that ideally will allow class merging (Mumby and Harbourne 1999). To some
extent this will depend on available ground truth data. For classification approaches it is important to perform sunglint correction if glint patterns are visible in the
imagery (Sect. 4.2.5) but the majority of atmospheric correction schemes will
make little or no difference, and in practice atmospheric correction is often
omitted. Calculation of depth invariant indices (Sect. 4.2.6) can increase accuracy,
but strictly speaking if ground truth data is available across the full range of depths
this should be unnecessary and in fact the depth may carry habitat information.
Note that working with more classes demands more detailed ground truth data for
calibration and validation. So there may be associated costs in fully exploiting
hyperspectral data that extend beyond the cost of data acquisition alone.
4.3.2 Band-Specific Analysis
A number of coral reef remote sensing studies have investigated the concept of
mapping benthic types such as live coral or algae by evaluating band differences,
ratios, or ‘derivatives’, the slope of the spectral profile (Tsai and Philpot 1998), at
specific wavelengths. These ‘wavelength feature’ approaches offer the potential to
map reefs to the level of individual reef components, rather than that of habitats as
typically elucidated by classification. While actual remote sensing applications
that have successfully used wavelength features for mapping benthic type are
scarce and have used differing methodologies, when successful they have produced results at the fore-front of coral reef remote sensing capability (Hochberg
and Atkinson 2000; Isoun et al. 2003; Mumby et al. 2001, 2004).
The basis of using wavelength features is informed by analysis of libraries of
in situ reflectance profiles, where the spectral reflectances of benthic types are
taken using a spectroradiometer either underwater or exposed (i.e., ‘field spectroscopy’ as it is termed in terrestrial environments). For example, Holden and
LeDrew (1999) took a number of spectral reflectance measurements of live and
bleached coral and identified three wavelength regions between 500 and 650 nm
that could be used for discrimination. Hochberg and Atkinson (2000, 2003c) and
Hochberg et al. (2003a) have extensively studied in situ spectral libraries and
identified wavelength regions where discrimination of benthic types may be
possible (see also Wettle et al. 2003). Hedley and Mumby (2002) review a number
of previous studies on the same theme and attempt to establish a causal biological
link to the observed spectral features in terms of pigments.
To use wavelength specific features for mapping requires hyperspectral data,
first because a free choice is required of which wavelengths to use, and secondly
because narrow bands are necessary to elucidate the spectral features. Effectively,
data in wavelength regions that are not of interest are discarded. In ocean color
remote sensing, important wavelength regions for phytoplankton pigments have
long been established; so operational satellite sensors such as MERIS have narrow
bands located at these wavelengths. Coral reef remote sensing is a dynamic field
4 Hyperspectral Applications
97
extent this will depend on available ground truth data. For classification approaches it is important to perform sunglint correction if glint patterns are visible in the
imagery (Sect. 4.2.5) but the majority of atmospheric correction schemes will
make little or no difference, and in practice atmospheric correction is often
omitted. Calculation of depth invariant indices (Sect. 4.2.6) can increase accuracy,
but strictly speaking if ground truth data is available across the full range of depths
this should be unnecessary and in fact the depth may carry habitat information.
Note that working with more classes demands more detailed ground truth data for
calibration and validation. So there may be associated costs in fully exploiting
hyperspectral data that extend beyond the cost of data acquisition alone.
4.3.2 Band-Specific Analysis
A number of coral reef remote sensing studies have investigated the concept of
mapping benthic types such as live coral or algae by evaluating band differences,
ratios, or ‘derivatives’, the slope of the spectral profile (Tsai and Philpot 1998), at
specific wavelengths. These ‘wavelength feature’ approaches offer the potential to
map reefs to the level of individual reef components, rather than that of habitats as
typically elucidated by classification. While actual remote sensing applications
that have successfully used wavelength features for mapping benthic type are
scarce and have used differing methodologies, when successful they have produced results at the fore-front of coral reef remote sensing capability (Hochberg
and Atkinson 2000; Isoun et al. 2003; Mumby et al. 2001, 2004).
The basis of using wavelength features is informed by analysis of libraries of
in situ reflectance profiles, where the spectral reflectances of benthic types are
taken using a spectroradiometer either underwater or exposed (i.e., ‘field spectroscopy’ as it is termed in terrestrial environments). For example, Holden and
LeDrew (1999) took a number of spectral reflectance measurements of live and
bleached coral and identified three wavelength regions between 500 and 650 nm
that could be used for discrimination. Hochberg and Atkinson (2000, 2003c) and
Hochberg et al. (2003a) have extensively studied in situ spectral libraries and
identified wavelength regions where discrimination of benthic types may be
possible (see also Wettle et al. 2003). Hedley and Mumby (2002) review a number
of previous studies on the same theme and attempt to establish a causal biological
link to the observed spectral features in terms of pigments.
To use wavelength specific features for mapping requires hyperspectral data,
first because a free choice is required of which wavelengths to use, and secondly
because narrow bands are necessary to elucidate the spectral features. Effectively,
data in wavelength regions that are not of interest are discarded. In ocean color
remote sensing, important wavelength regions for phytoplankton pigments have
long been established; so operational satellite sensors such as MERIS have narrow
bands located at these wavelengths. Coral reef remote sensing is a dynamic field
4 Hyperspectral Applications
97