4.3 Hyperspectral Algorithms
4.3.1 Classification
Currently the majority of practical coral reef mapping from remote sensing is
performed to habitat-level using classification algorithms, such as the k-means or
maximum likelihood algorithms, as implemented in software packages like ENVI
(Exelis VIS 2012) or ERDAS IMAGINE (ERDAS 2011) (Bertels et al. 2007;
Caplosini et al. 2003; Harborne et al. 2006). Classification algorithms seek to
group pixels in an image into a number of classes based on their spectral similarity.
The approach can be unsupervised, the user tells the algorithm how many classes
are desired and the algorithm will automatically try to group spectrally similar
pixels, or supervised, a set of pixels are identified as being of known classes and
the algorithm proceeds on that basis. In either case it is necessary to have some a
priori knowledge of the classes of some pixels (i.e., ‘ground truth’ or ‘calibration’
data). For the unsupervised classification normally two to three times as many
classes as required are requested and the user subsequently identifies what the
classes represent, merging down to the required number. In the absence of field
survey data it is possible for ‘ground truth’ to simply be visual interpretation of the
image combined with local knowledge. It is recommended to retain some ground
truth data for a subsequent independent accuracy assessment (sometimes called
‘validation’). However, caution should be exercised in the selection and structure
of the calibration and validation data. Having more points is not necessarily better
if the structure is such that data from the same areas is simply duplicated across the
calibration and validation data sets, which will inflate the reported accuracy with
respect to uncharacterized areas of the imagery. Contextual editing may also be
used to improve map accuracy; this simply means changing the class of pixels
where visual interpretation indicates they are clearly in error (Mumby et al. 1998).
The level of detail that classification can provide from different imagery sources
has been investigated a number of times (Mumby et al. 1997; Caplosini et al.
2003) and the conclusions are fairly consistent. Multispectral broad band data such
as Landsat TM/ETM+ data cannot derive much more than three or four classes of
coarse scale reef structures, while more and narrower bands may permit up to ten
or more habitat level classes (Fig. 4.5a). ‘Habitat level classes’ means areas
described by phrases such as ‘sparse corals with high algal cover’ or ‘sand with
occasional branching red algae’ (Mumby et al. 1997; Harborne et al. 2006).
Appropriate choices of classes will be site-dependent. Classification techniques
have not been demonstrated as able to map individual reef components such as
‘live coral’ or ‘macroalgae’. This distinction is important to realize; it is often
assumed coral reef remote sensing can map ‘coral cover’ but what can realistically
be achieved is mapping of ‘habitats containing coral’. Nevertheless the level of
detail obtainable from hyperspectral data can be used to derive interesting products, such as the beta-diversity map of Fig. 4.5b. This was produced from a 19band CASI image of U.S. Virgin Islands reefs, which was first classified to 19
4 Hyperspectral Applications
95
4.3.1 Classification
Currently the majority of practical coral reef mapping from remote sensing is
performed to habitat-level using classification algorithms, such as the k-means or
maximum likelihood algorithms, as implemented in software packages like ENVI
(Exelis VIS 2012) or ERDAS IMAGINE (ERDAS 2011) (Bertels et al. 2007;
Caplosini et al. 2003; Harborne et al. 2006). Classification algorithms seek to
group pixels in an image into a number of classes based on their spectral similarity.
The approach can be unsupervised, the user tells the algorithm how many classes
are desired and the algorithm will automatically try to group spectrally similar
pixels, or supervised, a set of pixels are identified as being of known classes and
the algorithm proceeds on that basis. In either case it is necessary to have some a
priori knowledge of the classes of some pixels (i.e., ‘ground truth’ or ‘calibration’
data). For the unsupervised classification normally two to three times as many
classes as required are requested and the user subsequently identifies what the
classes represent, merging down to the required number. In the absence of field
survey data it is possible for ‘ground truth’ to simply be visual interpretation of the
image combined with local knowledge. It is recommended to retain some ground
truth data for a subsequent independent accuracy assessment (sometimes called
‘validation’). However, caution should be exercised in the selection and structure
of the calibration and validation data. Having more points is not necessarily better
if the structure is such that data from the same areas is simply duplicated across the
calibration and validation data sets, which will inflate the reported accuracy with
respect to uncharacterized areas of the imagery. Contextual editing may also be
used to improve map accuracy; this simply means changing the class of pixels
where visual interpretation indicates they are clearly in error (Mumby et al. 1998).
The level of detail that classification can provide from different imagery sources
has been investigated a number of times (Mumby et al. 1997; Caplosini et al.
2003) and the conclusions are fairly consistent. Multispectral broad band data such
as Landsat TM/ETM+ data cannot derive much more than three or four classes of
coarse scale reef structures, while more and narrower bands may permit up to ten
or more habitat level classes (Fig. 4.5a). ‘Habitat level classes’ means areas
described by phrases such as ‘sparse corals with high algal cover’ or ‘sand with
occasional branching red algae’ (Mumby et al. 1997; Harborne et al. 2006).
Appropriate choices of classes will be site-dependent. Classification techniques
have not been demonstrated as able to map individual reef components such as
‘live coral’ or ‘macroalgae’. This distinction is important to realize; it is often
assumed coral reef remote sensing can map ‘coral cover’ but what can realistically
be achieved is mapping of ‘habitats containing coral’. Nevertheless the level of
detail obtainable from hyperspectral data can be used to derive interesting products, such as the beta-diversity map of Fig. 4.5b. This was produced from a 19band CASI image of U.S. Virgin Islands reefs, which was first classified to 19
4 Hyperspectral Applications
95