ArticlePDF Available

A Geostatistical Approach to Estimate High Resolution Nocturnal Bird Migration Densities from a Weather Radar Network

Abstract and Figures

Quantifying nocturnal bird migration at high resolution is essential for (1) understanding the phenology of migration and its drivers, (2) identifying critical spatio-temporal protection zones for migratory birds, and (3) assessing the risk of collision with artificial structures. We propose a tailored geostatistical model to interpolate migration intensity monitored by a network of weather radars. The model is applied to data collected in autumn 2016 from 69 European weather radars. To validate the model, we performed a cross-validation and also compared our interpolation results with independent measurements of two bird radars. Our model estimated bird densities at high resolution (0.2 • latitude-longitude, 15 min) and assessed the associated uncertainty. Within the area covered by the radar network, we estimated that around 120 million birds were simultaneously in flight (10-90 quantiles: 107-134). Local estimations can be easily visualized and retrieved from a dedicated interactive website. This proof-of-concept study demonstrates that a network of weather radar is able to quantify bird migration at high resolution and accuracy. The model presented has the ability to monitor population of migratory birds at scales ranging from regional to continental in space and daily to yearly in time. Near-real-time estimation should soon be possible with an update of the infrastructure and processing software.
Content may be subject to copyright.
remote sensing
A Geostatistical Approach to Estimate High
Resolution Nocturnal Bird Migration Densities
from a Weather Radar Network
Raphaël Nussbaumer 1,2,* , Lionel Benoit 2, Grégoire Mariethoz 2, Felix Liechti 1,
Silke Bauer 1and Baptiste Schmid 1
1Swiss Ornithological Institute, 6204 Sempach , Switzerland; (F.L.); (S.B.); (B.S.)
2Institute of Earth Surface Dynamics, University of Lausanne, 1015 Lausanne, Switzerland; (L.B.); (G.M.)
Received: 25 July 2019; Accepted: 19 September 2019; Published: 25 September 2019
Quantifying nocturnal bird migration at high resolution is essential for (1) understanding
the phenology of migration and its drivers, (2) identifying critical spatio-temporal protection zones
for migratory birds, and (3) assessing the risk of collision with artificial structures. We propose a
tailored geostatistical model to interpolate migration intensity monitored by a network of weather
radars. The model is applied to data collected in autumn 2016 from 69 European weather radars.
To validate the model, we performed a cross-validation and also compared our interpolation results
with independent measurements of two bird radars. Our model estimated bird densities at high
resolution (0.2
latitude–longitude, 15 min) and assessed the associated uncertainty. Within the area
covered by the radar network, we estimated that around 120 million birds were simultaneously in
flight (10–90 quantiles: 107–134). Local estimations can be easily visualized and retrieved from a
dedicated interactive website. This proof-of-concept study demonstrates that a network of weather
radar is able to quantify bird migration at high resolution and accuracy. The model presented has
the ability to monitor population of migratory birds at scales ranging from regional to continental in
space and daily to yearly in time. Near-real-time estimation should soon be possible with an update
of the infrastructure and processing software.
aeroecology; bird migration; geostatistical modeling; interactive visualization; kriging;
radar network; spatio-temporal interpolation map; weather radar
1. Introduction
Every year, several billions of birds undergo migratory journeys between their breeding and
non-breeding grounds [
]. These migratory movements link ecosystems and biodiversity on a global
scale [
], and their understanding and protection require international efforts [
]. Indeed, declines in
many migratory bird populations [
] resulted from the rapid changes in their habitats, including
the aerosphere [
]. Changes in aerial habitats are diverse, and their consequences still poorly known.
Climate change may alter global wind patterns and consequently the wind assistance provided to
migrants [
]. Likely to be more severe, the impact of direct anthropogenic changes, including light
pollution that reroutes migrants [
], buildings [
], wind energy production [
], and aviation [
causes billions of fatalities every year [13].
In the face of these threats and to set up efficient management actions, we need to quantify bird
migration at various spatial and temporal scales. Fine-scale monitoring is crucial for understanding
the phenology of migration and its drivers, identifying critical spatio-temporal protection zones to
Remote Sens. 2019,11, 2233; doi:10.3390/rs11192233
Remote Sens. 2019,11, 2233 2 of 24
support conservation actions, and assessing collision risks with artificial structures and aviation to
inform stakeholders. However, the great majority of migratory landbirds fly at night [
], rendering
the quantification of the sheer scale of bird migration a challenging exercise.
