28 AI MAGAZINE
We thank the thousands of eBird participants, K.
Rosenberg, S. Sukhnanand, and the Lab of
Ornithology IS eBird team — K. Webb, M. Iliff, J.
Gerbracht, T. Lenz, W. Morris, B. Sullivan, and C.
Wood. This work was supported by the Leon Levy
Foundation, the Wolf Creek Foundation, and the
National Science Foundation (OCI-0830944, CCF-
0832782, ITR-0427914, DBI-1049363, DBI-0542868,
DUE-0734857, IIS-0748626, IIS-0844546, IIS-
0612031, IIS-1050422, IIS-0905385, IIS-0746500,
IIS-1209589, AGS-0835821, CNS-0751152, CNS-
0855167, IIS-1017793, CDI-1125098) with comput-
ing support from CNS-1059284, OCI-1053575 and
Banerjee, S.; Carlin, B.; and Gelfand, A. 2004. Hierarchical
Modeling and Analysis for Spatial Data, vol. 101. London:
Chapman & Hall/CRC.
Bonter, D. N.; Donovan, T. M.; Brooks, E. W.; and Hobson,
K. A. 2009. Daily Mass Changes in Landbirds During Migration Stopover on the South Shore of Lake Ontario. The Auk
124( 1): 122–133. dx.doi.org/10.1642/0004-8038(2007)124
[122:DMCILD] 2.0.CO; 2
Breiman, L. 1996. Bagging Predictors. Machine Learning
24( 2): 123–140. dx.doi.org/10.1007/BF00058655
Breiman, L.; Friedman, J.; Stone, C.; and Olshen, R. 1984.
Classification and Regression Trees. London: Chapman &
Crainiceanu, C.; Ruppert, D.; Carroll, R.; Joshi, A.; and
Goodner, B. 2007. Spatially Adaptive Bayesian Penalized
Splines with Heteroscedastic Errors. Journal of Computational and Graphical Statistics 16( 2): 265–288. dx.doi.org/
Cressie, N. 1993. Statistics for Spatial Data. New York: John
Wiley, 2nd ed.
Cressie, N. 1986. Kriging Nonstationary Data. Journal of the
American Statistical Association 81(395): 625–634.
Cressie, N., and Johannesson, G. 2008. Fixed Rank Kriging
for Very Large Spatial Data Sets. Journal of the Royal Statistical. Society B 70( 1): 209–226. dx.doi.org/10.1111/j.1467-
Cressie, N., and Wikle, C. K. 2012. Statistics for Spatio-Tem-poral Data. New York: John Wiley.
DeLuca, W.; Holberton, R.; Hunt, P. D.; and Eliason, B. C.
2013. Blackpoll Warbler (Setophaga striata). In The Birds of
North America Online, ed. A. Poole. Ithaca, NY: Cornell
Laboratory of Ornithology.
Dickinson, J. L.; Zuckerberg, B.; and Bonter, D. N. 2010. Cit-
izen Science as an Ecological Research Tool: Challenges and
Benefits. Annual Review of Ecology, Evolution, and Systematics
from the binary classification problem itself — vari-
ability is greatest where predicted probabilities are
near 0.5 and smallest when the predicted probabili-
ties are closest to 0 or 1. Another reason for relative-
ly large standard errors in data dense regions is the
fact that AdaSTEM’s adaptive partitioning tends to
minimize bias in data dense regions, potentially lead-
ing to higher variance.
Using AdaSTEM, we have produced the first hemi-
spherewide population-level distribution estimates of
long-distance migrations using crowdsourced data
from eBird. These estimates demonstrate how
AdaSTEM can automatically adapt to patterns across
several orders of magnitude. While the hemispheric
extent of analysis extends over 10,000 kilometers
north to south, we found that the spatial resolution
of the distribution estimates was less than 100 kilo-
meters within most of the continental United States
and Southern Canada. In several data rich regions of
North America, the spatial resolution was found to
be less than 10 kilometers.
The simple adaptive divide and recombine strategy employed by AdaSTEM provides sufficient flexibility to model complex spatiotemporal processes
across a range of scales. AdaSTEM as a class of models will be useful in other spatial and spatiotemporal
domains where data are irregularly and sparsely distributed, such as applications based on geographic
surveys and geolocated data collected by volunteers
through crowdsourcing platforms.
The three species whose data were analyzed in this
article span the rich variation in avian distributional
dynamics that characterize bird species’ annual
cycles. For long-distance migrants, these dynamics
extend well beyond the conterminous USA, where
research and conservation efforts are often focused,
including our own efforts based on eBird data (La
Sorte et al. , for example). By modeling occurrences across the entire Western Hemisphere,
AdaSTEM provides novel information on how these
dynamics are structured for entire populations of
multiple species across their entire annual cycles,
even for species that are panhemispheric migrants.
This information has tremendous potential to generate novel inferences in avian ecology and evolution,
and to benefit national and international efforts in
avian conservation. For example, we can now gain a
more detailed understanding of theprocess of migration — how fast birds travel, which routes they take,
whether the same routes are followed northward and
southward, and whether there are discrete collections
of species that travel along the same flyways — that
have previously only been studied over smaller
region (La Sorte et al. 2013) or for very small numbers
of individual birds (Stutchbury et al. 2009). By
expanding and improving our existing knowledge