Since the Lee et al. (1998, 1999) original publications numerous other workers
have devised their own variants of the semi-analytical algorithm. Wettle and Brando
(2006) use ‘specific inherent optical properties’ (SIOPs), where water samples from
the site are required to characterize the concentration-dependent, or ‘specific’,
spectral absorptions of water constituents. The suggested advantages are two-fold:
(1) the method is calibrated for local water constituents, and (2) the outputs are actual
constituent concentrations whereas the original formulation of Lee et al. (1999)
returned concentration proxies that evade direct interpretation. Wettle and Brando
(2006) also introduced a linear mixing model for substrate reflectance; this concept
has been used by several other workers (Klonowski et al. 2007; Hedley et al. 2009a).
A number of these method variants were tested in a cross-comparison exercise with a
Caribbean and Australian data set (Dekker et al. 2011). It should be noted that these
methods require high quality atmospheric correction (Goodman et al. 2008).
4.4 Conclusions
The remote sensing community in general increasingly recognizes the need to
provide confidence intervals on remotely sensed products. For classification
approaches on reefs the normal practice is to describe overall accuracy in terms of
misclassification rates, but this does not express confidence at the pixel level and is
based on regions where ground truth data is available. In reality inaccuracies may
be dependent on the location in the image. For example, determination of benthic
type will be increasingly uncertain in deeper water. Inaccuracies also result from
the various assumptions and simplifications inherent to the calibration, correction
and image processing workflow (e.g., atmospheric correction, sunglint correction,
radiative transfer modeling, etc.). It is important to gain an understanding of how
these inaccuracies are manifest in the classification and mapping output. Analysis
methods based on wavelength features or model inversion will always produce ‘an
answer’ at each pixel, even if there is almost no information to extract. Confidence
in the output values of any method may vary widely across an image, and this is
especially true in coral reef applications where the environment displays heterogeneity in multiple aspects: depth, water constituents, and sea surface state (e.g.,
inside and outside of a lagoon).
One way to characterize uncertainty is to consider the noise-equivalent deltareflectance, NEDR (Brando and Dekker 2003). In shallow water applications,
perturbation in pixel-to-pixel spectral reflectance caused by the top of the water
surface, atmospheric turbulence, and sensor noise can be grouped to ‘environmental noise’ and evaluated from the band-wise standard deviation over an area of
homogenous deep water in an image (Brando and Dekker 2003). Brando et al.
(2009) used this measure directly to provide pixel quality assurance in a shallow
water model inversion, excluding pixels where depth was such that the reflectance
was below a NEDR determined threshold. However, this requires absolute depth to
104
J. D. Hedley
have devised their own variants of the semi-analytical algorithm. Wettle and Brando
(2006) use ‘specific inherent optical properties’ (SIOPs), where water samples from
the site are required to characterize the concentration-dependent, or ‘specific’,
spectral absorptions of water constituents. The suggested advantages are two-fold:
(1) the method is calibrated for local water constituents, and (2) the outputs are actual
constituent concentrations whereas the original formulation of Lee et al. (1999)
returned concentration proxies that evade direct interpretation. Wettle and Brando
(2006) also introduced a linear mixing model for substrate reflectance; this concept
has been used by several other workers (Klonowski et al. 2007; Hedley et al. 2009a).
A number of these method variants were tested in a cross-comparison exercise with a
Caribbean and Australian data set (Dekker et al. 2011). It should be noted that these
methods require high quality atmospheric correction (Goodman et al. 2008).
4.4 Conclusions
The remote sensing community in general increasingly recognizes the need to
provide confidence intervals on remotely sensed products. For classification
approaches on reefs the normal practice is to describe overall accuracy in terms of
misclassification rates, but this does not express confidence at the pixel level and is
based on regions where ground truth data is available. In reality inaccuracies may
be dependent on the location in the image. For example, determination of benthic
type will be increasingly uncertain in deeper water. Inaccuracies also result from
the various assumptions and simplifications inherent to the calibration, correction
and image processing workflow (e.g., atmospheric correction, sunglint correction,
radiative transfer modeling, etc.). It is important to gain an understanding of how
these inaccuracies are manifest in the classification and mapping output. Analysis
methods based on wavelength features or model inversion will always produce ‘an
answer’ at each pixel, even if there is almost no information to extract. Confidence
in the output values of any method may vary widely across an image, and this is
especially true in coral reef applications where the environment displays heterogeneity in multiple aspects: depth, water constituents, and sea surface state (e.g.,
inside and outside of a lagoon).
One way to characterize uncertainty is to consider the noise-equivalent deltareflectance, NEDR (Brando and Dekker 2003). In shallow water applications,
perturbation in pixel-to-pixel spectral reflectance caused by the top of the water
surface, atmospheric turbulence, and sensor noise can be grouped to ‘environmental noise’ and evaluated from the band-wise standard deviation over an area of
homogenous deep water in an image (Brando and Dekker 2003). Brando et al.
(2009) used this measure directly to provide pixel quality assurance in a shallow
water model inversion, excluding pixels where depth was such that the reflectance
was below a NEDR determined threshold. However, this requires absolute depth to
104
J. D. Hedley