2012 were eliminated due to a large cloud obstructed area. In order to minimize the
phenological effect during change detection analysis, data acquired in summer
season were preferred. Most of the data are obtained from June to September. All
used Landsat data were retrieved from USGS Global Visualization Viewer
(GloVis) interface (http://glovis.usgs.gov/).
To help select training samples of the supervised classification and reference
samples for accuracy assessment, two full-colour digital orthoimages with 12 cm
spatial resolution have been acquired from University Geospatial Centre. These two
orthoimages cover the entire area of Waterloo Region in 2006 and 2010 respectively. With 12 cm spatial resolution, the orthoimages have been projected in UTM
coordinates and they are stored in MrSID image format accompanying with SDW
world files. The datum used is the North American Datum of 1983 (NAD83).
Another data that can be used for aiding choosing training samples is a land use
shapefile of Waterloo Region in 2007. It also has been obtained from the University
of Waterloo Geospatial Centre. This shapefile parcels the study area into polygons
based on different land use types. Additionally, Google Map is providing high
spatial resolution aerial or satellite images of the world. It is an auxiliary source for
training samples selection and accuracy assessment. Selecting appropriate training
samples is very critical for satisfactory classification results. Even though there is
no reference data for every year, the two orthoimages (i.e. the land use shapefile and
images from Google Maps) are the effective reference data for understanding the
land surface information of this study area.
4.3.2 Classification Scheme
Image classification is the most important process for obtaining accurate LULC
information. To generate consistent classification results, an appropriate classification algorithm needs to be determined. For this case study a support vector machine
(SVM) was used for the classification process. It is because of its outstanding
performance in contrast to traditional classifier such as MLC for LULC classification using remote sensing techniques (Huang et al. 2002; Pal 2005; Frohn and
Arellano-Neri 2005; Gislason et al. 2006; Kotsiantis 2007; Benediktsson
et al. 2007; Mellor et al. 2013). SVM is based on statistical learning theory and
used structural risk minimization method proposed that was by Vapnik to discriminate class members (Nemmour and Chibani 2011; Song et al. 2012). SVM employs
optimization algorithms to decide the location of optimal boundaries that can best
separate the classes (Huang et al. 2002; Pal and Mather 2005). Thus a minimal
generalization error can be obtained by minimizing the probability of misclassification of the unseen data points. The workflow of classification process is shown
in Fig. 4.4.
An appropriate classification system and sufficient representative training samples are very critical for a successful classification (Lu and Weng 2007). As
referring to the USGS “Land-Use/Land-Cover Classification System for Use with
74
A. Fu et al.
phenological effect during change detection analysis, data acquired in summer
season were preferred. Most of the data are obtained from June to September. All
used Landsat data were retrieved from USGS Global Visualization Viewer
(GloVis) interface (http://glovis.usgs.gov/).
To help select training samples of the supervised classification and reference
samples for accuracy assessment, two full-colour digital orthoimages with 12 cm
spatial resolution have been acquired from University Geospatial Centre. These two
orthoimages cover the entire area of Waterloo Region in 2006 and 2010 respectively. With 12 cm spatial resolution, the orthoimages have been projected in UTM
coordinates and they are stored in MrSID image format accompanying with SDW
world files. The datum used is the North American Datum of 1983 (NAD83).
Another data that can be used for aiding choosing training samples is a land use
shapefile of Waterloo Region in 2007. It also has been obtained from the University
of Waterloo Geospatial Centre. This shapefile parcels the study area into polygons
based on different land use types. Additionally, Google Map is providing high
spatial resolution aerial or satellite images of the world. It is an auxiliary source for
training samples selection and accuracy assessment. Selecting appropriate training
samples is very critical for satisfactory classification results. Even though there is
no reference data for every year, the two orthoimages (i.e. the land use shapefile and
images from Google Maps) are the effective reference data for understanding the
land surface information of this study area.
4.3.2 Classification Scheme
Image classification is the most important process for obtaining accurate LULC
information. To generate consistent classification results, an appropriate classification algorithm needs to be determined. For this case study a support vector machine
(SVM) was used for the classification process. It is because of its outstanding
performance in contrast to traditional classifier such as MLC for LULC classification using remote sensing techniques (Huang et al. 2002; Pal 2005; Frohn and
Arellano-Neri 2005; Gislason et al. 2006; Kotsiantis 2007; Benediktsson
et al. 2007; Mellor et al. 2013). SVM is based on statistical learning theory and
used structural risk minimization method proposed that was by Vapnik to discriminate class members (Nemmour and Chibani 2011; Song et al. 2012). SVM employs
optimization algorithms to decide the location of optimal boundaries that can best
separate the classes (Huang et al. 2002; Pal and Mather 2005). Thus a minimal
generalization error can be obtained by minimizing the probability of misclassification of the unseen data points. The workflow of classification process is shown
in Fig. 4.4.
An appropriate classification system and sufficient representative training samples are very critical for a successful classification (Lu and Weng 2007). As
referring to the USGS “Land-Use/Land-Cover Classification System for Use with
74
A. Fu et al.