46 AI MAGAZINE
Our approach to reconstructing the velocity fields of
migrating birds detected by WSR-88D radars is a sig-
nificant improvement over previous methods. Our
algorithm will allow us to overcome a fundamental
challenge of analyzing radar data to tap the potential
information about bird migration available from the
continent-scale WSR-88D network. By creating an AI
tool to automate velocity data processing, we can
extract information about bird migration more accu-
rately and at a substantially larger scale than previ-
ously possible, and make advances in our knowledge
of bird movements for science and conservation. Of
particular value is the potential for using this newly
possible understanding of bird migration from our
radar studies as a platform to monitor and assess
ecosystem health. Many species of birds are excellent
bioindicators (Holt and Miller 2011), and quantify-
ing bird migration at regional and continental scales
across 10–20 year intervals provides an invaluable
opportunity to examine patterns of movements rela-
tive to changes in landscape and climate.
Collaborations between computer scientists and
domain scientists in sustainability-related fields can
yield powerful insights and advances in domains of
mutual interest. In our case, uniting radar ornithology with machine-learning research has yielded previously inaccessible and unavailable information about
the behaviors of birds migrating over the northeastern United States. This powerful example can be a
model for scaling up analyses to much larger spatial
and temporal scales. Many different future analyses
will benefit from having a robust quantification of
bird densities and velocities aloft: for example, comparisons among radar stations in different parts of
the United States, comparisons with weather and climate data, and comparisons with species observations from eBird and other sources. Future work will
expand on our existing dataset of 35,000 scans from
12 radar stations operating from 1 August to 30
November in 2010 and 2011 in the northeastern
United States to include the entire continent. Our
approach also highlights additional algorithms necessary to continue the process of exploring this massive dataset. For example, a remaining challenge is to
automate the screening of nonbiological targets and
to perform these analyses in real time. A detailed
understanding of the velocity field is a key first step
toward this goal (Dokter et al. 2011); a complete solution will likely involve other applications of
machine-learning algorithms as well.
This article has described a case study in the AI
techniques underlying automated analysis of bird
migration at large scales using radar data. A fully
automated continent-scale analysis of bird migration
may be game changing in terms of understanding
this global ecological process and protecting the sys-
tem from growing anthropogenic threats. Recent
advances in the new field of computational sustain-
ability have the potential to make this a reality. We
believe it is now possible to study migration at con-
tinent scale by integrating radar data together with
complementary sources of information such as citi-
zen science data from eBird and acoustic monitoring.
Each of these data sources is available at the conti-
nent scale but provides only partial information
about bird migration. The synthesis of these diverse
data sources using novel analysis approaches will
allow us to reconstruct, understand, and forecast
nightly bird migrations at the continental scale.
This work was supported in part by the National Sci-
ence Foundation under Grant No. 1125228 and by
Leon Levy Foundation.
1. We will assume the velocity in the vertical direction is
negligible, though the models can easily be extended to
include this as well.
2. The phase and amplitude are determined by the
unknown velocity parameters u and v.
3. The aliasing operation of a value is mathematically equivalent to the modulus operation in modular arithmetic,
except it follows the convention that the result lies in the
interval [–Vmax, Vmax] instead of [0,2Vmax].
4. When the variance is larger, the appearance of the
wrapped normal density is less like its normal counterpart.
As variance approaches infinity, the distribution approaches a uniform distribution on (–Vmax, Vmax).
Able, K. P. 1999. How Birds Migrate: Flight Behavior, Ener-
getics, and Navigation. In Gatherings of Angels: Migrating
Birds and Their Ecology, ed. K. P. Able, 11–27. Ithaca: Cornell
Bonter, D. N.; Donovan, T. M.; and Brooks, E. W. 2009. Daily Mass Changes in Landbirds During Migration Stopover
on the South Shore of Lake Ontario. Auk 124( 1): 122–133.
Breitenberger, E. 1963. Analogues of the Normal Distribution on the Circle and the Sphere. Biometrika 50( 1): 81–88.
Buler, J. J., and Diehl, R. H. 2009. Quantifying Bird Density
During Migratory Stopover Using Weather Surveillance
Radar. IEEE Transactions on Geoscience and Remote Sensing
47( 8): 2741–2751. doi: 10.1109/TGRS.2009.2014463.
Buler, J. J.; Randall, L. A.; Fleskes, J. P.; Barrow W. C.; Bogart, T.; and Kluver, D. 2012. Mapping Wintering Waterfowl
Distributions Using Weather Surveillance Radar. PloS One
7( 7): e41571. dx.doi.org/10.1109/TGRS.2009.2014463
Carey, J. 2012. Drones at Home. Scientific American 307( 6):
Crawford, R., and Engstrom, T. 2001. Characteristics of
Avian Mortality at a North Florida Television Tower: A 29-
Year Study. Journal of Field Ornithology 72( 3): 380–388.
Crum, T. D., and Alberty, R. L. 1993. The WSR-88D and the
WSR-88D Operational Support Facility. Bulletin of the Amer-