Radar monitoring has the potential to quantify birds’ migratory movements at the continental
scale [
]. Initially limited to single dedicated short-range radars, radar aeroecology truly took off
when it was able to leverage existing weather radar networks, thus providing continuous monitoring
over large geographical areas such as Europe or North America [
]. One important challenge
in using networks of weather radars is the interpolation of their signals in space and time. Recent
studies [
] have used relatively simple interpolation methods as they targeted patterns at coarse
spatial and temporal scales. However, these methods are insufficient if higher spatial or temporal
resolution is needed, such as for the fundamental and applied challenges outlined above.
To achieve a high-resolution interpolation of migration intensity derived from weather radars
( 20 km–15 min), we propose a tailored geostatistical framework able to model the spatio-temporal
patterns of bird migration. Starting from time series of bird densities measured by a radar network,
our geostatistical model produces a continuous map of bird densities over time and space. A major
strength of this method is its ability to provide the full range of uncertainty and thus to evaluate
complex statistics, for instance, the probability that bird densities reach a given threshold. In addition
to the estimation map, the method also produces simulation maps which are essential for several
applications such as quantification of the total number of birds.
As a proof-of-concept, we applied our geostatistical model to a three-weeks dataset from the
European Network of weather radars [
] and validated the results with independent dedicated bird
radars. In addition to insights into the spatio-temporal scales of broad-front migration, our approach
provides high-resolution (0.2
latitude and longitude, 15 min) interactive maps of the densities of
migratory birds.
2. Materials and Methods
2.1. Weather Radar Dataset
Our dataset originates from measurements of 69 European weather radars, spread from Finland to
the Pyrenees (eight countries) and covering the period from 19 September to 10 October 2016 (Figure 1).
It thus encompasses a large part of the Western European flyway during fall migration 2016.
Based on the reflectivity measurements of these weather radars, we used the bird densities as
calculated and stored on the repository of the European Network for the Radar surveillance of Animal
Movement (ENRAM) ( ([
] for details on the conversion
procedure). We inspected the vertical profiles and manually cleaned the bird densities data (see
detailed procedure in Appendix Aand resulting vertical profiles in Supplementary Material S1).
As we targeted a 2D model, we vertically integrated the cleaned bird densities from the
radar elevation to 5000 m above sea level. Because we aimed at quantifying nocturnal migration,
we restricted our data to night-time, between local dusk and dawn (civil twilight, sun 6
horizon). Furthermore, as rain could contaminate and distort the bird densities calculated from radar
data, a mask for rain was created when the total column of rain water exceeds a threshold of 1 mm/h
(ERA5 dataset from [22]). In the end, the resulting dataset consisted of a time series of nocturnal bird
densities [bird/km2] at each radar site with a resolution of 15 min (Figure 2).
Remote Sens. 2019,11, 2233 3 of 24
Temporal resolution
09/09 - 10/10/2016
143 km
closest radar
25km(64) - 40km(5)
Radar range
Figure 1.
) Locations of weather radars of the European Network for the Radar surveillance of
Animal Movement (ENRAM) network, whose fall 2016-data were used in this study (yellow dots) and
the two dedicated bird radars for validation (red dots). (
) Key characteristics of the dataset used.
50 (a)
Sep 27 Sep 28
Bird densities [bird/km2]
Sep 22 Sep 28 Oct 04 Oct 10
Bird densities [bird/km2]
Figure 2. Illustration of the spatio-temporal variability of bird densities measured by a weather radar
network. (
) Average bird densities measured by each radar over the whole study period. (
) Time
series of bird densities measured by the radar with the corresponding outer ring color in panel (
) Zoom on a two-days period. A strong continental trend appears in panel (
) as well as a correlation
at the multi-night scale when comparing (
). These spatial correlations are even stronger at the
regional scale when comparing within a subplot (
). The intra-night scale shows an obvious bell-shape
curve pattern during each night (g).
