**4. Satellite estimates of temperature versus ground measurements**

In this Section, a methodology is presented in which the temperature estimates from the MODIS sensor onboard the Terra Satellite is contrasted with ground measurements. The methodology consists of a neural network approach in which measurements on the ground are used as input to the neural network, whereas, the temperature estimate from the satellite is considered as the network's output.

The neural network methodology adopted has successfully been implemented before in tackling several climatological problems in Cyprus: the prediction of maximum daily total solar irradiance (Kalogirou et al., 2002), the prediction of the daily average solar radiation (Tymvios et al., 2002, 2005a), the modeling of photosynthetic radiation (Tymvios et al., 2005b) and others.

#### **4.1. Data**

**Figure 7.** The height pattern at 500hPa (Cluster 5) from 24 November 1962

8 Remote Sensing of Environment: Integrated Approaches

of the atmosphere.

850hPa, 700hPa).

Although it appears that some Clusters are associated with heat events over Cyprus, the connection between heat events and atmospheric circulation at 500hPa did not give definite results that any of these patterns dominate heat event occurrences (as it was possible to demonstrate in previous studies on rainfall and extreme rainfall events). There might be two reasons for this inadequacy. The first is that the window that was chosen for the classifica‐ tion does not include the synoptic patterns that influence the area sufficiently; the second reason is that, although upper air patterns at 500hPa contribute significantly to the evolution of certain surface features (such as dynamical or extreme rainfall), such an association is not so clear for the temperature field. In the search for associations of the temperature fields with synoptic patterns in the Mediterranean, it is important to consider also the lower parts

Future research concerning the connection of the weather classification patterns will be fo‐ cused into a new, much larger window that will include Northern Africa and the Middle East and a combination of classification of patterns at lower levels of the atmosphere (e.g., For the needs of this research, data from MODIS onboard the Terra satellite have been used. More specifically, the level-3 product MOD11A1 (version 5) for the period 2000-2009 was exploited, at a resolution of about 1km by 1km (0.927km). Using the available Land Surface Temperature (LST) fields derived from MODIS, a time series was established corresponding to the position of ground stations. Wan & Dozier (1996) have developed the Generalized Split Window (GSW) algorithm for the retrieval of LST, using the thermal (infrared) channels of MODIS and under different atmospheric condi‐ tions (see also, Wan, 1999, 2008). This algorithm retrieves LST on the basis of emissivi‐ ties in bands 31 and 32 of MODIS. The accuracy in estimating LST was found to be better than 1K, whereas in most cases it was better than 0.5K (Hulley & Hook, 2009; Coll et al., 2009).

The data base for the surface measurements used in this research consist of the hourly re‐ corded temperature at each of the automatic meteorological stations of the network operat‐ ed by the Cyprus Meteorological Service (see Fig. 4), in the period 2000-2007. Based on these data, the maximum temperature recorded in the time period 1100 – 1300 UTC (local stand‐ ard time=UTC+2 hours) was considered as the day's maximum and was subsequently used in the study.

The training of the neural network implemented requires that there are no missing data in the time series used, because the data are used in groups and are therefore inter-dependent. Therefore, the estimated LST (by the neural network implemented) is based on the data of a whole day and missing values result in the rejection of that day. Following quality control based on the above constraint, the number of automatic stations that were subsequently used was reduced to twelve, as shown in Table 1.

#### **4.2. Methodology**

Artificial Neural networks (ANN) are small autonomous computational units (algorithms) with inter-connections which, to a large extent, resemble the functioning of natural compu‐ tational units, namely, the neurons of the human brain. ANN can be trained and learn through repeated examples so that they can reach conclusions and results without human intervention. Since their invention, ANN covered a wide spectrum of research and disci‐ plines and their application has been phenomenal. A few of the numerous examples of ANN applications are mentioned here: medical systems' automation for the recognition of malignant tumors, control of military equipment and aircraft, estimation of environmental variables, quality verification in production factories, forecasting of financial indices, weath‐ er diagnosis and forecasting etc.

