually

is the representative spectrum of endmember

as a linear combination of the endmember spectra *ρ<sup>j</sup>*

*<sup>M</sup>* <sup>∑</sup>*<sup>l</sup>*=*<sup>j</sup> <sup>M</sup> b j*,*l* (*i*) <sup>∙</sup> *<sup>ρ</sup><sup>j</sup> ρl*

in the pixel *i* and the endmember spectra are used, for all the

(*i*) coefficients to produce synthetic spectra.

in the pixel *i*. The linear model assumes

http://dx.doi.org/10.5772/intechopen.71986

(*i*) ∙ *ρ<sup>j</sup> ρ<sup>l</sup>* (2)

(*i*) used during the spectral synthesis.

accounts for the non-

(*i*) represent the areal

(*i*), mixed spectra *ρ<sup>i</sup>*

is

137

are

**Figure 5.** The two-level hierarchical urban classification scheme used in this study.

Endmembers are collected from the image corresponding to the second level of the hierarchy. This ensures a variety of different endmember spectra in the library, representing different surface cover types, rather than a single category. It has to be noted here that the endmembers collected from medium resolution imagery do not correspond to material spectra, but rather spectra of large homogeneous surfaces. The spectral resolution of Landsat (30 m) is not enough to discriminate between types of the second level of hierarchy. The collected endmembers are then grouped to match the surface cover types of the first level.

The median value of each Landsat pixel surface reflectance for the total six images was estimated. The reason for doing this was to create a single image corresponding to the whole year and to avoid extreme reflectance values. The thermal and panchromatic bands were not included in the analysis. The median reflectance product is an image covering 1 year (April 1, 2013–April 30, 2014). This image is used for the collection of the endmember. The median is selected as a statistic to avoid extreme values, which might cause confusions in the network training and the unmixing process. A number of endmember spectra, corresponding to the cover types of **Figure 5**, is collected from the median reflectance product by visual inspection and using high resolution imagery from Google Earth as reference.

The endmember spectra are then used to produce synthetically mixed training data. These are generated using models corresponding to the first level of the classification scheme (**Figure 5**). Two- and three-endmemeber mixture models are considered, with repeated surface cover types allowed inside a model. **Table 1** shows the models that are considered for the spectral unmixing. No water endmembers are considered in the models and the generation of synthetic spectra, because water is generally dark and highly degenerate in terms of spectral mixture [15]. Both linear and non-linear mixture models are then considered for the generation of synthetic spectra.

The linear mixing model assumed is described by:

$$\rho\_i = \sum\_{j=1}^{M} a\_j'(i) \cdot \rho\_{j'} \,\forall a\_j'(i) \ge 0 \, and \, \sum\_{j=1}^{M} a\_j = 1 \tag{1}$$


**Table 1.** Mixture models assumed for spectral synthesis.

Endmembers are collected from the image corresponding to the second level of the hierarchy. This ensures a variety of different endmember spectra in the library, representing different surface cover types, rather than a single category. It has to be noted here that the endmembers collected from medium resolution imagery do not correspond to material spectra, but rather spectra of large homogeneous surfaces. The spectral resolution of Landsat (30 m) is not enough to discriminate between types of the second level of hierarchy. The collected end-

The median value of each Landsat pixel surface reflectance for the total six images was estimated. The reason for doing this was to create a single image corresponding to the whole year and to avoid extreme reflectance values. The thermal and panchromatic bands were not included in the analysis. The median reflectance product is an image covering 1 year (April 1, 2013–April 30, 2014). This image is used for the collection of the endmember. The median is selected as a statistic to avoid extreme values, which might cause confusions in the network training and the unmixing process. A number of endmember spectra, corresponding to the cover types of **Figure 5**, is collected from the median reflectance product by visual inspection

The endmember spectra are then used to produce synthetically mixed training data. These are generated using models corresponding to the first level of the classification scheme (**Figure 5**). Two- and three-endmemeber mixture models are considered, with repeated surface cover types allowed inside a model. **Table 1** shows the models that are considered for the spectral unmixing. No water endmembers are considered in the models and the generation of synthetic spectra, because water is generally dark and highly degenerate in terms of spectral mixture [15]. Both linear and non-linear mixture models are then considered for the genera-

(*i*) ≥ 0 *and* ∑

*j*=1 *M*

*aj* = 1 (1)

members are then grouped to match the surface cover types of the first level.

