still under development; hence sensors with reef specific bands are unlikely to be
launched in the near future, or ever. For now hyperspectral data facilitates
wavelength feature analysis even when only a few wavelengths are required.
One of the most successful demonstrations of measuring a proxy for ‘reef
health’ by a remote sensing study was that of Mumby et al. (2001, 2004) where
derivatives from CASI data were used to discriminate live and dead Porites coral
in a shallow French Polynesian atoll and to quantify live coral among several other
benthic types. In particular the spectral slope 506–565 nm was key to the discrimination of live from dead coral. Isoun et al. (2003) mapped the cover of living
coral at a Hawaiian site using three narrow wavelength bands at 488, 551 and
577 nm. In another Hawaiian example, Hochberg and Atkinson (2000) used an
automated procedure to establish optimal separating wavelengths from in situ
reflectance measurements and then applied them to map coral, algae, and sand. For
these methods depth correction and atmospheric correction is recommended,
especially if a direct comparison to in situ spectra is part of the method. Again
there is little consistency between publications. The site for Mumby et al. (2004)
consisted of very shallow reef (\ 4 m) and a basic depth correction was applied
whereas Isoun et al. (2003) used a form of depth invariant indices. Due to the scant
availability of validated applications wavelength feature methods should be considered the most experimental reef mapping approach and with the most questionable transferability to other sites.
4.3.3 Spectral Unmixing
Reefs are spatially heterogeneous at scales smaller than even the highest resolution
sensors. Spectrally and functionally diverse benthos and substrates such as corals,
macroalgae, rubble and sand can co-exist at sub-meter scales. This mixing clearly
presents a challenge for analysis techniques that attempt to characterize each pixel
as a single class. Spectral unmixing is an approach that attempts to tease apart the
hyperspectral reflectance measurement of a pixel to quantify the proportions of the
individual sub-pixel constituents. Unmixing relies on having a spectral reflectance
library of the ‘pure endmembers’ and the mathematics of the approach assumes the
pixel reflectance is the sum of the reflectances of the components, weighted by
their proportion in the pixel.
The unmixing approach has been successful in mineralogical applications
(Adams et al. 1986) but in mineralogy the endmember reflectances are well defined
and the physical mixing often occurs on a fine scale or on relatively flat surfaces.
In contrast for coral reefs endmember spectra are not well defined (Fig. 4.3),
the overlying water column complicates the reflectance, and the three-dimensional
structure of the reef may invalidate the linear mixing assumption (Hedley 2008).
Nevertheless modeling and experimental studies indicate unmixing has potential
(Hedley and Mumby 2003; Hedley et al. 2004) and only a few published
image analyses have included an unmixing component (Goodman et al 2007;
98
J. D. Hedley
launched in the near future, or ever. For now hyperspectral data facilitates
wavelength feature analysis even when only a few wavelengths are required.
One of the most successful demonstrations of measuring a proxy for ‘reef
health’ by a remote sensing study was that of Mumby et al. (2001, 2004) where
derivatives from CASI data were used to discriminate live and dead Porites coral
in a shallow French Polynesian atoll and to quantify live coral among several other
benthic types. In particular the spectral slope 506–565 nm was key to the discrimination of live from dead coral. Isoun et al. (2003) mapped the cover of living
coral at a Hawaiian site using three narrow wavelength bands at 488, 551 and
577 nm. In another Hawaiian example, Hochberg and Atkinson (2000) used an
automated procedure to establish optimal separating wavelengths from in situ
reflectance measurements and then applied them to map coral, algae, and sand. For
these methods depth correction and atmospheric correction is recommended,
especially if a direct comparison to in situ spectra is part of the method. Again
there is little consistency between publications. The site for Mumby et al. (2004)
consisted of very shallow reef (\ 4 m) and a basic depth correction was applied
whereas Isoun et al. (2003) used a form of depth invariant indices. Due to the scant
availability of validated applications wavelength feature methods should be considered the most experimental reef mapping approach and with the most questionable transferability to other sites.
4.3.3 Spectral Unmixing
Reefs are spatially heterogeneous at scales smaller than even the highest resolution
sensors. Spectrally and functionally diverse benthos and substrates such as corals,
macroalgae, rubble and sand can co-exist at sub-meter scales. This mixing clearly
presents a challenge for analysis techniques that attempt to characterize each pixel
as a single class. Spectral unmixing is an approach that attempts to tease apart the
hyperspectral reflectance measurement of a pixel to quantify the proportions of the
individual sub-pixel constituents. Unmixing relies on having a spectral reflectance
library of the ‘pure endmembers’ and the mathematics of the approach assumes the
pixel reflectance is the sum of the reflectances of the components, weighted by
their proportion in the pixel.
The unmixing approach has been successful in mineralogical applications
(Adams et al. 1986) but in mineralogy the endmember reflectances are well defined
and the physical mixing often occurs on a fine scale or on relatively flat surfaces.
In contrast for coral reefs endmember spectra are not well defined (Fig. 4.3),
the overlying water column complicates the reflectance, and the three-dimensional
structure of the reef may invalidate the linear mixing assumption (Hedley 2008).
Nevertheless modeling and experimental studies indicate unmixing has potential
(Hedley and Mumby 2003; Hedley et al. 2004) and only a few published
image analyses have included an unmixing component (Goodman et al 2007;
98
J. D. Hedley