Other notable mapping examples covering a large spatial extent include
22,000 km
2 of seagrass mapping in Shark Bay, Western Australia (Bruce 1997)
and a 345,400 km
2 coral reef habitat map based on Landsat TM data for the Great
Barrier Reef, Australia by the Great Barrier Reef Marine Park Authority (Lewis
et al. 2003). However, despite their scale, these last two studies were not included
in the literature review as they are not published in readily accessible scientific
journals or reports.
This review of mapping approaches was followed by an experiment where eight
different mapping approaches were applied to create habitat maps for three different reef environments. The mapping approaches incorporated different combinations of remote sensing image type, calibration/validation method, and the
processing method applied. The approaches were assessed according to cost, time,
and the knowledge and skills needed to produce each map. The final map products
were analyzed with respect to perceived relevance by map users (Roelfsema et al.
2008). Results demonstrated how variations in coral reef environments, and differences in intended map purpose, affected the user’s selection of a suitable
mapping approach.
14.3.2 Sampling Design
Of the 80 papers reviewed, 19 used a sampling scheme based on stratified random
sampling, 16 used a stratified and non-random sampling, 5 used simple random
sampling, and the remaining 32 papers did not provide sufficient information to
determine the type of sampling scheme used (Table 14.4). Of the 40 papers that
reported sampling schemes, maps presenting the location of the calibration and
validation samples were included in only 36. Ideally, the number and distribution
of samples would be determined based on the statistical rules previously suggested
(Sect. 14.2.1); however, that is not always practical due to logistical challenges. It
was therefore not surprising that the sampling schemes were mostly stratified, with
fewer sample locations used than was statistically required, which is the inherent
trade off when faced with having to overcome the logistical challenges.
A total of 15 papers reported that they used more than 50 recommended validation samples per mapping category, 28 reported using less than 50, and the
remaining 37 did not provide the number of samples. The statistically optimum
number of required sample points per mapping category cannot always be
achieved, particularly for mapping categories that cover small areas and are thus
harder to actually find for sampling purposes. This was the case for Chauvaud et al.
(1998) who created a map with 32 categories over an area extent of 240 km
2 with
95.7 % overall accuracy. Statistically at least 1,600 sampling points should have
been visited (50 for each category), however it was not possible to identify sufficient sample areas for every category and just 111 samples were assessed.
386
C. M. Roelfsema and S. R. Phinn
22,000 km
2 of seagrass mapping in Shark Bay, Western Australia (Bruce 1997)
and a 345,400 km
2 coral reef habitat map based on Landsat TM data for the Great
Barrier Reef, Australia by the Great Barrier Reef Marine Park Authority (Lewis
et al. 2003). However, despite their scale, these last two studies were not included
in the literature review as they are not published in readily accessible scientific
journals or reports.
This review of mapping approaches was followed by an experiment where eight
different mapping approaches were applied to create habitat maps for three different reef environments. The mapping approaches incorporated different combinations of remote sensing image type, calibration/validation method, and the
processing method applied. The approaches were assessed according to cost, time,
and the knowledge and skills needed to produce each map. The final map products
were analyzed with respect to perceived relevance by map users (Roelfsema et al.
2008). Results demonstrated how variations in coral reef environments, and differences in intended map purpose, affected the user’s selection of a suitable
mapping approach.
14.3.2 Sampling Design
Of the 80 papers reviewed, 19 used a sampling scheme based on stratified random
sampling, 16 used a stratified and non-random sampling, 5 used simple random
sampling, and the remaining 32 papers did not provide sufficient information to
determine the type of sampling scheme used (Table 14.4). Of the 40 papers that
reported sampling schemes, maps presenting the location of the calibration and
validation samples were included in only 36. Ideally, the number and distribution
of samples would be determined based on the statistical rules previously suggested
(Sect. 14.2.1); however, that is not always practical due to logistical challenges. It
was therefore not surprising that the sampling schemes were mostly stratified, with
fewer sample locations used than was statistically required, which is the inherent
trade off when faced with having to overcome the logistical challenges.
A total of 15 papers reported that they used more than 50 recommended validation samples per mapping category, 28 reported using less than 50, and the
remaining 37 did not provide the number of samples. The statistically optimum
number of required sample points per mapping category cannot always be
achieved, particularly for mapping categories that cover small areas and are thus
harder to actually find for sampling purposes. This was the case for Chauvaud et al.
(1998) who created a map with 32 categories over an area extent of 240 km
2 with
95.7 % overall accuracy. Statistically at least 1,600 sampling points should have
been visited (50 for each category), however it was not possible to identify sufficient sample areas for every category and just 111 samples were assessed.
386
C. M. Roelfsema and S. R. Phinn