Imagery analysis
Image data were geo-coded using Ordnance Survey maps (root
mean square error < width of a single image pixel) and radio-
metrically corrected to account for sensor calibration, time of year
and atmospheric conditions (see Price 1987; Tanre et al. 1990).
One of the most commonly cited diculties with remote sensing
of underwater environments is the confounding in¯uence of vari-
able depth on bottom re¯ectance (e.g. Cracknell et al. 1987). For
example, the spectra (spectral signature) of sand at 20 m may be
similar to that of seagrass at (say) 3 m. The eects of variable depth
were compensated using the model derived by Lyzenga (1978, 1981).
For the ®rst ®eld survey, 180 ®eld sites were located on the
imagery. At each site (pixel), ®eld data identi®ed the habitat type
and the overall extent of the habitat. A spectral signature was
created from the image data at each site using the software Erdas
Imagine 8.2. Signatures were developed using the region-growing
tool, which allows neighbouring pixels to be incorporated into the
signature. A geographic constraint was set on this process so that
only those pixels found within the overall extent of each habitat
were selected. For example, if the habitat at a site was considered to
have a diameter of at least 100 m (7860 m2) and a signature was
created for SPOT XS imagery whose pixels cover 400 m2 each, up
to 20 (7860 ¸ 400) pixels were allowed to contribute to the sig-
nature for that site. Pixels further from the ®eld site could not be
expected to represent the same habitat type reliably.
The habitat type at each site was de®ned to its ®nest descriptive
resolution. Individual signatures for each habitat type were then
progressively merged to provide characteristic habitat spectra.
Spectra were then used to train a supervised image classi®cation
which is a multivariate discriminant function (Mather 1987). Pixels
were assigned to habitat classes using the maximum-likelihood
decision rule (Mather 1987). The resulting thematic map of habitats
was evaluated visually and obvious areas of pixel mis-assignment
were identi®ed (e.g. pixels classi®ed as Montastrea spp. reef which
were situated in known seagrass beds). The spectra of habitats
which had over-classi®ed (in this example, Montastrea spp. reef)
were then down-weighted and the supervised image classi®cation
was repeated. All habitat spectra had equal weighting in the ®rst
classi®cation (P 1). Down-weighting was achieved by reducing
the probability that pixels would be assigned to speci®c habitat
classes (in this example, the new probability, P, for Montastrea spp.
was 0.7). This heuristic procedure was repeated and re®ned up to
six times, beyond which no further improvements were noticed by
visual inspection.
The success of a supervised image classi®cation is dependent on
the separability of spectra for dierent habitats in the imagery.
Similar spectra may lead to confusion in the supervised classi®ca-
tion and misclassi®cations in the output-image map. If the sources
of misclassi®cation are known, it is possible to improve map
accuracy by contextual editing (Groom et al. 1996). This process is
perhaps best thought of as ``the application of common sense to
habitat mapping''. Contextual rules may be applied to pairs of
habitats which have similar spectra but exist in dierent, yet pre-
dictable, physical environments, such as seagrass beds and forereef
escarpments. Pixels which classi®ed as seaward patches of seagrass
were reclassi®ed to the appropriate reef categories. Similar reclas-
si®cation was carried out for fringing reef pixels which had been
incorrectly classi®ed as sheltered communities of calci®ed rho-
dophytes with sponge (Class F7, Table 2).