42
Jane Elith
section of the program produces ranked predictions for each site of interest, and
the ranks are based on user-specified bounds of the species profile. These predictions define the climatic suitability of the site for the species. The fact that a
bounding box is used to describe the potential distribution rather than a more
restricted polygon means that large areas are rated as suitable even though the
species is likely to be absent, because on an ecological basis one would not expect
species to occupy extreme combinations of the ranges (Nix 1986; Carpenter et al.
1993). The program only deals with climate variables, although the approach
could theoretically be extended to a broader range of environmental variables.
This may be especially important if the method is to be applied to animal species.
DOMAIN is one example of the application of multivariate distance measures
to mapping. Conceptually, it takes the opposite approach to ANUCLIM, by defining sites of similarity rather than by determining bounds. It provides a measure of
similarity for each site of interest, in which similarity is the environmental similarity between the site of interest and the most similar known record site, and is
calculated from the Gower metric (Legendre and Legendre 1998). Environmental
differences between sites are scaled by the range of each environmental variable
at the sites at which the species has been recorded. DOMAIN can be used to
specify an environmental envelope by selecting a lower threshold of similarity or
to map similarities on a continuous scale. Only presence data are required as
species records.
GLMs are a broad class of statistical models that include ordinary regression
and analysis of variance. All GLMs have a random component (the response
variable), a systematic component (the predictor or explanatory variables), and a
link function that describes the relationship between the expected value of the
response and the predictors. The predictors and their coefficients are always
combined in a linear form, even if individual explanatory variables are presented
in nonlinear form (see, e.g., Austin et al. 1994a). The particular forms of GLMs
that are useful for modeling species distributions are logistic regression for
presence-absence data and Poisson regression for abundance data. These methods
estimate the probability of presence of the species. The predicted probability of
presence may be used as an index of habitat quality.
GAMs are a nonparametric extension of GLMs, in which the linear or polynomial functions in a GLM are replaced by smoothed data-dependent functions in
a GAM. They are considered a useful tool in modeling biological systems because
the response is not limited to a parametric function, which means that the fitted
response surface may be a more realistic representation of the true response shape.
GAMs retain many of the features of GLMs, including additivity of the predictor
effects—this means, in practice, that the roles of the different predictor variables
can still be assessed in the model. GAMs can include both parametric and nonparametric terms. The predicted probability surface may be mapped and interpreted as an estimate of habitat suitability, as it may be for GLMs.
Genetic algorithms are general purpose optimization techniques based on a set
of logical learning rules. The “genetic” analogy suggests evolution in which
Jane Elith
section of the program produces ranked predictions for each site of interest, and
the ranks are based on user-specified bounds of the species profile. These predictions define the climatic suitability of the site for the species. The fact that a
bounding box is used to describe the potential distribution rather than a more
restricted polygon means that large areas are rated as suitable even though the
species is likely to be absent, because on an ecological basis one would not expect
species to occupy extreme combinations of the ranges (Nix 1986; Carpenter et al.
1993). The program only deals with climate variables, although the approach
could theoretically be extended to a broader range of environmental variables.
This may be especially important if the method is to be applied to animal species.
DOMAIN is one example of the application of multivariate distance measures
to mapping. Conceptually, it takes the opposite approach to ANUCLIM, by defining sites of similarity rather than by determining bounds. It provides a measure of
similarity for each site of interest, in which similarity is the environmental similarity between the site of interest and the most similar known record site, and is
calculated from the Gower metric (Legendre and Legendre 1998). Environmental
differences between sites are scaled by the range of each environmental variable
at the sites at which the species has been recorded. DOMAIN can be used to
specify an environmental envelope by selecting a lower threshold of similarity or
to map similarities on a continuous scale. Only presence data are required as
species records.
GLMs are a broad class of statistical models that include ordinary regression
and analysis of variance. All GLMs have a random component (the response
variable), a systematic component (the predictor or explanatory variables), and a
link function that describes the relationship between the expected value of the
response and the predictors. The predictors and their coefficients are always
combined in a linear form, even if individual explanatory variables are presented
in nonlinear form (see, e.g., Austin et al. 1994a). The particular forms of GLMs
that are useful for modeling species distributions are logistic regression for
presence-absence data and Poisson regression for abundance data. These methods
estimate the probability of presence of the species. The predicted probability of
presence may be used as an index of habitat quality.
GAMs are a nonparametric extension of GLMs, in which the linear or polynomial functions in a GLM are replaced by smoothed data-dependent functions in
a GAM. They are considered a useful tool in modeling biological systems because
the response is not limited to a parametric function, which means that the fitted
response surface may be a more realistic representation of the true response shape.
GAMs retain many of the features of GLMs, including additivity of the predictor
effects—this means, in practice, that the roles of the different predictor variables
can still be assessed in the model. GAMs can include both parametric and nonparametric terms. The predicted probability surface may be mapped and interpreted as an estimate of habitat suitability, as it may be for GLMs.
Genetic algorithms are general purpose optimization techniques based on a set
of logical learning rules. The “genetic” analogy suggests evolution in which