**Ground station**

**Latitude (N)**

**Longitude (E)**

**Altitude Above sea level (m)**

**Astromeritis** 35°03´ 32°26´ 175 6.81 -7.31 0.22 2.35

**Athalassa** 35°04´ 33°58´ 162 7.14 -7.58 0.63 2.77

**Athienou** 35°03´ 33°32´ 185 6.45 -5.74 0.23 2.27

**Dasaki** 35°03´ 33°47´ 50 5.74 -4.52 0.60 2.17

**Kannaviou** 34°55´ 32°35´ 419 7.03 -6.23 0.48 2.41

**Kathikas** 34°55´ 32°26´ 650 8.25 -5.30 0.63 2.53

**Kato Pyrgos** 35°11´ 32°41´ 5 7.19 -10.12 0.62 2.99

**Malia** 34°49´ 32°47´ 645 7.59 -7.75 -0.11 2.56

**Mennogeia** 34°51´ 33°26´ 140 7.36 -7.78 0.28 3.07

**Paphos** 34°47´ 32°26´ 82 8.34 -7.45 0.37 2.83

**Paralimni** 35°04´ 33°58´ 65 7.51 -8.42 0.68 2.61

**Polis** 35°03´ 32°26´ 20 6.87 -4.83 0.40 2.40

In this research, an attempt has been made to relate ground measurements of temperature with the temperature as it is estimated from MODIS and develop a neural network method‐ ology that can be used in the estimation of ground temperatures by using the satellite im‐ agery. Although the methodology performs sufficiently, overall, it seems that further refinement is needed in order to improve the approach. The adoption of a single network for all the time series of data seems to limit the application of the methodology. For example, the present single neural network developed for each station does not take into account the large seasonal variations in the parameter concerned. It could be more effective if several

**Table 1.** Errors of LST estimation for the independent set of data for each ground station

neural networks are developed based on seasonally grouped data.

**Maximum (°C)**

Satellite and Ground Measurements for Studying the Urban Heat Island Effect in Cyprus

**Minimum (°C)**

**Average (°C)**

http://dx.doi.org/10.5772/39313

**Standard deviation (°C)**

11

For the implementation of the ANN methodology in the present research, the Multi-Layer Perceptron (MLP) was adopted (see Haykin, 1998). The input to this network is the surface temperature recorded at the ground stations and the output is the temperature estimated by the satellite (LST). Fig. 8 displays the MLP implemented.

**Figure 8.** The Multi-Layer Perceptron (MLP) network implemented for the prediction of LST

The data from the twelve ground stations and the respective MODIS estimations of LST were used as follows: 60% were used for Training of the network, 20% for Validation and the remaining 20% as an independent set for Testing.

#### **4.3. Results**

Table 1 displays the errors in the estimation of LST with the neural network, by using the independent set of data. In this table, the maximum, minimum and average errors along with the standard deviation are shown for each of the ground stations. Overall, the perform‐ ance of the neural network is considered as very satisfactory. However, there are cases where the error is unacceptable and this requires further investigation.

The relation between input data (ground temperature) and satellite estimated temperature LST (target) is shown in Fig. 9. The results are shown for all the data but also separately for the Training, Validation and Testing data sets. For the Training set, the correlation coeffi‐ cient is R=0.96991, for the Validation set R=0.89692 and for the Testing set R=0.9145, whereas, for all the data R=0.94747. Based on these findings, the performance of the network in pre‐ dicting LST is considered as satisfactory.

Satellite and Ground Measurements for Studying the Urban Heat Island Effect in Cyprus http://dx.doi.org/10.5772/39313 11


**Table 1.** Errors of LST estimation for the independent set of data for each ground station

tational units, namely, the neurons of the human brain. ANN can be trained and learn through repeated examples so that they can reach conclusions and results without human intervention. Since their invention, ANN covered a wide spectrum of research and disci‐ plines and their application has been phenomenal. A few of the numerous examples of ANN applications are mentioned here: medical systems' automation for the recognition of malignant tumors, control of military equipment and aircraft, estimation of environmental variables, quality verification in production factories, forecasting of financial indices, weath‐

For the implementation of the ANN methodology in the present research, the Multi-Layer Perceptron (MLP) was adopted (see Haykin, 1998). The input to this network is the surface temperature recorded at the ground stations and the output is the temperature estimated by

The data from the twelve ground stations and the respective MODIS estimations of LST were used as follows: 60% were used for Training of the network, 20% for Validation and

Table 1 displays the errors in the estimation of LST with the neural network, by using the independent set of data. In this table, the maximum, minimum and average errors along with the standard deviation are shown for each of the ground stations. Overall, the perform‐ ance of the neural network is considered as very satisfactory. However, there are cases

The relation between input data (ground temperature) and satellite estimated temperature LST (target) is shown in Fig. 9. The results are shown for all the data but also separately for the Training, Validation and Testing data sets. For the Training set, the correlation coeffi‐ cient is R=0.96991, for the Validation set R=0.89692 and for the Testing set R=0.9145, whereas, for all the data R=0.94747. Based on these findings, the performance of the network in pre‐

er diagnosis and forecasting etc.

