**3. Development of methodology for the optimization of classification accuracy of Landsat TM/ETM+ imagery in a catchment area in Cyprus**

ries, according to their level of humidity: 0% (dry sample); 25%; 50%; > 50%. With regard to tile and roof specimens, the results were divided into "dry" or "humid" categories due to

Integrated Remote Sensing and GIS Applications for Sustainable Watershed Management: A Case Study from Cyprus

http://dx.doi.org/10.5772/39307

101

**Figure 2.** Collection of soil data (left). Spectroradiometric measurements of material samples at the premises of the

Based on the results of the scatter-plots, it was found that in the case of dry samples there is a strong spectral confusion between the chalk A response and the urban fabric (roof and tile) materials. The "moisture" scatter plot (humidity > 50%) highlights the different spectral re‐ sponse between artificial materials (roof and tile) and natural materials (chalk A, B, C). In this plot, the spectral difference between different samples is increased and two major clus‐ ters are created with complete contrary spectral response (increase of chalk A spectral re‐ sponse and substantial decrease of tile and house roof -constructed from clay and cement

The results highlighted the different spectral response of materials under different humidity levels. Specifically, reflectance values of chalk samples (samples A and C) tend to be sepa‐

**Figure 3.** Scatter-plots of the different targets examined in this study for Band 1 – Band 4 (humidity 0%) (left) and

rated from those of urban samples (tile and roof) as humidity increases.

Remote Sensing and Geomatics Laboratory of CUT (right)

consecutively- spectral response, see Fig.3).

Band 1 - Band 4 of Landsat (humidity > 50%) (right)

the difficulty to measure the scaling levels of humidity in those kinds of materials.

#### **3.1. Introduction**

An important tool for the detection and quantification of land cover changes across catch‐ ment areas is the classification of multispectral satellite imagery, as such results are very im‐ portant for hydrological analysis and flood scenarios.

This study aimed at testing different material samples in the Yialias region (central Cy‐ prus) in order to examine: a) their spectral behavior under different precipitation rates and b) to introduce an alternative methodology to optimize the classification results de‐ rived from single satellite imagery with the combined use of satellite, spectroradiometric and precipitation data.

#### **3.2. Data and methodology**

#### *3.2.1. Ground sample*

According to preliminary classification results (Alexakis et al., 2011), spectral mixing be‐ tween urban areas and specific geological formations was observed. Thus, samples of re‐ golith and construction material were collected and tested for their spectral response under different conditions of humidity with the use of spectroradiometer in the premises of the Remote Sensing and Geomatics Laboratory of Cyprus University of Technology (Alexakis et al., 2012).

#### *3.2.2. Satellite and precipitation data*

For the purposes of the study, specific tools and data were incorporated:


In order to investigate the different spectral response of each sample under different mois‐ ture conditions, all samples were immersed in water in a step-by-step process and measured for the rate of their humidity with a soil moisture meter. The specific hand-held instrument used in this study was able to measure moisture values from 0 to 50% within an accuracy of 0.1%. The final under investigation regolith samples were divided in four different catego‐ ries, according to their level of humidity: 0% (dry sample); 25%; 50%; > 50%. With regard to tile and roof specimens, the results were divided into "dry" or "humid" categories due to the difficulty to measure the scaling levels of humidity in those kinds of materials.

**3. Development of methodology for the optimization of classification accuracy of Landsat TM/ETM+ imagery in a catchment area in Cyprus**

An important tool for the detection and quantification of land cover changes across catch‐ ment areas is the classification of multispectral satellite imagery, as such results are very im‐

This study aimed at testing different material samples in the Yialias region (central Cy‐ prus) in order to examine: a) their spectral behavior under different precipitation rates and b) to introduce an alternative methodology to optimize the classification results de‐ rived from single satellite imagery with the combined use of satellite, spectroradiometric

According to preliminary classification results (Alexakis et al., 2011), spectral mixing be‐ tween urban areas and specific geological formations was observed. Thus, samples of re‐ golith and construction material were collected and tested for their spectral response under different conditions of humidity with the use of spectroradiometer in the premises of the Remote Sensing and Geomatics Laboratory of Cyprus University of Technology

For the purposes of the study, specific tools and data were incorporated:

