103
focused on a multidisciplinary approach to bridge the gap between remote sensing
scientists and ecologists, allowing funding to be used effectively for both monitoring and management practices. The project also developed the Earth Observation
for Habitat Monitoring (EODHaM) system (Lucas et al. 2015), which provided a
standardised framework for consistent land cover and habitat mapping and for monitoring Natura 2000 sites. A key component of the system is the inclusion of decision rules within a hierarchical classification structure generated by experienced
ecologists and remote sensing scientists, and the system’s use of the Land Cover
Classification System (LCCS). The LCCS was developed by the Food and
Agricultural Association (FAO) and represents a common classification scheme that
can be used anywhere on global to local scales (Fig. 3).
The EODHaM system was utilised for the first stage of mapping (Fig. 4) as it
allows the use of multiple data sources and types and is run within open source
software that is freely available.
For the first stage of classification (LCCS level 3), a rule-based classification
based primarily on thresholds of spectral data and indices is used. First, vegetated
areas were separated from non-vegetated areas, which created a classification with
an overall accuracy of 98%. All discrepancies encountered were explained by the
presence of short sparse vegetation in primarily sandy areas known as dune annuals
communities, shifting dune and semi-fixed dune habitat. However, as the spectral
diversity within the vegetated areas is high, a more robust method was developed to
select the data layers that would provide the best separation for distinguishing the
more complicated classes such as lifeform.
An Analysis of Variance (ANOVA) was performed on all possible data layers,
from the 134 indices calculated per Worldview-2 scene to the DTM and derived
measures such as slope. The training dataset allowed analyses to test if a significant
difference (at ρ < 0.05) existed between data layers for each lifeform category. The
ANOVA coefficient or F-ratio is an extension of Fisher’s discriminant and provides
a measure of separability between multiple classes (Scheffe 1959). The magnitude
and significance level associated with each ANOVA F value were used to determine
the most effective layers in separating categories where the larger the F value, the
more likely it is that the null hypothesis of no differences between group means is
false, indicating greater separation. Although, one of the assumptions of this method
is that data are normally distributed, numerous studies have demonstrated that the
data’s distribution has very little effect on the F-ratio (Tiku 1971; Box and Watson
1961). Once the best layer or layers were selected, boxplots were then used to determine the optimal threshold values for classification.
An example of a successful category classified using this approach is the woody
(trees) lifeform class. The class is separable by using two data layers and was classified with an overall accuracy of 84%. However, some classes were deemed inseparable based on ANOVA as the distribution of classes in n-dimensional space
becomes too complicated Therefore, alternative methods of classification were considered, and described in the next section (Figs. 5 and 6).
Mapping Coastal Habitats in Wales
focused on a multidisciplinary approach to bridge the gap between remote sensing
scientists and ecologists, allowing funding to be used effectively for both monitoring and management practices. The project also developed the Earth Observation
for Habitat Monitoring (EODHaM) system (Lucas et al. 2015), which provided a
standardised framework for consistent land cover and habitat mapping and for monitoring Natura 2000 sites. A key component of the system is the inclusion of decision rules within a hierarchical classification structure generated by experienced
ecologists and remote sensing scientists, and the system’s use of the Land Cover
Classification System (LCCS). The LCCS was developed by the Food and
Agricultural Association (FAO) and represents a common classification scheme that
can be used anywhere on global to local scales (Fig. 3).
The EODHaM system was utilised for the first stage of mapping (Fig. 4) as it
allows the use of multiple data sources and types and is run within open source
software that is freely available.
For the first stage of classification (LCCS level 3), a rule-based classification
based primarily on thresholds of spectral data and indices is used. First, vegetated
areas were separated from non-vegetated areas, which created a classification with
an overall accuracy of 98%. All discrepancies encountered were explained by the
presence of short sparse vegetation in primarily sandy areas known as dune annuals
communities, shifting dune and semi-fixed dune habitat. However, as the spectral
diversity within the vegetated areas is high, a more robust method was developed to
select the data layers that would provide the best separation for distinguishing the
more complicated classes such as lifeform.
An Analysis of Variance (ANOVA) was performed on all possible data layers,
from the 134 indices calculated per Worldview-2 scene to the DTM and derived
measures such as slope. The training dataset allowed analyses to test if a significant
difference (at ρ < 0.05) existed between data layers for each lifeform category. The
ANOVA coefficient or F-ratio is an extension of Fisher’s discriminant and provides
a measure of separability between multiple classes (Scheffe 1959). The magnitude
and significance level associated with each ANOVA F value were used to determine
the most effective layers in separating categories where the larger the F value, the
more likely it is that the null hypothesis of no differences between group means is
false, indicating greater separation. Although, one of the assumptions of this method
is that data are normally distributed, numerous studies have demonstrated that the
data’s distribution has very little effect on the F-ratio (Tiku 1971; Box and Watson
1961). Once the best layer or layers were selected, boxplots were then used to determine the optimal threshold values for classification.
An example of a successful category classified using this approach is the woody
(trees) lifeform class. The class is separable by using two data layers and was classified with an overall accuracy of 84%. However, some classes were deemed inseparable based on ANOVA as the distribution of classes in n-dimensional space
becomes too complicated Therefore, alternative methods of classification were considered, and described in the next section (Figs. 5 and 6).
Mapping Coastal Habitats in Wales