**Figure 5.** The two-level hierarchical urban classification scheme used in this study.

136 Multi-purposeful Application of Geospatial Data

and using high resolution imagery from Google Earth as reference.

*j*=1 *M aj* (*i*) ∙ *ρ<sup>j</sup>* , ∀*aj*

tion of synthetic spectra.

*ρ<sup>i</sup>* = ∑

The linear mixing model assumed is described by:

where *ρ<sup>i</sup>* is the observed spectrum of pixel *i*, *ρ<sup>j</sup>* is the representative spectrum of endmember *<sup>j</sup>*, and *M* is the number of endmembers in the mixture model. The abundance coefficients *aj* (*i*) represent the areal fraction of the endmembers *ρ<sup>j</sup>* in the pixel *i*. The linear model assumes that the surface corresponding to a pixel is flat and uniformly irradiated. This simple model representing the spectrum *ρ<sup>i</sup>* as a linear combination of the endmember spectra *ρ<sup>j</sup>* is rather popular [23]. Using the endmember spectra for each cover type, for all the models in **Table 1** and assuming combinations of different mixture levels *aj* (*i*), mixed spectra *ρ<sup>i</sup>* are produced.

Although many studies assume linear mixing effects, it has been known for some time that non-linear spectral mixing effects can be a crucial component in many real-world scenarios, including the urban scenes [24]. The buildings and the street canyons create a complicated 3D structure in the cities at a meter scale, which induces multiple scattering of light between surfaces. Recently, a physics-based approach of a spectral mixture model suitable for urban scenes was presented and validated against simulated urban scenes using a quadratic mixing model [25]:

$$\rho\_i = \sum\_{j=1}^{M} a\_j(\mathbf{i}) \cdot \rho\_j + \sum\_{j=1}^{M} \sum\_{i=1}^{M} b\_{ji}(\mathbf{i}) \cdot \rho\_j \rho\_i \tag{2}$$

where the ∑*<sup>j</sup>*=1 *<sup>M</sup> aj* (*i*) <sup>∙</sup> *<sup>ρ</sup><sup>j</sup>* accounts for the linear mixing, while ∑*<sup>j</sup>*=1 *<sup>M</sup>* <sup>∑</sup>*<sup>l</sup>*=*<sup>j</sup> <sup>M</sup> b j*,*l* (*i*) <sup>∙</sup> *<sup>ρ</sup><sup>j</sup> ρl* accounts for the nonlinear interactions in the urban structure. The abundance coefficients *aj* (*i*) represent the areal fraction of the endmember *ρ<sup>j</sup>* in the pixel *i* and the endmember spectra are used, for all the models in **Table 1**, with randomly generated *b j*,*l* (*i*) coefficients to produce synthetic spectra.

A three-layer feed-forward neural network with a sigmoid activation function was used in this study. The input layer has six neurons, one for each Landsat band and four output neurons, one for each of the land cover types of the first level (**Figure 5**) and a hidden layer with 12 neurons. The number of hidden neurons is set to 12, because the network performance would not improve much by adding more neurons. The network is trained using a Levenberg-Marquardt backpropagation algorithm and endmember and synthetic spectra from the developed spectral library. The network training data output is set to a combination of values in the interval [0,1], matching the coefficients *aj* (*i*) used during the spectral synthesis. Once trained, the network is then applied to estimate the surface cover fractions individually for the Landsat scenes.

The developed spectral library contains 15 endmember spectra, 2 representing buildings/ roofs bright materials, 2 representing buildings/roofs dark materials, 3 transportation areas (roads, parking lots, and airport runways), 1 sport infrastructure, 3 green vegetation, 1 nonphotosynthetic vegetation, 2 bare soil, and 1 rocks. Synthetic spectra were generated from these endmembers using the two- and three-endmember models presented in **Table 1**. In the end, a set of 72,030 synthetic spectra were used to train the neural network. Randomly chosen 70% of the spectral library data were used for training, 15% for validation and 15% for testing the neural network. The topology of the network was determined to be 6-12-3 inputhidden-output neurons and the Levenberg-Marquardt backpropagation algorithm was used to estimate the weights and bias values of the network. Derived estimates were then applied independently in the series of 14 Landsat images, and cover fraction images for the four land cover types assumed in this study were generated.

