Modeling Species Distribution with GIS
407
Obviously the strong wind of the example should be regarded as a stochastic
event and thus be treated as an outlier in the definition of a possible GIS distribution model. In other words, observations should be analyzed for their content of unconstrained selection by the species.
We will see, when dealing with the issues of scale, that GIS distribution
models tend to describe only the deterministic components that drive a
species’ distribution pattern, so stochastic events must be either averaged on
the long term or eliminated as outliers. When observations are carried out for
a limited time and the biology of the species under investigation is scarcely
known, this problem can become increasingly important because the identification of outliers will be virtually impossible.
Statistical assumptions
Most of the statistical techniques used to define species–environment relationships rely on the identification of two observation sets: one that identifies locations in which the species is present and one in which it is absent. Even though
this cannot be identified properly as a statistical assumption, it is probably the
most important factor limiting the applicability of the statistical techniques
that rely on the two groups of observations.
The most common way to define the two subsets is to compare locations
of known presence with a random sample of locations not pertaining to the
previous set. Obviously some of the random locations can represent a suitable
environment for the species, thus introducing, for that particular environment, a bias that underestimates the species–environment association.
To overcome this problem, data sets can be screened for outliers (Jongman
et al. 1995), using for instance a scatter plot of the variables taken two by two.
Once an outlier is identified, it can be checked to identify possible reasons for
the absence of the species and, if necessary, removed from the analysis. Similar
results can be achieved through analyses such as decision trees, where additional rules can be introduced to predict outliers (Walker 1990; Skidmore et
al. 1996).
Another way to get around the problem is to eliminate the absence subgroup. Skidmore et al. (1996), for example, used both the BIOCLIM approach
and the supervised nonparametric classifier, which use only observation sites
to derive distribution patterns. The same result can also be achieved by using
distance (or similarity) measures from the environmental characteristics of
locations in which the species has been observed. A measure of distance that
seems particularly promising for this application is the Mahalanobis distance
407
Obviously the strong wind of the example should be regarded as a stochastic
event and thus be treated as an outlier in the definition of a possible GIS distribution model. In other words, observations should be analyzed for their content of unconstrained selection by the species.
We will see, when dealing with the issues of scale, that GIS distribution
models tend to describe only the deterministic components that drive a
species’ distribution pattern, so stochastic events must be either averaged on
the long term or eliminated as outliers. When observations are carried out for
a limited time and the biology of the species under investigation is scarcely
known, this problem can become increasingly important because the identification of outliers will be virtually impossible.
Statistical assumptions
Most of the statistical techniques used to define species–environment relationships rely on the identification of two observation sets: one that identifies locations in which the species is present and one in which it is absent. Even though
this cannot be identified properly as a statistical assumption, it is probably the
most important factor limiting the applicability of the statistical techniques
that rely on the two groups of observations.
The most common way to define the two subsets is to compare locations
of known presence with a random sample of locations not pertaining to the
previous set. Obviously some of the random locations can represent a suitable
environment for the species, thus introducing, for that particular environment, a bias that underestimates the species–environment association.
To overcome this problem, data sets can be screened for outliers (Jongman
et al. 1995), using for instance a scatter plot of the variables taken two by two.
Once an outlier is identified, it can be checked to identify possible reasons for
the absence of the species and, if necessary, removed from the analysis. Similar
results can be achieved through analyses such as decision trees, where additional rules can be introduced to predict outliers (Walker 1990; Skidmore et
al. 1996).
Another way to get around the problem is to eliminate the absence subgroup. Skidmore et al. (1996), for example, used both the BIOCLIM approach
and the supervised nonparametric classifier, which use only observation sites
to derive distribution patterns. The same result can also be achieved by using
distance (or similarity) measures from the environmental characteristics of
locations in which the species has been observed. A measure of distance that
seems particularly promising for this application is the Mahalanobis distance