19 Satellite Surveys of Lagoon and Coastal Waters . . .
387
primarily structural parameters (distances, areas, morphological changes), typical
of land features, an accurate spatial registration of the images used is essential for
the application of any change-detection method. This demands the use of geometric
rectification algorithms, which register the images to each other and/or to a standard
map projection. The change-detection algorithms that generated the results shown
here are based upon an unsupervised classification technique (Olmanson et al. 2001).
The basic premise in multi-spectral classification is that the objects being observed
manifest sufficiently different reflectance properties (commonly indicated as spectral signatures) in different regions of the electromagnetic spectrum. Based on these
spectral signatures, natural and artificial surface features can be discriminated and a
new output image is created, identifying a specific number of classes or categories.
The same approach is taken when the observations address parameters related to
water surfaces, by performing first a land-water classification (as described e.g. in
Olmanson et al. 2002). Since water has spectral characteristics different from those
of land features, water pixels can be easily placed into one or a few distinct classes
(notwithstanding the possible classification errors generated by bottom reflection or
floating vegetation). Then, proper masking of terrestrial features allows creating a
water-only image, in which a second classification can be performed on the water pixels alone. Average brightness values from this second classification can be graphed
to show the spectral signature of each class. The various signatures, along with the
location where the pixels occur, can be used to differentiate classes containing clear,
turbid, or shallow water. Based on this information, classes are redefined so that vegetation, bottom and other land effects can be identified, when selecting water pixels
within the lagoon area. In the case of the images shown here, steps were also taken
to identify the type of aquatic vegetation present in the lagoons. The first step was to
categorize the lagoon surface in terms of emerging and submerged vegetation features. The second consisted of two main procedures for aquatic plant discrimination:
separation of image features into discrete units and recognition of the pixels in each
unit. This was accomplished by performing yet another unsupervised classification
(see Olmanson et al. 2008 for details). Vegetation classes were spectrally analyzed to
select 7 different plant species, later identified and named according to field survey
information. The amount and percentage of each plant type was also identified.
The results of the above classifications have been interpreted with the aid of supporting data from topographic maps, aerial photography, local knowledge, and the
spectral and/or temporal properties of the classes themselves, obtained by virtue of
field spectroscopy campaigns. In addition, several geospatial ancillary data sets (e.g.,
digital terrain models, geological maps, land use maps, urban plans, housing census,
population distribution, native vegetation species, agricultural production) were also
used to guide data analysis, collect ground truth and validate the results obtained.
Further, in situ sampling at fixed stations, covering the entire coastal lagoons area,
allowed obtaining complementary data on lake water, sediments and vegetation. The
final database was completed with information on the nature and location of anthropic
activities in the region, urban/suburban runoff from all drainage areas, diffuse pollution sources and runoff volume based on rainfall rates and canal discharges, as
well as erosion and sedimentation sites. This information was finally assembled in
387
primarily structural parameters (distances, areas, morphological changes), typical
of land features, an accurate spatial registration of the images used is essential for
the application of any change-detection method. This demands the use of geometric
rectification algorithms, which register the images to each other and/or to a standard
map projection. The change-detection algorithms that generated the results shown
here are based upon an unsupervised classification technique (Olmanson et al. 2001).
The basic premise in multi-spectral classification is that the objects being observed
manifest sufficiently different reflectance properties (commonly indicated as spectral signatures) in different regions of the electromagnetic spectrum. Based on these
spectral signatures, natural and artificial surface features can be discriminated and a
new output image is created, identifying a specific number of classes or categories.
The same approach is taken when the observations address parameters related to
water surfaces, by performing first a land-water classification (as described e.g. in
Olmanson et al. 2002). Since water has spectral characteristics different from those
of land features, water pixels can be easily placed into one or a few distinct classes
(notwithstanding the possible classification errors generated by bottom reflection or
floating vegetation). Then, proper masking of terrestrial features allows creating a
water-only image, in which a second classification can be performed on the water pixels alone. Average brightness values from this second classification can be graphed
to show the spectral signature of each class. The various signatures, along with the
location where the pixels occur, can be used to differentiate classes containing clear,
turbid, or shallow water. Based on this information, classes are redefined so that vegetation, bottom and other land effects can be identified, when selecting water pixels
within the lagoon area. In the case of the images shown here, steps were also taken
to identify the type of aquatic vegetation present in the lagoons. The first step was to
categorize the lagoon surface in terms of emerging and submerged vegetation features. The second consisted of two main procedures for aquatic plant discrimination:
separation of image features into discrete units and recognition of the pixels in each
unit. This was accomplished by performing yet another unsupervised classification
(see Olmanson et al. 2008 for details). Vegetation classes were spectrally analyzed to
select 7 different plant species, later identified and named according to field survey
information. The amount and percentage of each plant type was also identified.
The results of the above classifications have been interpreted with the aid of supporting data from topographic maps, aerial photography, local knowledge, and the
spectral and/or temporal properties of the classes themselves, obtained by virtue of
field spectroscopy campaigns. In addition, several geospatial ancillary data sets (e.g.,
digital terrain models, geological maps, land use maps, urban plans, housing census,
population distribution, native vegetation species, agricultural production) were also
used to guide data analysis, collect ground truth and validate the results obtained.
Further, in situ sampling at fixed stations, covering the entire coastal lagoons area,
allowed obtaining complementary data on lake water, sediments and vegetation. The
final database was completed with information on the nature and location of anthropic
activities in the region, urban/suburban runoff from all drainage areas, diffuse pollution sources and runoff volume based on rainfall rates and canal discharges, as
well as erosion and sedimentation sites. This information was finally assembled in