Steven Phillips, Rob Anderson, & Rob
From the Special Topic of
In the Research Front Map
Change and Species Distributions," which is part of our
Special Topics analysis of Climate Change research over the
past decade, the paper "Maximum entropy modeling of species
geographic distributions," (Phillips SJ, Anderson RP,
Schapire RE, Ecol. Model. 190[3-4]: 231-59, 25
January 2006) has 192 cites. This paper is also a Highly
Cited Paper in the field of Environment & Ecology
Essential Science IndicatorsSM from
Authors Dr. Steven Phillips, Dr. Robert P. Anderson, and Dr. Robert
Schapire are also in the top 1% among scientists in this field in the
database. Dr. Phillips is a Lead Member of Technical Staff, Research, at
AT&T Labs in Florham Park, NJ; Dr. Anderson is an Associate Professor
in the Department of Biology at the City College of the City University of
New York in New York City; and Dr. Schapire is Professor of Computer
Science at Princeton University in Princeton, NJ.
Below, they talk with ScienceWatch.com
about their paper, its impact on conservation
research, and the future of ecological
Would you please describe the significance of
your paper and why it is highly cited?
Our paper introduced a mathematically rigorous method (Maxent) for modeling
species geographic distributions, based on known occurrences and
environmental (especially climatic) predictor variables. Applying such
models to predicted future climatic conditions enables prediction of
climate-change impact on individual species, so ecologists can better
understand the likely scope of climate change as a threat to biodiversity.
Maxent relies on data for species presences (such as from natural history
museums and herbaria) rather than presence/absence surveys, which allows
for modeling of many more species. Although other techniques are available
for this, ours has a clear mathematical formulation with explicit
assumptions and links to ecological theory, and is available in
user-friendly, free software.
Steven Phillips in a baobab in
A Maxent model of the distribution of Bradypus
How did you become involved in this research, and
were there any particular successes or obstacles that stand out?
Steven Phillips was seeking ways for a computer scientist to contribute to
conservation. He learned of the uses of species distribution models while
volunteering at the Center for Biodiversity and Conservation at the
American Museum of Natural History, where Rob Anderson was modeling mammal
distributions. Rob Schapire, a colleague of Steven's at AT&T Research
at the time, is a machine-learning specialist. Together, we decided to see
how machine learning could help to model species distributions.
Two major successes:
1. Early on, our research team was joined by Miroslav Dudík, a Ph.D.
student working with Rob Schapire. Miro made major contributions to the
theory and implementation of Maxent that followed the publication of this
2. Our Maxent software participated in a species-distribution modeling
"bakeoff"—a comparison of the performance of a wide variety of
modeling methods (Elith J, et al.,Ecography, 2006).
Maxent and boosted regression trees emerged as the two highly performing
methods in that study.
1. Machine learning and ecology use different and specialized vocabulary,
so the collaboration required each of us to learn a new language, in some
cases even new meanings for the same word from our original discipline.
2. Biological data are particularly challenging: by machine-learning
standards, the amount of occurrence data available for modeling many
species is miniscule, especially for rare species which are of special
concern. In addition, there are multiple biases in the data, and absence
data are missing for most species, making both modeling and model
evaluation difficult. These challenges have all yielded fruitful topics for
new research in machine learning.
Where do you see your research and the broader
field leading in the future?
Recent publications have shown that predictions from different
species-distribution models can vary widely, especially when predicting
species distributions under climate change. Such variations are driven by
climate-model uncertainty; by data issues such as bias, small sample sizes
and missing predictors; by mathematical differences among modeling methods;
and by the fact that models consider only a portion of the inherent
complexity of ecological systems.
In addition to better addressing each of these sources of error, we need
better ways to evaluate predictions of future distributions, and ways to
estimate and communicate the degree of confidence in predictions. Finally,
since many important applications (including climate change and invasive
species) require models of species distributions that are transferable
across space and/or time, additional tests of model performance in this
arena (rather than of species' current distributions) are necessary.
Both for ourselves and our respective fields, we look forward to further
fruitful collaborations and applications of machine learning, computer
science, and statistics to problems in ecology.
What are the implications of your work for this
The technique has and will continue to be used for myriad applications
related to climate change and species distributions. Examples include the
study of niche evolution and speciation, discovery of new populations of
rare species, large-scale conservation planning and site prioritization,
prediction of the spread of invasive species, and anticipation of outbreaks
of zoonotic and other kinds of diseases.
Dr. Steven Phillips
Florham Park, NJ, USA
Dr. Robert P. Anderson
Department of Biology
City College of the City University of New York
New York, NY, USA
Dr. Robert Schapire
Department of Computer Science
Princeton, NJ, USA