An example of the resulting surface cover fraction maps for May and August is shown in **Figure 6**. The general pattern shown in **Figure 6** matches the Urban Atlas polygons (**Figure 4**). Moreover, the fraction image corresponding to May reveals more vegetation abundance than the one corresponding to August both in the outskirts as well as in the urban core.

#### **3.3. Urban land surface temperature**

Although LST is routinely derived by satellite TIR observations, currently there is no spaceborne sensor capable of providing frequent thermal imagery at spatial resolution needed in urban studies. Current and forthcoming TIR remote sensing is confronting the trade-off between spatial and temporal resolution. A synergistic method that unmixes the low-resolution TIR measurements using high spatial information on the surface cover for estimating high spatial resolution LST is applied here [26]. The method is a multistep procedure described in detail along with its validation in Ref. [26]. For the method to be applied, information on the surface cover fractions is necessary. Representative emissivity values are then assigned to each of the cover types in **Figure 5**, using information derived from the ASTER Spectral Library [27]. Samples from the library, which are representative for the study area, are selected and convolved with the sensor's spectral response function and the emissivity *ε<sup>k</sup>* (*H*) for each pixel high resolution (*H*) pixel *i* is estimated by:

$$
\varepsilon\_k^{\prime \prime \prime \prime} = \sum\_{l=1}^{n} \varepsilon\_l \cdot a\_{lk}^{\prime \prime \prime \prime} \tag{3}
$$

where *P* is the number of high resolution pixels (*j* ∈ *P*) corresponding to each low resolution one (*k*). Each low resolution pixel is then unmixed, using the contextual information of the

Earth Observation for Urban Climate Monitoring: Surface Cover and Land Surface Temperature

http://dx.doi.org/10.5772/intechopen.71986

139

**Figure 6.** Pseudo color composition of the derived fraction images for the May 13, 2013 (a) and August 22, 2013 (b).

= *A*(*L*) ⏟*<sup>w</sup>*×*<sup>n</sup>* ∙ *E n*⏟×1

contributions of surface types to those pixels, and *E* is the thermal radiances under consider-

is a vector of the thermal radiances of the pixels in the window, *A* (*L*)

ation. A regularization term is also used to prevent large deviations in optimization:

+ *er* (5)

is a matrix of the

⏟*<sup>w</sup>*×1

neighboring pixels in a window (*a* window of size *w*):

*S*(*L*)

where *S* (*L*)

where *n* is the number of surface cover types, *ε<sup>i</sup>* is the representative emissivity value for the surface cover type *i*, and *aik* (*H*) are the estimated fractions of surface cover types.

Spatial-spectral unmixing is then used to enhance the spatial resolution of the low resolution thermal bands. The contribution of the land cover components is estimated for each thermal pixel *k*, by summing the estimated fractions *aj* (*H*) :

*i*

$$A\_{\mathbb{A}\_{\parallel}}^{\quad(\text{L})} = \frac{1}{P} \sum\_{\parallel \in P} a\_{\parallel}^{\quad(\text{H})} \tag{4}$$

Earth Observation for Urban Climate Monitoring: Surface Cover and Land Surface Temperature http://dx.doi.org/10.5772/intechopen.71986 139

The developed spectral library contains 15 endmember spectra, 2 representing buildings/ roofs bright materials, 2 representing buildings/roofs dark materials, 3 transportation areas (roads, parking lots, and airport runways), 1 sport infrastructure, 3 green vegetation, 1 nonphotosynthetic vegetation, 2 bare soil, and 1 rocks. Synthetic spectra were generated from these endmembers using the two- and three-endmember models presented in **Table 1**. In the end, a set of 72,030 synthetic spectra were used to train the neural network. Randomly chosen 70% of the spectral library data were used for training, 15% for validation and 15% for testing the neural network. The topology of the network was determined to be 6-12-3 inputhidden-output neurons and the Levenberg-Marquardt backpropagation algorithm was used to estimate the weights and bias values of the network. Derived estimates were then applied independently in the series of 14 Landsat images, and cover fraction images for the four land

An example of the resulting surface cover fraction maps for May and August is shown in **Figure 6**. The general pattern shown in **Figure 6** matches the Urban Atlas polygons (**Figure 4**). Moreover, the fraction image corresponding to May reveals more vegetation abundance than

