50
Jane Elith
The full set of 391 validation sites was spread throughout the whole RFA,
approximately 1.1 million ha in area. It has been shown (Elith and Burgman, in
review) that discrimination between sites located across the whole region may be
good even though discrimination between sites at a finer scale is poor. In this
context, “finer scale” could be understood either in a geographic sense (compare
sites within a subcatchment rather than within a whole region) or in an ecological
sense (compare sites that are at least within possible ecological bounds rather than
sites that are clearly outside the environmental tolerances of the species). Earlier
publications of these data investigated the second concept (Elith et al. 1998),
whereas here we focus on the first interpretation and create different validation
sets for each species. The locations of the species in the full modeling set (3,522
presence-absence sites) were taken to indicate the geographic extent of the species
in the RFA, and a mask was created that enclosed most presence locations and
extended to the estimated bounds of the inhabited subcatchments. The masks were
created through a visual interpretation of the region’s digital elevation model and
were intended to select subcatchments that might be expected to contain the
species. The final validation sets for the seven species reported here vary in size
from 78 to 161 sites, with species frequencies ranging from 0.07 to 0.62 (Table
4.2).
Results
Data presented here (Fig. 4.2) refer to observations of the presence of seven
species within the 250-m cells (i.e., to the first two species types, as described
under the heading Validation). The data for these species are representative of the
data for all 29 species evaluated in the larger study (Elith et al. 1998). The
majority of models do not discriminate well at a fine (subcatchment) scale. The
predictions of any one method were not consistently better than the others, and
there was no significant difference (p < .05) between the mean ROC areas for any
of the methods across the seven species. In the larger data set, GAMs were more
frequently represented among the models with ROC area greater than 0.75 than
any other method (Elith et al. 1998). Species that occurred frequently in the
original 3,522 quadrats were no better modeled than species that were relatively
scarce.
Modeling success varied between species. Three species for which sufficiently
discriminatory models could not be found were Tetratheca stenocarpa, Wittsteinia
vacciniacea, and Phebalium bilobum (Fig. 4.2). These models were based on 54,
131, and 119 records, respectively. The species with the least successful models,
Wittsteinia vacciniacea, is confined to high rainfall montane and subalpine areas.
Although it is ubiquitous in some patches in the region, there are other areas where
suitable habitat is restricted, and the available variables did not adequately
describe the important variation in the environment. The “successful” models
were for Grevillea barklyana (all methods), Helichrysum scorpioides and Lep-
Jane Elith
The full set of 391 validation sites was spread throughout the whole RFA,
approximately 1.1 million ha in area. It has been shown (Elith and Burgman, in
review) that discrimination between sites located across the whole region may be
good even though discrimination between sites at a finer scale is poor. In this
context, “finer scale” could be understood either in a geographic sense (compare
sites within a subcatchment rather than within a whole region) or in an ecological
sense (compare sites that are at least within possible ecological bounds rather than
sites that are clearly outside the environmental tolerances of the species). Earlier
publications of these data investigated the second concept (Elith et al. 1998),
whereas here we focus on the first interpretation and create different validation
sets for each species. The locations of the species in the full modeling set (3,522
presence-absence sites) were taken to indicate the geographic extent of the species
in the RFA, and a mask was created that enclosed most presence locations and
extended to the estimated bounds of the inhabited subcatchments. The masks were
created through a visual interpretation of the region’s digital elevation model and
were intended to select subcatchments that might be expected to contain the
species. The final validation sets for the seven species reported here vary in size
from 78 to 161 sites, with species frequencies ranging from 0.07 to 0.62 (Table
4.2).
Results
Data presented here (Fig. 4.2) refer to observations of the presence of seven
species within the 250-m cells (i.e., to the first two species types, as described
under the heading Validation). The data for these species are representative of the
data for all 29 species evaluated in the larger study (Elith et al. 1998). The
majority of models do not discriminate well at a fine (subcatchment) scale. The
predictions of any one method were not consistently better than the others, and
there was no significant difference (p < .05) between the mean ROC areas for any
of the methods across the seven species. In the larger data set, GAMs were more
frequently represented among the models with ROC area greater than 0.75 than
any other method (Elith et al. 1998). Species that occurred frequently in the
original 3,522 quadrats were no better modeled than species that were relatively
scarce.
Modeling success varied between species. Three species for which sufficiently
discriminatory models could not be found were Tetratheca stenocarpa, Wittsteinia
vacciniacea, and Phebalium bilobum (Fig. 4.2). These models were based on 54,
131, and 119 records, respectively. The species with the least successful models,
Wittsteinia vacciniacea, is confined to high rainfall montane and subalpine areas.
Although it is ubiquitous in some patches in the region, there are other areas where
suitable habitat is restricted, and the available variables did not adequately
describe the important variation in the environment. The “successful” models
were for Grevillea barklyana (all methods), Helichrysum scorpioides and Lep-