95
simple to understand, and the accuracy of the results can be assessed quantitatively
and qualitatively (McDermid et al. 2005).
Traditional classification methods can generally be referred to as either unsupervised or supervised. As stated from the name, unsupervised classification does not
require prior knowledge of the area, whereas supervised classification does require
a priori knowledge for training the classifier. Classes can then be assigned based on
predicted variables measured from the training dataset (Cerna and Chytry 2005).
Examples of unsupervised algorithms include K-means and ISODATA, and are
often used for thematic mapping on a large scale as they do not require spatially
detailed ground data initially, and produce useful information by clustering spectrally similar pixels (Tso and Olsen 2005; Giri et al. 2011). The algorithms are also
often widely available in image processing and statistical software packages
(Langley et al. 2001). However, the benefits of unsupervised techniques are often
outweighed by the difficulty of post classification labelling, which does require
ground information (McDermid et al. 2005).
The most widely used supervised classification technique is the Maximum
Likelihood Classifier (MLC). It has been used to successfully map areas with high
classification accuracies (MacAlister and Mahaxay 2009; Laba et al. 2008) but
works on the assumption that input data follow a Gaussian distribution. MLC may
perform poorly in the presence of non-parametric distributions (Peddle 1995) as it
is heavily reliant on the distribution of the training data. Another method includes
the calculation of spectral indices to characterise specific attributes of plants
(DeFries et al. 1995). The most widely used spectral index is the NDVI and is based
on the fact that vegetation is highly reflective in the near infrared and has a high
absorption rate in the visible red wavelengths. The contrast between these wavelengths can be used as a biophysical parameter that correlates with photosynthetic
activity of vegetation, which is an indication of ‘greenness’ (Wang and Tenhunen
2004). Therefore, NDVI is a good indicator to reflect dynamic changes of different
vegetation groups (Geerken et al. 2005). This area of spectral indices has expanded
rapidly and numerous indices are available for different biophysical parameters
(e.g, soil moisture, plant senescence).
Another approach to mapping in comparison to pixel based classification is
object-based image analysis (OBIA). Environmental objects are parts of the real
world for which information can or should become available (e.g, a tree). As soon
as the real world part becomes a formal object in a spatial dimension then it can be
subject to some kind of classification (Bock et al. 2005). However, object based
analysis is very much dependent on the spatial resolution of available imagery. The
follow-up map for the UK LCM (Fuller et al. 2002) was one of the principal drivers
in the development of OBIA in ecological remote sensing.
The advantage of the object-based approach is that it offers new possibilities for
image analysis as image objects can be characterised not only by spectral values but
by texture, shape, context, relationship and thematic information supplied by ancillary data (Bock et al. 2005). Additional layers of knowledge can be valuable to a
Mapping Coastal Habitats in Wales
simple to understand, and the accuracy of the results can be assessed quantitatively
and qualitatively (McDermid et al. 2005).
Traditional classification methods can generally be referred to as either unsupervised or supervised. As stated from the name, unsupervised classification does not
require prior knowledge of the area, whereas supervised classification does require
a priori knowledge for training the classifier. Classes can then be assigned based on
predicted variables measured from the training dataset (Cerna and Chytry 2005).
Examples of unsupervised algorithms include K-means and ISODATA, and are
often used for thematic mapping on a large scale as they do not require spatially
detailed ground data initially, and produce useful information by clustering spectrally similar pixels (Tso and Olsen 2005; Giri et al. 2011). The algorithms are also
often widely available in image processing and statistical software packages
(Langley et al. 2001). However, the benefits of unsupervised techniques are often
outweighed by the difficulty of post classification labelling, which does require
ground information (McDermid et al. 2005).
The most widely used supervised classification technique is the Maximum
Likelihood Classifier (MLC). It has been used to successfully map areas with high
classification accuracies (MacAlister and Mahaxay 2009; Laba et al. 2008) but
works on the assumption that input data follow a Gaussian distribution. MLC may
perform poorly in the presence of non-parametric distributions (Peddle 1995) as it
is heavily reliant on the distribution of the training data. Another method includes
the calculation of spectral indices to characterise specific attributes of plants
(DeFries et al. 1995). The most widely used spectral index is the NDVI and is based
on the fact that vegetation is highly reflective in the near infrared and has a high
absorption rate in the visible red wavelengths. The contrast between these wavelengths can be used as a biophysical parameter that correlates with photosynthetic
activity of vegetation, which is an indication of ‘greenness’ (Wang and Tenhunen
2004). Therefore, NDVI is a good indicator to reflect dynamic changes of different
vegetation groups (Geerken et al. 2005). This area of spectral indices has expanded
rapidly and numerous indices are available for different biophysical parameters
(e.g, soil moisture, plant senescence).
Another approach to mapping in comparison to pixel based classification is
object-based image analysis (OBIA). Environmental objects are parts of the real
world for which information can or should become available (e.g, a tree). As soon
as the real world part becomes a formal object in a spatial dimension then it can be
subject to some kind of classification (Bock et al. 2005). However, object based
analysis is very much dependent on the spatial resolution of available imagery. The
follow-up map for the UK LCM (Fuller et al. 2002) was one of the principal drivers
in the development of OBIA in ecological remote sensing.
The advantage of the object-based approach is that it offers new possibilities for
image analysis as image objects can be characterised not only by spectral values but
by texture, shape, context, relationship and thematic information supplied by ancillary data (Bock et al. 2005). Additional layers of knowledge can be valuable to a
Mapping Coastal Habitats in Wales