2.2. Interpolation Approach
Bird densities are strongly correlated spatially (Figure 2a) and temporally at both nightly
(Figure 2b–d) and sub-nightly scales (Figure 2e–g). These strong spatio-temporal correlations motivated
the use of a Gaussian process regression to interpolate bird densities measured by weather radars at
high temporal resolution. In this framework, the spatio-temporal structure of bird migration is first
Remote Sens. 2019,11, 2233 4 of 24
learned from the punctual radar measurements. This model is then combined with the measurements
to estimate bird densities at any location in space and time. In this paper, we adopt the terminology
and notations of Geostatistics [
], and mention its correspondence with Gaussian processes in the
field in machine learning [25].
Because of the multi-scale temporal structure of bird migration (Figure 2), we consider here an
additive model combining two temporal scales: first, a multi-night process that models bird density
averaged over the night and, second, an intra-night process that models variations within each night.
Subsequently, each scale-specific process is further split into two terms: a smoothly-varying (in space
time) deterministic trend and a stationary Gaussian process.
2.3. Geostatistical Model
The bird density B(s,t)observed at location sand at time tis modeled by
B(s,t)p=µ(s) + M(s,d(t))
| {z }
Multi-night scale
+ι(s,t) + I(s,t)
| {z }
Intra-night scale
, (1)
are deterministic trends at the multi-night and intra-night resolution, respectively,
are random effects at the multi-night and intra-night resolution, respectively;
a step function that maps the continuous time
to the discrete day
of the closest night. A power
is applied on bird densities to transform the highly skewed marginal distribution
into a Gaussian distribution (Figure A2 in Appendix B). Figure 3illustrates the decomposition of
Equation (1) during three nights.
Sep 27, 00:00 Sep 27, 12:00 Sep 28, 00:00 Sep 28, 12:00 Sep 29, 00:00
Bird densities [bird/km2]
Figure 3.
Illustration of the proposed mathematical model decomposition of Equation (1) with the
exception that the power transformation was not applied. Note that the values of
, and
can be
either positive or negative.
2.3.1. Multi-Night Scale
At the multi-night scale, we model bird densities averaged overnight as a space-time Gaussian
process with a spatial trend (also called mean function), due to the general increasing bird densities
southwards (Figure 2a). Because of the relatively short duration of the dataset, no temporal component
is added in the trend. The trend at the multi-night scale is therefore modeled as a planar function,
µ(s=[slat,slon]) =wlatslat +w0, (2)
Remote Sens. 2019,11, 2233 5 of 24
are the latitude and longitude of location
is the slope coefficient in latitude;
is the value of the trend at the origin. Because no longitudinal trend is observed in the
data (Figure 1a), only latitude is used to configure the planar function (see Figure A3 in Appendix B).
If longer periods are considered, Equation (2) can be replaced by a spatio-temporal polynomial function
in order to handle the emerging patterns of long-term nonstationarity.
With the spatial trend accounted for by
can be modeled as a Gaussian random process with
is assumed to be 2
order stationary, so that its covariance function
(also called
autocovariance or kernel function) depends only on s,t,
CM(M(s,t),M(s+s,t+t))=CM(s,t). (3)
The covariance function CMis modeled with the Gneiting type function [26]
. (4)
In Equation (4), the scale parameters
(in space and time, respectively) control the
decorrelation distances and, thus, the average extent and duration of the space-time patterns of
The regularity parameters 0
1 (in space and time, respectively) control the shape of
the covariance function close to the origin. Values of
close to 0 lead to sharp variations at
short lags, whereas values close to 1 lead to smooth variations of
. The separability parameter
controls the space-time interactions. When
0 the space-time interactions vanish and the covariance
function becomes space-time separable.
controls the amplitude of the covariance function. Finally,
is a nugget effect which accounts for the uncorrelated variability of
, and
is the Kronecker
delta function
C0δs=(C0if s=0
0 otherwise . (5)
Note that in contrast to a usual nugget
, here we use a nugget in the space dimension
. This nugget accounts for the uncorrelated variability of
over space which can be caused
by persistent local geographical features affecting bird migration (e.g., topography and water body)
or possible bias of radar observation (e.g., ground scattering). The total variance of
is defined as
CM(0, 0) = C0+CG.
