tribution surface per week for 52 weeks of the year.
The surface is the probability of occurrence on the
given day estimated at 130 thousand locations from
a geographically stratified random design. All effort
predictors were held constant to remove variation in
detectability. The precise quantity estimated is the
relative probability that a typical eBird participant
will detect the species on a search at a given location
from 7 to 8 AM while traveling 1 kilometer on the
given day of the year.
Figure 3 shows the distribution estimates for Barn
Swallow (top), Blackpoll Warbler (middle), and
Black-throated Blue Warbler (bottom) on June 28
(left), October 11 (center), and December 20 (right).
For the three species these dates fall in the breeding
season, during autumn migration, and in the nonbreeding season, respectively. To control for seasonal variation in detectability when comparing distributions across dates and species, we standardized
the predicted probability of occurrence across distributions.
Across their annual cycle, Barn Swallow occur
throughout most of the longitudinal extent of the
terrestrial Western Hemisphere with broad and complex movements during autumn migration (figure 3,
top and center), which occur over several months.
This contrasts with the Blackpoll Warbler, a neotropical migrant that breeds in large numbers in the
boreal forests of North America and undertakes one
of the longest migrations of any North American
warbler (DeLuca et al. 2013). During autumn migration, it travels first eastward to the North American
coast from which the majority of individuals make a
transatlantic flight through the eastern Caribbean to
Northern South America. Figure 3 (center) shows
how the Blackpoll population occurrence is concentrated along the northeastern coast of the United
States in October, with much lower rates of occurrence along the southeast coast and a second high
concentration in the Western Caribbean islands.
These distributional patterns are in agreement with
studies of Blackpoll Warbler migration that have
relied on a variety of different data sources, based
primarily on observations of individual birds (
DeLuca et al. 2013). Finally, the Black-throated Blue Warbler, one of the most extensively studied passerine
species in North America, migrates south from the
eastern deciduous forest of North America along a
broad front from the eastern seaboard to the
Appalachians where it reaches its wintering grounds
in the Caribbean (Holmes, Rodenhouse, and Sillett
Assessing the Scale
and Quality of AdaSTEM
High-quality species distribution and movement
information is useful for a variety of ecological and
conservation applications across a range of scales.
However, distribution estimates by themselves are
not sufficient for most applications because they do
not convey information about the scale or quality of
the estimates. More often than not, distribution estimates from spatially explicit models are computed at
arbitrarily fine resolutions. Visualizations and maps
generated from these products risk communicating
the existence of fine-scale patterns where none may
be supported by the data. Without understanding the
spatial resolution of an estimate it is easy to overin-terpret the results and make inference about fine-scale patterns where this is not warranted.
In this section we present a set of model diagnostics to assess and visualize spatial patterns of scale,
bias, and uncertainty. First, we determine the spatial
scale of the distribution estimates so that spurious
inferences about fine-scale patterns can be avoided.
Understanding the spatial scale of estimates is also
necessary for constructing statistical comparisons
between regions. Second, we present an analysis of
Understanding spatial patterns of bias is especially useful when using crowdsourced data. Finally, we
provide a quantitative assessment of the uncertainty attached to distribution estimates. This information is essential for making decisions in the face of
Together these three diagnostics provide useful
information to interpret and apply summaries of distribution estimates to real-world sustainability problems. All of the diagnostics discussed here are for the
June 28 Barn Swallow AdaSTEM distribution estimate
(figure 3, top, left).
Assessing Spatial Scale
We formalize the notion of scale as the effective
range, the shortest distance at which the correlation
between pairs of measurements within a neighborhood becomes negligibly small (Banerjee, Carlin, and
Gelfand 2004). To assess scale of a distribution estimate we measure the effective range of the residuals.
The spatial variation in scale is visualized by computing effective ranges across a half degree grid and
Figure 4 (left) shows the interpolated effective
range for the distribution of Barn Swallow based on
residuals from June 24–July 1. The portion of the
study area with insufficient residual density to estimate the effective range is shown in grey and the
effective range of this Barn Swallow distribution estimate can be seen to vary from less than 10 kilometers
to over 100 kilometers depending on location. For
example, in regions with very short ranges, like Ithaca, New York, in the northeastern United States, estimates of occurrence separated by as little as 5 kilometers are independent. At the same time, in the
state of Montana, located in the northcentral United
States, occurrence estimates must be separated by at
least 60 kilometers to be independent of each other.