Modeling Species Distribution with GIS - Research Techniques in Animal Ecology - page 421

A number of possible solutions are suggested for modeling accumulation of
error in a GIS . These include reliability diagrams and a probabilistic approach.
Wright (1942) suggested that reliability diagrams should accompany all
maps. He also emphasized that the sources used to generate different regions
of the map have varying accuracy and these sources should be stated clearly on
the map. For example, one region may have been mapped using low-altitude
aerial photography and controlled ground survey, and would therefore be
more accurate than another region mapped using high-altitude photography
and only reconnaissance survey. This theme was taken up by Chrisman (1987)
and MacEachren (1985), who suggested that such a reliability diagram show-
ing map pedigree should be included as an additional layer accompanying
each map layer in a GIS . However, for the purposes of error accumulation mod-
eling, reliability diagrams do not provide a quantitative statement about the
accuracy (or error) of the map.
The supervised nonparametric classifier described by Skidmore and Turner
(1988) and Skidmore et al. (1996) classified remotely sensed and GIS digital
data. The classifier gives for all cells the empirical probability of correct classi-
fication for each class according to the training area data and thereby gives an
indication of map accuracy.
The few methods proposed for modeling error accumulation are limited in
their application. Working with ideal data, these methods do allow some con-
clusions to be drawn about error accumulation during GIS overlay operations.
However, the methods break down when used with map layers created under
different conditions than assumed by the methods.
Newcomer and Sjazgin (1984) used probability theory to model error
accumulation, whereas Heuvelink and Burrough (1993) modeled the accu-
mulation of error in Boolean models, using surfaces interpolated by kriging as
the estimated error source.
Sensitivity analysis
When no information is available on the extent of the errors of the original
data sets or on the type of error propagation function applicable to the model,
a way of defining levels of reliability of the output is to analyze its variability
subject to changes in the input parameters. In the this chapter we have seen
possible sources of variability and uncertainties that can arise in the deductive
and inductive approaches in species distribution modeling. They range from
subjective errors introduced by the specialist who defines the species-environ-
ment relationship to locational errors of species observation caused by possible