2.3.2. Intra-Night Scale
At the intra-night scale, the main trend visible in the dataset is a bell-shape curve pattern
(Figure 2e–g) that results from the onset and sharp increase of migration activity after sunset, and its
slow decrease towards sunrise [
]. This trend is modeled with a curve template
for all nights and
locations, defined by a polynomial of degree 8,
ι(s,t) =
aiNNT(s,t)i, (6)
are the coefficients of the polynomial and
(Normalized Night Time) is a proxy of the
progression of night, defined such that the local sunrise and sunset occur at
1 and
1 ,
I(s,t) = I(s,t)
σI(s,t), (7)
Remote Sens. 2019,11, 2233 6 of 24
where σI(s,t)is a polynomial function with coefficients bithat models the variation of the variance
σI(s,t) =
biNNT(s,t)i. (8)
This normalization allows to use a stationary covariance function for ˜
. (9)
Note that modeling
through the covariance function of its normalized variable
is equivalent to
modeling Idirectly with a nonstationary covariance function, which includes σI(s,t).
2.4. Bird Migration Mapping
The geostatistical model presented in Section 2.3 is used to interpolate bird density observations
derived from weather radars, and produces high resolution maps of both estimation (with
corresponding uncertainty) and simulation (see Section 2.4.2). In the case study presented in this paper,
the interpolation map is calculated on a spatio-temporal grid with a resolution of 0.2
in latitude (43
to 68
) and longitude (
to 30
) and 15 min in time, resulting in 127
2017 nodes. Over this
large data cube, the estimation and simulation are only computed at the nodes located (1) over land,
(2) within 200km of the nearest radar and (3) during night-time ( 1<NNT(t,s)<1).
2.4.1. Estimation
The estimation is performed by applying universal kriging at both the multi- and intra-night
scales. We employ a two-step approach of universal kriging where the trends
are first estimated
by ordinary least squares (see Appendix B.2 and B.3), and then subtracted from the observations. The
resulting random effects
are parameterized by fitting their covariance function (defined in
Equations (4) and (9)) to the empirical covariance computed based on the detrended observations (see
Appendix B.4). Finally,
are interpolated at any space-time location of interest,
, by simple
kriging (see Appendix C), denoted as
. The final estimation of bird density
Bp(s0,t0)is reconstructed based on Equation (1) as,
I(s0,t0). (10)
An important advantage of using kriging is that it expresses the estimation as a Gaussian
distribution, thus providing not only the “most likely value” (i.e., mean or expected value), but also a
measure of uncertainty with the variance of estimation. As
are constructed independently,
the variance of estimation of Bp(s0,t0)can be computed with
var Bp(s0,t0)=var M(s0,t0)+σI(s0,t0)2+var ˜
I(s0,t0). (11)
In the Gaussian process framework,
var Bp(s0,t0)
are referred to as the
posteriori mean and variance, respectively.
The conversion of the transformed variable
(i.e., the expected value Equation (10) and variance
of Equation (11)) into bird density
, is possible through the use of a quantile function,
which returns the bird density value bcorresponding to a given quantile ρ:
QB(ρ;s0,t0)=inf{b: Pr(B(s0,t0)<b)ρ}. (12)
Remote Sens. 2019,11, 2233 7 of 24
The quantile function allows to describe
because the quantile value
is preserved through a
power transform. Therefore, the quantile function of Bis computed with
Bp(ρ)1/p, (13)
is the cumulative distribution function of the Gaussian variable
by its mean (Equation (10)) and variance (Equation (11)).
Consequently, we choose to characterize the estimation of the bird density
by its median
and its quantiles 10 and 90 (i.e., uncertainty range).
2.4.2. Simulation
The geostatistical simulation of the random variable
consists of randomly drawing a realization
among the set of all possible outcomes defined by the probability of
conditional to the radar
observations (see Equations (10) and (11)) e.g., [
]. This is identical to sampling the posterior
probability in the Gaussian process framework.
Although kriging estimation is known to produce accurate point estimates, it leads to excessively
smooth interpolation maps [
], and thus fails to reproduce the fine-scale texture of the process at
hand. This causes problems for applications in which the space-time structure of the interpolation
map matters, such as when a nonlinear transformation is applied to the interpolated map. This is
the case in our model, as the back power transformation creates skewed distributions of bird density.
Computing the total number of birds migrating is a prime example of the necessity of simulation.
Indeed, integrating the bird density estimation map would greatly underestimates this number, because
of its inability to reproduce peaks of bird densities. Instead, integrating multiple realizations would
produce an accurate distribution of the total number of birds.