Although LST is routinely derived by satellite TIR observations, currently there is no spaceborne sensor capable of providing frequent thermal imagery at spatial resolution needed in urban studies. Current and forthcoming TIR remote sensing is confronting the trade-off between spatial and temporal resolution. A synergistic method that unmixes the low-resolution TIR measurements using high spatial information on the surface cover for estimating high spatial resolution LST is applied here [26]. The method is a multistep procedure described in detail along with its validation in Ref. [26]. For the method to be applied, information on the surface cover fractions is necessary. Representative emissivity values are then assigned to each of the cover types in **Figure 5**, using information derived from the ASTER Spectral Library [27]. Samples from the library, which are representative for the study area, are selected and convolved with the sensor's spectral response function and the emissivity *ε<sup>k</sup>*

> (*H*) = ∑ *i*=1 *n ε<sup>i</sup>* ∙ *aik*

Spatial-spectral unmixing is then used to enhance the spatial resolution of the low resolution thermal bands. The contribution of the land cover components is estimated for each thermal

> *i* (*H*) :

(*L*) = \_\_1 *<sup>P</sup>* ∑ *j*∈*P aj i*

are the estimated fractions of surface cover types.

(*H*)

(*H*) (3)

(*H*) (4)

is the representative emissivity value for the

the one corresponding to August both in the outskirts as well as in the urban core.

cover types assumed in this study were generated.

for each pixel high resolution (*H*) pixel *i* is estimated by:

(*H*)

*ε<sup>k</sup>*

surface cover type *i*, and *aik*

where *n* is the number of surface cover types, *ε<sup>i</sup>*

pixel *k*, by summing the estimated fractions *aj*

*Aki*

**3.3. Urban land surface temperature**

138 Multi-purposeful Application of Geospatial Data

**Figure 6.** Pseudo color composition of the derived fraction images for the May 13, 2013 (a) and August 22, 2013 (b).

where *P* is the number of high resolution pixels (*j* ∈ *P*) corresponding to each low resolution one (*k*). Each low resolution pixel is then unmixed, using the contextual information of the neighboring pixels in a window (*a* window of size *w*):

$$\underbrace{\mathbf{S}^{(\mathcal{L})}}\_{n \le 1} = \underbrace{\mathbf{A}^{(\mathcal{L})}}\_{n \le n} \cdot \underbrace{\mathbf{E}}\_{n \le 1} + er \tag{5}$$

where *S* (*L*) is a vector of the thermal radiances of the pixels in the window, *A* (*L*) is a matrix of the contributions of surface types to those pixels, and *E* is the thermal radiances under consideration. A regularization term is also used to prevent large deviations in optimization:

$$\min\_{\mathbf{E}} \left\| \mathbf{S}^{(\mathsf{L})} - \mathbf{A}^{(\mathsf{L})} \cdot \mathbf{E} + b \frac{w^2}{n} (\mathbf{E} - \mathbf{S}^{\prime(\mathsf{L})}) \right\|\_{2}^{2} \tag{6}$$

Given high resolution brightness temperature products for two thermal bands (*Ti*

*LST* = *T<sup>i</sup>* + *c*<sup>0</sup> + *c*<sup>1</sup> + *c*<sup>2</sup> <sup>2</sup> + (*c*<sup>3</sup> + *c*<sup>4</sup> *WV*)(1 − ) + (*c*<sup>5</sup> + *c*<sup>6</sup> *WV*) (7)

The LST downscaling method was applied to the series of daily MODIS and a time series of daily high resolution LST (90 m) was derived for the case study. **Figure 7** shows an example of the methodology application for a cloud-free day (30 August, 2013) for which the ASTER LST product was also available. **Figure 7a** shows the high resolution LST, derived using the above-described downscaling procedure. The ASTER (**Figure 7b**) and MODIS (**Figure 7c**) LST products corresponding to the same date are also presented for comparison. The general temperature pattern of the downscaled LST product is similar to the ASTER LST product. The level of detail that appears in the downscaled product (**Figure 7a**) compared to the ASTER LST product (**Figure 7b**) is because the downscaled product is of 30 m spatial resolution (matching the Landsat-derived surface cover fractions), while the ASTER LST product is of