10 Remote Sensing of Environment: Integrated Approaches

the satellite (LST). Fig. 8 displays the MLP implemented.

**Figure 8.** The Multi-Layer Perceptron (MLP) network implemented for the prediction of LST

where the error is unacceptable and this requires further investigation.

the remaining 20% as an independent set for Testing.

dicting LST is considered as satisfactory.

**4.3. Results**

In this research, an attempt has been made to relate ground measurements of temperature with the temperature as it is estimated from MODIS and develop a neural network method‐ ology that can be used in the estimation of ground temperatures by using the satellite im‐ agery. Although the methodology performs sufficiently, overall, it seems that further refinement is needed in order to improve the approach. The adoption of a single network for all the time series of data seems to limit the application of the methodology. For example, the present single neural network developed for each station does not take into account the large seasonal variations in the parameter concerned. It could be more effective if several neural networks are developed based on seasonally grouped data.

**5. Land surface temperature analysis**

Cyprus during the extreme heat wave of August 2010.

**5.1. MODIS LST temporal evolution and temperature anomaly maps**

thermal bands.

2005; Wan, 2008).

and day Aqua passes.

The MODIS sensor, onboard Terra and Aqua polar satellites, provides one day and one night image under clear sky conditions. MODIS is particularly suitable for the land sur‐ face temperature (LST) product due to its global coverage, radiometric resolution and dy‐ namic ranges for a variety of land cover types and high calibration accuracy in multiple

Satellite and Ground Measurements for Studying the Urban Heat Island Effect in Cyprus

http://dx.doi.org/10.5772/39313

13

MODIS LST product is based on the generalized split-window (GSW) algorithm (Wan & Dozier, 1996) using as input the MODIS thermal bands 31 and 32. The parameters in the MODIS GSW depend on the satellite zenith view angles, column water vapor and also on the low atmosphere boundary temperature. The band emissivities rely on the classificationbased method (Snyder et al., 1998) according to land cover types in the pixel (Monteiro et al., 2007). Temperatures are extracted in Kelvin; accuracy of 1 Kelvin is yielded for materials with known emissivities (Wan, 1999), while a number of studies have also tested the accura‐ cy of the MODIS LST product with favorable results (Wan, 2002; Wan et al., 2004; Coll et al.,

The MODIS Aqua product MYD11A1 (V5) and MODIS Terra product MOD11A1 (V5) – Land Surface Temperature and Emissivity Daily L3 Global 1 km Grid SIN were used. Terra and Aqua overpass times for the study area are considered at approximately 1030 and 1330 UTC for day passes, and at approximately 2230 and 0130 UTC for night passes, respectively. The use of MODIS LST data for examining the temporal evolution and the retrieved temper‐ ature anomaly maps for a heat wave event occurred on 24 June 2007 is presented. Moreover, MODIS LST data are used for calculating the urban heat island (UHI) at four urban areas of

MODIS LST data were initially used for generating mean monthly climatology LST maps for June in the period of 2003-2008. The mean and maximum Aqua day and night LST values for June are presented in Fig. 10 for the period 2003-2008 for two urban areas (Nicosia, Lar‐ naca) and one rural area (Ag. Marina). The curves show that the mean night LST values for the two urban areas are similar, while for the area of Ag. Marina, the temperature levels are 3-4 °C lower. For all sites, a minimum was observed for year 2005. The situation is different though regarding day LST values. The coastal site of Larnaca exhibited the lowest values among the three areas, while Nicosia and Ag. Marina exhibited similar patterns and temper‐

ature levels. The overall trend over time for the three areas showed a positive trend.

The intense heat wave event of 24 June 2007 was next examined in order to study the LST behavior during such events since satellite derived LST is controlled by land cover and topographic effect factors. In Fig.11, temperature anomaly maps, in terms of temper‐ ature deviation from the long-term monthly mean values (calculated for the period 2003-2008), are presented for the heat wave event under consideration and for both night

**Figure 9.** Neural network performance for the Training, Validation, Test and All data sets