The simulation of both
is performed using Sequential Gaussian Simulation [
The simulation results in multiple realizations denoted as
}. The final realization of
is simply computed using,
I(`)(s0,t0)1/p. (14)
2.5. Validation
2.5.1. Cross-Validation
We tested the internal consistency of the model by cross-validation. It consists of sequentially
omitting the data of a single radar, then estimating bird densities at this radar location with the
model, and, finally, comparing the model-estimated value
to the observed data
The model is assessed by its ability to provide both the smallest misfit errors, i.e.,
and uncertainty ranges matching the magnitude of these errors. Because it is more convenient to
quantify these two aspects with a normal variable, we used the transformed variable
and quantified
the model performance with the normalized error of estimation, defined as
pvar (Bp(s,t)), (15)
var (Bp(s,t))
is the variance of the estimation as defined in Equation (11), which quantifies the
uncertainty of the estimation. The numerator of Equation (15) measures the misfit of the model
estimation, and the denominator normalizes this misfit according to the estimation uncertainty
provided by the model. For instance, a normalized error of 1 corresponds to the estimation value being
one estimated standard deviation above the measured value. Consequently, an ideal estimation should
produce normalized errors of estimation that follow a standard normal distribution, because (1) the
Remote Sens. 2019,11, 2233 8 of 24
estimation should be unbiased (mean of zero) and (2) the uncertainty provided by the model should
correspond to the observed error (variance of 1).
In addition, we performed the same cross-validation procedure using a nearest-neighbor
interpolation method instead, thus allowing to benchmark the proposed approach.
The root-mean-square error (RMSE) and coefficient of determination R
are used to assess
the performance of both the proposed model and the nearest-neighbor interpolation.
2.5.2. Comparison with Dedicated Bird Radars
A second validation of our modeling framework (from data acquisition by weather radars to
geostatistical interpolation) requires comparing the model-predicted birdmigration intensities with the
measurements of two dedicated bird radars (Swiss BirdRadar Solution AG, https://swiss-birdradar. located in Herzeele, France (50
N 2
E), and Sempach,
Switzerland (47
N 8
E) (see Figure 1). These radars are located 50 km and 84 km,
respectively, to the closest weather radar. By comparison, the distance of the grid nodes to their closest
weather radar is on average 92 km (10–90 quantiles: 35–166 km).
These bird radars continuously register echoes transiting through a conical shaped beam (17.5
nominal beam angle). The diameter of the radar beam cross-section varies from 50 m at 50 m agl to
500 m at 1500 m agl [
]. The individual echoes are aggregated over an hour to produce migration
traffic rates (MTR) (bird/km/h) and an average speed of birds aloft (km/h). The bird density measured
by the bird radar is computed by dividing the MTR by the mean speed [32].
Using the model presented, we estimate bird density at the exact locations of the bird radars.
In order to account for the difference of time resolution, the estimations are first computed every
15 min and then averaged over one hour. We subsequently assess the quality of the estimation and
uncertainty provided by the model by computing the normalized errors of estimation (Equation (9)).
In addition, we also compare our approach with a simple nearest-neighbor interpolation using the
RMSE and R2.
3. Results
3.1. Validation
3.1.1. Cross-Validation
The normalized error of estimation over all radars has a near-Gaussian distribution with a mean
of 0.01 and a variance of 1.08 (Figure A7 in Appendix D). The near-zero mean of the error distribution
indicates that the model provides nonbiased estimations of bird densities, where the near-one standard
deviation demonstrates that the model provides appropriate uncertainty estimates. The performance of
the cross-validation shows radar-specific biases (i.e., constant under- or overpredictions; Figure A8 in
Appendix Dand time series of each radar in Supplementary Material S2). The biases do not show any
clear spatial pattern (Figure A8 in Appendix D), suggesting that these radar-specific biases probably
originate from measurement errors, such as birds nonaccounted for (e.g., flying below the radar) or
errors in the cleaning procedure (e.g., ground scattering).
Compared to a nearest-neighbor interpolation, the model performs better in the cross-validation
with a RMSE of 17.67 bird/km
and R
of 0.59 against a RMSE of 23.17 bird/km
and R
of 0.29 for the
nearest neighbor.
