hyperplanes for SVM, it may be highly possible to reduce training sample size to a
small number of the most informative samples that are used to fit the decision
hyperplanes. Several studies have been conducted to identify these critical samples.
For example, Foody and Marthur (2004a, b, 2006) incorporated ancillary information of soil types and geographical boundary pixels of mixed spectral characteristics
of two crop types in the selection of useful training samples, which dramatically
reduced training samples before being applied to classification. They also examined
the usefulness of applying other ancillary information (e.g., landform, moisture, and
spatial texture) in targeting support vectors. Various techniques have been identified to automatically reduce the training sample size and hence help reduce the
computational burden for SVM. For example, clustering-based algorithms are
applied in training pattern selection to remove samples locating at the high density
regions or to detect support vectors at the clustering centers (Demir and Ertu ¨rk
2009; Su 2009). With these support vectors obtained from clustering preprocessing,
the computational load has been substantially reduced, while the classification
accuracy was much higher than using the full training samples.
13.3 Implementation of SVM for Land Cover Mapping
In order to demonstrate the effectiveness of SVM for heterogeneous land cover
mapping, we implemented SVM to map land cover types in an urban area. In this
section, we will discuss the specific procedures, including the study site and data
acquisition, classification scheme design, SVM configuration, and classification
and accuracy assessment (Fig. 13.1).
13.3.1 Study Site and Data Acquisition
The study site covers the entire Gwinnett County, a suburban county located at
northeastern Atlanta metropolitan area, Georgia, USA (Fig. 13.2). The county has
an area of about 1,122 km
2 and its population was 805,321 according to the 2010
census survey. The majority of topography is relatively flat and has primarily a
humid subtropical climate. Gwinnett has been one of America’s fastest-growing
counties and the second most populated county in Georgia. Its landscape is characterized by a mosaic of complex land use and land cover types, and therefore
Gwinnett is an ideal site to examine the effectiveness of SVM for heterogeneous
landscape mapping.
A cloud-free Landsat-5 Thematic Mapper (TM) image dated on 19 May 2007
was acquired from USGS EROS Data Center, and a subset of this scene covering
the entire Gwinnett County was actually used in our study (Fig. 13.3). The image
has been geometrically corrected at the EROS data center, and no further
preprocessing was conducted. The spatial resolution of this image is 30 m for all
13 Support Vector Machines for Land Cover Mapping from Remote Sensor Imagery
269
small number of the most informative samples that are used to fit the decision
hyperplanes. Several studies have been conducted to identify these critical samples.
For example, Foody and Marthur (2004a, b, 2006) incorporated ancillary information of soil types and geographical boundary pixels of mixed spectral characteristics
of two crop types in the selection of useful training samples, which dramatically
reduced training samples before being applied to classification. They also examined
the usefulness of applying other ancillary information (e.g., landform, moisture, and
spatial texture) in targeting support vectors. Various techniques have been identified to automatically reduce the training sample size and hence help reduce the
computational burden for SVM. For example, clustering-based algorithms are
applied in training pattern selection to remove samples locating at the high density
regions or to detect support vectors at the clustering centers (Demir and Ertu ¨rk
2009; Su 2009). With these support vectors obtained from clustering preprocessing,
the computational load has been substantially reduced, while the classification
accuracy was much higher than using the full training samples.
13.3 Implementation of SVM for Land Cover Mapping
In order to demonstrate the effectiveness of SVM for heterogeneous land cover
mapping, we implemented SVM to map land cover types in an urban area. In this
section, we will discuss the specific procedures, including the study site and data
acquisition, classification scheme design, SVM configuration, and classification
and accuracy assessment (Fig. 13.1).
13.3.1 Study Site and Data Acquisition
The study site covers the entire Gwinnett County, a suburban county located at
northeastern Atlanta metropolitan area, Georgia, USA (Fig. 13.2). The county has
an area of about 1,122 km
2 and its population was 805,321 according to the 2010
census survey. The majority of topography is relatively flat and has primarily a
humid subtropical climate. Gwinnett has been one of America’s fastest-growing
counties and the second most populated county in Georgia. Its landscape is characterized by a mosaic of complex land use and land cover types, and therefore
Gwinnett is an ideal site to examine the effectiveness of SVM for heterogeneous
landscape mapping.
A cloud-free Landsat-5 Thematic Mapper (TM) image dated on 19 May 2007
was acquired from USGS EROS Data Center, and a subset of this scene covering
the entire Gwinnett County was actually used in our study (Fig. 13.3). The image
has been geometrically corrected at the EROS data center, and no further
preprocessing was conducted. The spatial resolution of this image is 30 m for all
13 Support Vector Machines for Land Cover Mapping from Remote Sensor Imagery
269