Short communicationSensitivity of distributional prediction algorithms to geographic data completeness
Introduction
An emerging field in landscape ecology and conservation biology is that of building ecological niche models to predict geographic distributions of species (e.g. Breininger et al., 1991, Dewey et al., 1991, Pereira and Itami, 1991, Stoms et al., 1992, Herr and Queen, 1993, Mladenoff et al., 1995, Sperduto and Congalton, 1996, Bian and West, 1997, Li et al., 1997). Especially challenging has been developing such models from point data, especially data such as museum specimen locality data, which are rarely accompanied by data indicating absence or abundance at sites. Several analytical approaches have been applied to these challenges, including logistic regression, discriminant function analysis, approaches based on distance measures, and parallelipiped set-based approaches.
Although broad comparisons of efficiency have not yet been carried out, our preliminary trials (L.G. Ball et al., in preparation) indicate that an especially promising algorithm is an expert-system approach based on a genetic algorithm, developed by David Stockwell (Stockwell and Noble, 1991, Stockwell, 1993). This algorithm, called Genetic Algorithm for Rule-set Prediction (GARP) produces a model of species, niches in geographic space based on heterogeneous rule-sets (see Methods). GARP appears especially well-suited to reducing error in predicted distributions, both in the form of omission of real distributional areas and inclusion of areas not holding actual populations.
One dimension of the challenge of distributional prediction that has not seen careful testing is that of the depth or completeness of geographic information necessary for accurate predictions. The purpose of the present contribution is to analyze—for a particular geographic and taxonomic challenge—the completeness of geographic data necessary for accurate distributional prediction, and if possible, what particular thematic coverages are especially important.
Section snippets
Methods
The three species employed in this study were chosen to represent distinct distributional areas in diverse ecosystems across North America. The Pygmy Nuthatch (Sitta pygmaea) is distributed in coniferous forests throughout western North America south to central Mexico. The Barred Owl (Strix varia) is distributed in deciduous and coniferous forests almost exclusively in eastern North America, and locally south to southern Mexico. Finally, the Lark Bunting (Calamospiza melanocorys) is distributed
Results
Predictive accuracy was clearly related to the number of coverages, with low accuracy associated with few coverages, and highest accuracy generally achieved with the full or near-full complement of coverages. Fig. 1 illustrates the highly accurate prediction based on all eight coverages, compared with an example prediction based on a single coverage, for Pygmy Nuthatches. In each case, the decline in correspondence of the predicted distributional area with known distributional points in the
Methodology
This preliminary study assessed the amount of geographic information necessary for accurate distributional predictions. The conclusion was simple in the case of North America; five coverages, one of which being that of annual mean temperature, were necessary for consistently accurate prediction. Inclusion of additional coverages and themes did not greatly affect the resulting accuracy. This result, of course, depends to some degree on the particular species, and on the array of geographic
Acknowledgements
We thank Dr David Stockwell for his generous and capable assistance in all phases of this study. Thanks also to Dr Bruce Peterjohn, of the Breeding Bird Survey, for providing critical avian distributional data. Finally, this study was supported financially by the National Geographic Society, the National Science Foundation, the Kansas EPSCor, and the University of Kansas Honors Program.
References (15)
- et al.
A regression model for the spatial distribution of red-crown crane in Yancheng Biosphere Reserve, China
Ecol. Model.
(1997) LBS: Bayesian Learning System for rapid expert system development
Expert Syst. Appl.
(1993)- et al.
GIS modeling of elk calving habitat in a prairie environment with statistics
Photogr. Eng. Remote Sens.
(1997) - et al.
Mapping Florida Scrub Jay habitat for purposes of land-use management
Photogr. Eng. Remote Sens.
(1991) - et al.
Hierarchical partitioning
Am. Stat.
(1991) - et al.
Satellite remote sensing to predict potential distribution of dyers woad (Isatis tinctoria)
Weed Tech.
(1991) - et al.
Crane habitat evaluation using GIS and remote sensing
Photogr. Eng. Remote Sens.
(1993)