Remote Sens. 2019,11, 2233 9 of 24
3.1.2. Comparison with Dedicated Bird Radars
The daily migration patterns estimated by the model generally coincide well with the observations
derived from dedicated bird radars (Figure 4). Over the whole validation period, the normalized
estimation error has a mean of 0.49 and a variance of 0.77 at Herzeele radar location, and a mean
0.68 and a variance of 1.45 at Sempach radar location. These normalized estimation errors
indicate a tendency of the model to slightly overestimate bird densities in Herzeele (i.e., mean above 1)
and to underestimate them in Sempach, while providing overconfident uncertainty in Herzeele and
underconfident at Sempach.
These errors translate into a RMSE of 9.2 and 19.7 bird/km
and R
of 0.87 and 0.73 for the radar
in Herzeele and Sempach, respectively. Reference [
] reported relatively similar values of R
comparing the MTR of close-by weather radar and bird radar. By comparison, the nearest-neighbor
approach yields a RMSE of 9.7 and 31.7 bird/km2, and a R2of 0.85 and 0.59, respectively.
Bird densities [bird/km2]
Uncertainty range
Closest radar
Bird radar
Sep 19 Sep 22 Sep 25 Sep 28 Oct 01 Oct 04 Oct 07 Oct 10
Date 2016
Figure 4.
Comparison of the estimated bird densities (black line) and their uncertainty range
(10–90 quantiles in gray) with the bird densities (red dots) observed using dedicated bird radars
at two locations in (
) Herzeele, France (50
N 2
E), and (
) Sempach, Switzerland
N 8
E). Note that, because of the power transformation, model uncertainties are
larger when the migration intensity is high. It is therefore critical to account for the uncertainty ranges
(light gray) when comparing the interpolation results with the bird radars observations (red dots).
3.2. Application to Bird Migration Mapping
The main outcome of our model is to estimate bird densities at any time and location within the
domain of interest. This is illustrated by the estimation of bird density time series at specific locations
(e.g., Figure 4) and by the generation of bird density maps at different time steps (Figure 5).
Remote Sens. 2019,11, 2233 10 of 24
Although the estimation represents the most likely value of bird density (i.e., mean of estimation)
at each node of the grid (e.g., Figure 5), the simulation provides multiple values of bird density
according to their conditional distribution and reproduces more accurately the fine-scale patterns
of migration (e.g., Figure 6). As explained in Section 2.4.2, the amplitude of peak migration is more
adequately illustrated in the realizations (Figure 6), compared to the smooth estimation map (Figure 5).
For each of the 100 realizations, we computed the total number of birds flying over the whole
domain for each time step (Figure 7b). Within the time periods considered in this study, the peak
migration occurred in the night of 4–5 October with up to 121 million (10–90 quantiles: 118–124) birds
flying simultaneously. Computing this on subdomains, such as countries, highlights geographical
differences in migration intensity. For instance, during the same night, France had 37 (35–39) million
birds aloft (89 bird/km
), Poland had 14 (13–15) million (65 bird/km
), and Finland only 2 (1.9–2.2)
million (30 bird/km2) (Figure 7c).
1 10 100 250
Daily average bird density [bird/km2]
18:00 19:00 20:00 21:00 22:00 23:00
00:00 01:00 02:00 03:00 04:00 05:00
Figure 5.
Maps of bird density estimation every hour of a single night (3–4 October). The sunrise and
sunset fronts are visible at 18:00 and 05:00 with lower densities close to the fronts and no value after the
front. The resemblance from hour-to-hour illustrates the high temporal continuity of the model. A rain
cell above Poland blocked migration on the eastern part of the domain. In contrast, a clear pathway is
visible from Northern Germany to Southwestern France.
Bird densities [bird/km2]
Realization 1 Realization 2 Realization 3
Figure 6.
Snapshot of three different realizations showing peak migration (4 October 2016 21:00 UTC).
The total number of birds in the air for these realizations was 125, 126, and 122 million, respectively.
Comparing the similarities and differences of bird density patterns among the realizations illustrates
the variability allowed by the stochastic model. The texture of these realizations is more coherent with
the observations than the smooth estimation map in Figure 5.
The spatio-temporal dynamics of bird migration can be visualized with an animated and
interactive map (available online at with a user manual
Remote Sens. 2019,11, 2233 11 of 24
provided in Appendix E), produced with an open source script (
BMM-web). In the web app, users can visualize the estimation maps or a single simulation map
animated in time, as well as time series of bird densities of any location on the map. In addition, it is
also possible to compute the number of birds over a custom area and to download all data through a
dedicated API.
