Digital processing includes image classification or segmentation based on
spectral characteristics (i.e., digital number, radiance, or reflectance) and, in some
cases, texture. Spectral processing discriminates features based solely on multispectral signatures, while texture approaches also incorporate spatial variability of
the multispectral signatures. Unsupervised classification analyzes images without
user input, and then the different segments are assigned to a given benthic category, or class, according to expert knowledge of the user. In supervised classification, ground-truth for each class (i.e., user supplied input) is used to train the
classification scheme and identify these classes throughout the image. Classification using texture considers spatial patterns as a function of spectral variation
within a particular area. Some substrate types (e.g., corals and macroalgae) may be
spectrally similar but retain distinct texture properties by virtue of a systematic
variance in spectral characteristics at a discrete spatial scale. Therefore, the
boundary of ecotones can be detected (Andréfouët and Roux 1998), and analysis of
the texture may thus enhance detection or classification accuracy (LeDrew et al.
2004; Purkis et al. 2006; Lim et al. 2009). Further, high spatial resolution is
preferable to observe texture, and thus pan-sharpening of the image (Hanaizumi
et al. 2008) may be helpful for improving classification output using texture.
Contextual editing. Some misclassification of habitat categories is inevitable in
the digitally classified image. Accordingly, classification results can be edited to
take account of known patterns of habitat distribution (Mumby et al. 1998). For
example, pixels classified as seagrass, but present on a forereef slope where it is
known that seagrass is absent, can be reclassified to the appropriate reef categories
(Fig. 3.1). In addition to this type of a posteriori contextual editing, a priori
contextual editing is another attractive option (Andréfouët 2008). As indicated in
Bouvet et al. (2003) and Andréfouët et al. (2003), accuracy is enhanced if different
areas containing different habitats are processed separately. In these situations, it is
recommended that this a priori segmentation be based on geomorphology, because
geomorphological zonation is related to depth and wave exposure, which also
affects the distribution of habitats.
3.2.3 Time-Series Analysis
The recent rapid decline of coral reefs (e.g., Gardner et al. 2003) raises the
importance of time-series analysis of historical images to examine the timing and
extent of this decline. Two approaches have been proposed for change detection
using multi-temporal images (Lunetta and Elvidge 1998), post-classification and
pre-classification. The post-classification approach involves the analysis of differences among multi-temporal output products (e.g., habitat maps). The change
detection accuracy between the multi-temporal maps is a product of the accuracies
of the two maps. Therefore, availability of in situ data for past images is essential
to achieve high accuracy for change detection. If only recent in situ data is
62
H. Yamano
spectral characteristics (i.e., digital number, radiance, or reflectance) and, in some
cases, texture. Spectral processing discriminates features based solely on multispectral signatures, while texture approaches also incorporate spatial variability of
the multispectral signatures. Unsupervised classification analyzes images without
user input, and then the different segments are assigned to a given benthic category, or class, according to expert knowledge of the user. In supervised classification, ground-truth for each class (i.e., user supplied input) is used to train the
classification scheme and identify these classes throughout the image. Classification using texture considers spatial patterns as a function of spectral variation
within a particular area. Some substrate types (e.g., corals and macroalgae) may be
spectrally similar but retain distinct texture properties by virtue of a systematic
variance in spectral characteristics at a discrete spatial scale. Therefore, the
boundary of ecotones can be detected (Andréfouët and Roux 1998), and analysis of
the texture may thus enhance detection or classification accuracy (LeDrew et al.
2004; Purkis et al. 2006; Lim et al. 2009). Further, high spatial resolution is
preferable to observe texture, and thus pan-sharpening of the image (Hanaizumi
et al. 2008) may be helpful for improving classification output using texture.
Contextual editing. Some misclassification of habitat categories is inevitable in
the digitally classified image. Accordingly, classification results can be edited to
take account of known patterns of habitat distribution (Mumby et al. 1998). For
example, pixels classified as seagrass, but present on a forereef slope where it is
known that seagrass is absent, can be reclassified to the appropriate reef categories
(Fig. 3.1). In addition to this type of a posteriori contextual editing, a priori
contextual editing is another attractive option (Andréfouët 2008). As indicated in
Bouvet et al. (2003) and Andréfouët et al. (2003), accuracy is enhanced if different
areas containing different habitats are processed separately. In these situations, it is
recommended that this a priori segmentation be based on geomorphology, because
geomorphological zonation is related to depth and wave exposure, which also
affects the distribution of habitats.
3.2.3 Time-Series Analysis
The recent rapid decline of coral reefs (e.g., Gardner et al. 2003) raises the
importance of time-series analysis of historical images to examine the timing and
extent of this decline. Two approaches have been proposed for change detection
using multi-temporal images (Lunetta and Elvidge 1998), post-classification and
pre-classification. The post-classification approach involves the analysis of differences among multi-temporal output products (e.g., habitat maps). The change
detection accuracy between the multi-temporal maps is a product of the accuracies
of the two maps. Therefore, availability of in situ data for past images is essential
to achieve high accuracy for change detection. If only recent in situ data is
62
H. Yamano