**•** Four Landsat TM/ ETM+ multispectral images of medium resolution (30x30 m2

**•** Precipitation data obtained from the Meteorological Service of Cyprus (Pera Chorio Mete‐ orological Station : Lon - 35° 01', Latitude - 33° 23'). All of these data were compared with the satellite imagery data. Selected satellite imagery was retrieved a day after the record‐ ing of substantial scaling amount of precipitation from the Pera-Chorio Metereological

**•** Data derived from spectroradiometric field campaigns. For this reason a GER 1500 spec‐ troradiometer was used. This instrument can record electromagnetic radiation between

In order to investigate the different spectral response of each sample under different mois‐ ture conditions, all samples were immersed in water in a step-by-step process and measured for the rate of their humidity with a soil moisture meter. The specific hand-held instrument used in this study was able to measure moisture values from 0 to 50% within an accuracy of 0.1%. The final under investigation regolith samples were divided in four different catego‐

pixel size).

**3.1. Introduction**

and precipitation data.

*3.2.1. Ground sample*

(Alexakis et al., 2012).

Station.

*3.2.2. Satellite and precipitation data*

350 nm up to 1050 nm (Fig. 2).

**3.2. Data and methodology**

portant for hydrological analysis and flood scenarios.

100 Remote Sensing of Environment: Integrated Approaches

**Figure 2.** Collection of soil data (left). Spectroradiometric measurements of material samples at the premises of the Remote Sensing and Geomatics Laboratory of CUT (right)

Based on the results of the scatter-plots, it was found that in the case of dry samples there is a strong spectral confusion between the chalk A response and the urban fabric (roof and tile) materials. The "moisture" scatter plot (humidity > 50%) highlights the different spectral re‐ sponse between artificial materials (roof and tile) and natural materials (chalk A, B, C). In this plot, the spectral difference between different samples is increased and two major clus‐ ters are created with complete contrary spectral response (increase of chalk A spectral re‐ sponse and substantial decrease of tile and house roof -constructed from clay and cement consecutively- spectral response, see Fig.3).

The results highlighted the different spectral response of materials under different humidity levels. Specifically, reflectance values of chalk samples (samples A and C) tend to be sepa‐ rated from those of urban samples (tile and roof) as humidity increases.

**Figure 3.** Scatter-plots of the different targets examined in this study for Band 1 – Band 4 (humidity 0%) (left) and Band 1 - Band 4 of Landsat (humidity > 50%) (right)

#### *3.2.3. Satellite imagery data*

After the application of all necessary pre-processing steps (radiometric, atmospheric and geometric corrections,) spectral signature profiles were extracted for all of the different ma‐ terials during the acquisition dates of each satellite imagery (Fig. 4).

**Figure 4.** Scatter-plots of the different targets examined in this study for Band 1 – Band 4 (left) and Band 3 - Band 4 of Landsat (right)

The results of the scatter plots denoted the scaling optimization of spectral separability of satellite imagery data, from 0 to 23.7 mm of precipitation. Specifically, concerning the 0 mm precipitation case, a spectral confusion was indicated between the "urban" targets (roof and tile) and chalk A and C targets. This conflict was outreached gradually as the precipitation level increased. The samples started to have different spectral behaviour, with the chalk samples (except chalk B) standing gradually away from the "urban" sam‐ ples cluster in the scatter-plot. It is important to mention the quite different spectral re‐ sponse of chalk C sample in satellite images compared to its response in the laboratory specimens. This problem occurred due to the medium spatial resolution of Landsat im‐ ages (30x30 m2 pixel size) which increases the likelihood of the common mixing pixel phenomenon.

**Figure 5.** Detail of the "rainy" satellite image after the application of supervised classification algorithm

but with insufficient accuracy to be considered as credential.

**3.4. Conclusions**

analysis.