Sep 19 Sep 25 Oct 01 Oct 07
Date 2016
Total number of bird aloft [millions]
Full domain France Poland Finland Uncertainty range
(a) (b)
Figure 7.
Averaged time series of birds in migration and their associated uncertainty ranges (10–90
quantiles) over (b) the whole domain, and (c) France, Poland and Finland. (a) illustrates the location
and size of the three countries.
4. Discussion
The model developed here can estimate bird migration intensity and its uncertainty range on
a high-resolution space-time grid (0.2
lat. lon. and 15 min.). The highest total number of birds
flying simultaneously over the study area is estimated at 121 million, corresponding to an average
density of 52 bird/km
. This number illustrates the impressive magnitude of nocturnal bird migration,
and resembles the values of peak migration estimated over the USA with 500 million birds and a
similar average density of 51 bird/km
]. For more local results, interactive maps of the resulting
bird density are available on a website with a dedicated interface that facilitates the visualization and
export of the estimated bird densities with their associated uncertainty (https://birdmigrationmap., see Appendix Efor a user manual).
4.1. Advantages and Limitations
This paper presents the first spatio-temporal interpolation of nocturnal bird densities at the
continental scale that accounts for sub-daily fluctuations and provides uncertainty ranges. In contrast
to the methods based on covariates, deemed more reliable for extrapolation in the future [
our interpolation approach relies solely on the strong space-time correlation of bird migration and
consequently does not require any external covariate per se (e.g., temperature, rain, or wind). Similarly,
local features such as the proximity to the ocean or the presence of mountains were not explicitly
accounted for in the model. Yet, the influence of these meteorological and geographical features on bird
migration is largely captured by the measurements of weather radars, so that, in turn, the interpolation
implicitly accounts for them. However, to take full advantage of these covariates, often available at
extensive scale, the model could be adapted to consider linear relationships with covariates, using the
standard frameworks of regression kriging, or possibly full co-kriging [23].
Remote Sens. 2019,11, 2233 12 of 24
The fitted parameters of the model provide information on the broad scale bird migration (see
Appendix B). In particular, the covariance function of the model describes the general spatio-temporal
scale at which migration is happening (Figure A5 in Appendix B). For instance, even with the spatial
trend removed, the multi-night scale of bird migration (
) correlates at 50% at a distance of 300 km
(25%–500 km) and at 60% for one day to the next (25% at 3 days). These ranges qualitatively describe
the spatio-temporal extent of broad-front migration in the midst of the autumn migration season,
and highlight the importance of international cooperation for data acquisition and for the spread of
warning systems during peak migration events.
As a proof-of-concept, we used three weeks of bird density data available on the ENRAM data
repository (see Data Accessibility). As more data from weather radars become available, our analyses
can be extended to year-round estimations of migration intensity at the continental scale where weather
radar data are available with a good coverage, as is the case in Europe and North America. We also
significantly preprocessed the bird density data, i.e., restricted our model to nocturnal movements,
and applied a strict manual data cleaning. This is because the bird density data presently made
available can be strongly contaminated with the presence of insects during the day, and birds flying at
low altitude are not reliably recorded by radars because of ground clutter and the radar position in
relation to its surrounding topography. Once the quality of the bird density data has improved [
our model can be implemented in near-real-time and provide continuous information to stakeholders
and the public and private sectors.
Although the model introduced here is already a valuable tool for bird migration mapping, we see
several avenues for further development. For instance, in applications where the distribution of
bird density over altitudes is crucial, the model can be extended to explicitly incorporate the vertical
dimension. Furthermore, if fluxes of birds, i.e., migration traffic rates, are sought after, a similar
geostatistical approach can be used to interpolate the flight speeds and directions that are also derived
from weather radar data. If one has access to the raw radial velocity, the method developed by [
would produce more accurate results.
4.2. Applications
Many applied problems rely on high-resolution estimates of bird densities and migration
intensities, and the model developed here lays the groundwork for addressing these challenges.