Medium spatial resolution satellite data have been used in the past with spectral unmixing methods for mapping the urban surface cover [15]. This chapter demonstrated the use of image endmember and synthetic spectra to estimate sub-pixel information on the urban surface cover. The proposed methodology is fast in terms of computational time and affordable to implement and apply for urban studies. It is also easy to reproduce for other cities, if the relevant data are available. It is, thus, suitable for monitoring the surface cover and it can be used for change detection and time series analysis. The products are useful for various studies, related to surface cover properties, urban climate, urban climatology, and urban

An example use of this detailed urban surface cover information is the LST downscaling method. The method described and applied in this chapter is highly dependent on accurate surface cover information. It has been demonstrated than the uncertainty in the downscaled LST estimation is closely linked to the uncertainty related to the surface cover fractions [29]. The methodology for mapping the urban surface cover is applicable to Sentinel-2 imagery. The enhanced spatial and spectral resolution of Sentinel-2 compared to Landsat is expected to advance the method. The optical bands of Sentinel-2 are similar to the ones of Landsat 8, but the enhanced spatial resolution of 10 m provides better insights on the underlying objects. Further advances may include analysis of the 10 m bands for identifying pure spectra to be used as endmembers for spectral unmixing techniques. Moreover, the additional bands

, *εj*

where *WV* is the atmospheric water vapor content, *<sup>ε</sup>* <sup>=</sup> (*ε<sup>i</sup>*

window coefficients determined from the algorithm calibration.

respective emissivity products (*ε<sup>i</sup>*

window algorithm [28]:

90 m spatial resolution.

**4. Discussion**

expansion.

, *Tj*

http://dx.doi.org/10.5772/intechopen.71986

), LST is derived in high spatial resolution using a split-

Earth Observation for Urban Climate Monitoring: Surface Cover and Land Surface Temperature

<sup>+</sup> *<sup>ε</sup>j*)/2, <sup>=</sup> *<sup>ε</sup><sup>i</sup>*

− *ε<sup>j</sup>* , and *c* 0 -*c* 6 ) and the

141

are the split-

where *S*¯′ (*L*) are predefined spectra corresponding to surface cover types and *b* is a regularization parameter to ensure small spectral variations. The high spatial resolution thermal band is then constructed by applying Eq. (5) for high resolution (*H*).

**Figure 7.** An example of downscaled LST (K) for the August 30, 2013 (a). The ASTER (b) and MODIS (c) LST products corresponding to the same date are also presented for comparison.

Given high resolution brightness temperature products for two thermal bands (*Ti* , *Tj* ) and the respective emissivity products (*ε<sup>i</sup>* , *εj* ), LST is derived in high spatial resolution using a splitwindow algorithm [28]:

$$\mathbf{LST} = \mathbf{T}\_{\text{i}} + \mathbf{c}\_{\text{o}} + \mathbf{c}\_{\text{i}} \, \mathbf{AT} + \mathbf{c}\_{\text{z}} \, \mathbf{AT}^2 + \left(\mathbf{c}\_{\text{3}} + \mathbf{c}\_{\text{4}} \, \mathbf{WV}\right) (\mathbf{1} - \mathbf{c}) + \left(\mathbf{c}\_{\text{5}} + \mathbf{c}\_{\text{6}} \, \mathbf{WV}\right) \Delta \mathbf{e} \tag{7}$$

where *WV* is the atmospheric water vapor content, *<sup>ε</sup>* <sup>=</sup> (*ε<sup>i</sup>* <sup>+</sup> *<sup>ε</sup>j*)/2, <sup>=</sup> *<sup>ε</sup><sup>i</sup>* − *ε<sup>j</sup>* , and *c* 0 -*c* 6 are the splitwindow coefficients determined from the algorithm calibration.

The LST downscaling method was applied to the series of daily MODIS and a time series of daily high resolution LST (90 m) was derived for the case study. **Figure 7** shows an example of the methodology application for a cloud-free day (30 August, 2013) for which the ASTER LST product was also available. **Figure 7a** shows the high resolution LST, derived using the above-described downscaling procedure. The ASTER (**Figure 7b**) and MODIS (**Figure 7c**) LST products corresponding to the same date are also presented for comparison. The general temperature pattern of the downscaled LST product is similar to the ASTER LST product. The level of detail that appears in the downscaled product (**Figure 7a**) compared to the ASTER LST product (**Figure 7b**) is because the downscaled product is of 30 m spatial resolution (matching the Landsat-derived surface cover fractions), while the ASTER LST product is of 90 m spatial resolution.
