96
classification system especially when certain habitat types do not have very distinct
spectral features. Many segmentation algorithms are not adapted to detect the variety of geographical entities comprising a complex scene and while they perform
well in delineating some landscape objects, this is rarely true for all objects of interest (Marceau et al. 1994). The issue of scale is problematic for segmentation algorithms as segments in an image will never represent meaningful objects at all scales
which is required for many applications, although multi-scale segmentation
approaches may be able to combat this (Blaschke et al. 2001).
Traditional methods are able to learn automatically from clustering or training
data, while knowledge based classifiers require user-defined thresholds to determine
class relationship. While training data is not necessarily required, expert knowledge
of the area of interest is needed, which is why knowledge based classifiers require
information from ecologists for vegetation characterisation. The use of extensive
field knowledge and auxiliary data and the use of empirical rules to extract thematic
features has proved successful, and can even improve classification accuracy (Gad
and Kusky 2006; Shrestha and Zinck 2001). The most common knowledge-based
classifiers are therefore, rule-based, and are commonly accompanied by image segmentation and OBIA approaches. Lucas et al. (2011) applied a rule-based approach
to update the Phase 1 habitat of Wales, by utilising satellite imagery and ancillary
data. Approaches like this enable full user control but gathering specific knowledge
and obtaining ancillary data is often seen as an enormous task and can be very costly
and too time consuming (Xie et al. 2008). However, in countries such as the UK,
much of this data is already acquired by conservation bodies, and better communication between the remote sensing and ecological communities is encouraged
(Lucas et al. 2007).
In recent years, when faced with large dimensional and complex data spaces,
machine-learning algorithms have emerged as more accurate and efficient alternatives to the traditional classification techniques used within the remote sensing community (Roudriguez-Galiano et al. 2012). The algorithms used typically involve
statistical pattern recognition, the theory of which was mostly developed in the
1960s and 1970s. Some major developments include the Bayes decision theory
problem (Chow 1957), nearest neighbour decision rules (Cover and Hart 1967) and
supervised and unsupervised learning (Fukunaga 1990). During the latter part of the
1980s, artificial neural networks and support vector machines were developed as
statistical classifiers and these had a significant impact on the remote sensing community (Bishop 2006). The early assumption that each pixel in a multispectral
image has a histogram with an approximate Gaussian distribution was made by Fu
(1982) and became most popular in classifying multispectral data with use of the
maximum likelihood algorithm for example. Even with the development of new
sensors and the expanded applications of remote sensing, the Gaussian assumption
remains a good approximation (Chen and Ho 2008), which explains their popularity. However, most parametric classifiers are heavily reliant on this assumption and
data may not always represent normal distributions.
A variety of nonparametric machine algorithms exist including k-Nearest
Neighbour (Gabroswki et al. 2003), Bagging (Breiman 1996), Adaboost (Freund
G. Jones et al.
classification system especially when certain habitat types do not have very distinct
spectral features. Many segmentation algorithms are not adapted to detect the variety of geographical entities comprising a complex scene and while they perform
well in delineating some landscape objects, this is rarely true for all objects of interest (Marceau et al. 1994). The issue of scale is problematic for segmentation algorithms as segments in an image will never represent meaningful objects at all scales
which is required for many applications, although multi-scale segmentation
approaches may be able to combat this (Blaschke et al. 2001).
Traditional methods are able to learn automatically from clustering or training
data, while knowledge based classifiers require user-defined thresholds to determine
class relationship. While training data is not necessarily required, expert knowledge
of the area of interest is needed, which is why knowledge based classifiers require
information from ecologists for vegetation characterisation. The use of extensive
field knowledge and auxiliary data and the use of empirical rules to extract thematic
features has proved successful, and can even improve classification accuracy (Gad
and Kusky 2006; Shrestha and Zinck 2001). The most common knowledge-based
classifiers are therefore, rule-based, and are commonly accompanied by image segmentation and OBIA approaches. Lucas et al. (2011) applied a rule-based approach
to update the Phase 1 habitat of Wales, by utilising satellite imagery and ancillary
data. Approaches like this enable full user control but gathering specific knowledge
and obtaining ancillary data is often seen as an enormous task and can be very costly
and too time consuming (Xie et al. 2008). However, in countries such as the UK,
much of this data is already acquired by conservation bodies, and better communication between the remote sensing and ecological communities is encouraged
(Lucas et al. 2007).
In recent years, when faced with large dimensional and complex data spaces,
machine-learning algorithms have emerged as more accurate and efficient alternatives to the traditional classification techniques used within the remote sensing community (Roudriguez-Galiano et al. 2012). The algorithms used typically involve
statistical pattern recognition, the theory of which was mostly developed in the
1960s and 1970s. Some major developments include the Bayes decision theory
problem (Chow 1957), nearest neighbour decision rules (Cover and Hart 1967) and
supervised and unsupervised learning (Fukunaga 1990). During the latter part of the
1980s, artificial neural networks and support vector machines were developed as
statistical classifiers and these had a significant impact on the remote sensing community (Bishop 2006). The early assumption that each pixel in a multispectral
image has a histogram with an approximate Gaussian distribution was made by Fu
(1982) and became most popular in classifying multispectral data with use of the
maximum likelihood algorithm for example. Even with the development of new
sensors and the expanded applications of remote sensing, the Gaussian assumption
remains a good approximation (Chen and Ho 2008), which explains their popularity. However, most parametric classifiers are heavily reliant on this assumption and
data may not always represent normal distributions.
A variety of nonparametric machine algorithms exist including k-Nearest
Neighbour (Gabroswki et al. 2003), Bagging (Breiman 1996), Adaboost (Freund
G. Jones et al.