On the one hand, the results of the unsupervised algorithm performance for both dry and humid acquisition days could be described as poor and were not considered for further evaluation (Kappa coefficient of classification accuracy - (Kc) < 60%). On the other hand, the application of supervised algorithm to "rainy" image provided better accuracy results (Kc = 0.75). The product of "dry" image was substantially better than that of unsupervised case

Integrated Remote Sensing and GIS Applications for Sustainable Watershed Management: A Case Study from Cyprus

http://dx.doi.org/10.5772/39307

103

The results noted the importance of imagery acquisition date for optimization of classifi‐ cation results. Specifically, the overall accuracy of classification product was substantial‐ ly increased (more than 30% for supervised classification), especially for urban and marl/ chalk areas, during days where high precipitation measurements were recorded in the broader study area. The results were established either by laboratory or satellite imagery

#### **3.3. Results and verification**

The results from the laboratory and satellite imagery analysis methods highlighted the different spectral response of materials to different levels of humidity. For the direct comparison of the classification accuracy between images, where different levels of pre‐ cipitation have been recorded, two Landsat TM/ETM+ images acquired on 2 June 2005 (0 mm precipitation – "dry") and 23 July 2009 (23.7 mm precipitation – "rainy") were clas‐ sified and compared (Fig. 5). Both unsupervised (ISODATA) and supervised classifica‐ tion algorithms (Maximum Likelihood - ML) were used. Initially, the ISODATA classification technique was applied to both images with 95% convergence threshold. The following 5 classes were used for both the supervised and unsupervised algorithms: 1) urban Fabric, 2) marl - chalk formations, 3) vegetation, 4) bare soil and 5) forest.

Integrated Remote Sensing and GIS Applications for Sustainable Watershed Management: A Case Study from Cyprus http://dx.doi.org/10.5772/39307 103

**Figure 5.** Detail of the "rainy" satellite image after the application of supervised classification algorithm

On the one hand, the results of the unsupervised algorithm performance for both dry and humid acquisition days could be described as poor and were not considered for further evaluation (Kappa coefficient of classification accuracy - (Kc) < 60%). On the other hand, the application of supervised algorithm to "rainy" image provided better accuracy results (Kc = 0.75). The product of "dry" image was substantially better than that of unsupervised case but with insufficient accuracy to be considered as credential.

#### **3.4. Conclusions**

*3.2.3. Satellite imagery data*

102 Remote Sensing of Environment: Integrated Approaches

Landsat (right)

phenomenon.

**3.3. Results and verification**

After the application of all necessary pre-processing steps (radiometric, atmospheric and geometric corrections,) spectral signature profiles were extracted for all of the different ma‐

**Figure 4.** Scatter-plots of the different targets examined in this study for Band 1 – Band 4 (left) and Band 3 - Band 4 of

The results of the scatter plots denoted the scaling optimization of spectral separability of satellite imagery data, from 0 to 23.7 mm of precipitation. Specifically, concerning the 0 mm precipitation case, a spectral confusion was indicated between the "urban" targets (roof and tile) and chalk A and C targets. This conflict was outreached gradually as the precipitation level increased. The samples started to have different spectral behaviour, with the chalk samples (except chalk B) standing gradually away from the "urban" sam‐ ples cluster in the scatter-plot. It is important to mention the quite different spectral re‐ sponse of chalk C sample in satellite images compared to its response in the laboratory specimens. This problem occurred due to the medium spatial resolution of Landsat im‐ ages (30x30 m2 pixel size) which increases the likelihood of the common mixing pixel

The results from the laboratory and satellite imagery analysis methods highlighted the different spectral response of materials to different levels of humidity. For the direct comparison of the classification accuracy between images, where different levels of pre‐ cipitation have been recorded, two Landsat TM/ETM+ images acquired on 2 June 2005 (0 mm precipitation – "dry") and 23 July 2009 (23.7 mm precipitation – "rainy") were clas‐ sified and compared (Fig. 5). Both unsupervised (ISODATA) and supervised classifica‐ tion algorithms (Maximum Likelihood - ML) were used. Initially, the ISODATA classification technique was applied to both images with 95% convergence threshold. The following 5 classes were used for both the supervised and unsupervised algorithms: 1) urban Fabric, 2) marl - chalk formations, 3) vegetation, 4) bare soil and 5) forest.

terials during the acquisition dates of each satellite imagery (Fig. 4).

The results noted the importance of imagery acquisition date for optimization of classifi‐ cation results. Specifically, the overall accuracy of classification product was substantial‐ ly increased (more than 30% for supervised classification), especially for urban and marl/ chalk areas, during days where high precipitation measurements were recorded in the broader study area. The results were established either by laboratory or satellite imagery analysis.