For instance, such migration maps can help identify migration hotspots, i.e., areas through which many
aerial migrants move, and thus, assist in prioritizing conservation efforts. Furthermore, mitigating
collision risks of birds by turning off artificial lights on tall buildings or shutting down wind energy
installations requires information on when and where migration intensity peaks. The probability
distribution function of our model can provide such information as it estimates when and where
migration intensity exceeds a given threshold. Such information can be used in shut-down on demand
protocols for wind turbine operators or trigger alarms to infrastructure managers.
Supplementary Materials:
The illustrations of the cleaned vertical-integrated time series of nocturnal bird
densities for all radars are available in supplementary materials
2233/s1. The illustrations of the cross-validation of all radars are available in supplementary materials http:
// The MATLAB livescripts of this article are available at https:
// The cleaned vertical time series profile are available at doi: https://doi.
org/10.5281/zenodo.3243397 [
]. and the final interpolated maps are available at doi:
zenodo.3243466 [
]. The codes of the website (HTML, Js, NodeJs, Css) are available at
Author Contributions:
R.N., L.B., F.L., and B.S. conceived the study; R.N., L.B., and G.M. designed the
geostatistical model; R.N. developed and implemented the computational framework; and R.N., L.B., and B.S.
performed the analyses and wrote a first draft of the manuscript; G.M., F.L. and S.B. contributed substantially to
the manuscript.
Remote Sens. 2019,11, 2233 13 of 24
We acknowledge the financial support from the Globam project, funded by BioDIVERSA, including the
Swiss National Science Foundation (31BD20_184120), Netherlands Organisation for Scientific Research (NWO
E10008), and Academy of Finland (aka 326315), BelSPO BR/185/A1/GloBAM-BE.
This study contains modified Copernicus Climate Change Service Information 2019. Neither
the European Commission nor ECMWF is responsible for any use that may be made of the Copernicus
Information or Data it contains. We acknowledge the European Operational Program for Exchange of Weather
Radar Information (EUMETNET/OPERA) for providing access to European radar data, facilitated through a
research-only license agreement between EUMETNET/OPERA members and ENRAM (European Network for
Radar surveillance of Animal Movements). Mathieu Boos (Research Agency in Applied Ecology, Naturaconsta,
Wilshausen, France) kindly provided the BirdScan data for Herzeele in France.
Conflicts of Interest: The authors declare no conflicts of interest.
Appendix A. Data Preprocessing
Appendix Adescribes the full procedure applied to manually clean the time series of bird
densities. The raw dataset of bird density downloaded from the ENRAM data repository (https:
// has been previously published in [
]. The steps detailed
below are illustrated in Figure A1 for the radar located in Zaventem, Belgium (50
N, 4
Of the 84 radars contributing data during the study period, 11 radars are discarded because of
their poor quality due to S-band radar type, poor processing, or large gaps (temporal or altitude
cut). The same radars were removed in [
]. In addition, the four radars from Bulgaria and
Portugal were excluded because of their geographic isolation.
The full vertical profile was discarded when rain was present at any altitude bin (purple rectangle
in Figure A1). A dedicated MATLAB GUI was used to visualize the data and manually set bird
densities to “not-a-number” in such cases.
Zones of high bird densities can sometimes be incorrectly eliminated in the raw data (red rectangle
in Figure A1). To address this, reference [
] excluded problematic time or height ranges from
the data. Here, in order to keep as much data as possible, the data was manually edited to
replace erroneous data either with “not-a-number”, or by cubic interpolation using the dedicated
Due to ground scattering (brown rectangle in Figure A1), the lower altitude layers are sometimes
contaminated by errors or excluded in the raw data. We vertically interpolated bird density by
copying the first layer without error into to the lower ones. This approach is relatively conservative
as bird migration intensity usually decreases with height in the absence of obstacles, and more so
in autumn [39].
The vertical profiles were vertically integrated from the radar ground level (black line in
Figure A1c) and up to 5000 m asl.
The data recorded during daytime are excluded. Daytime is defined for each radar by the civil
dawn and dusk (sun 6below horizon).
Finally, the data of 10 radars with high temporal resolution (5–10 min) was downsampled to
15 min to preserve a balanced representation of each radar.
The resulting cleaned time series of nocturnal bird densities are displayed in Figure A1d.
We provide the same figure than Figure A1 for all radar data in Supplementary Materials http:
Remote Sens. 2019,11, 2233 14 of 24
09/20 09/22 09/24 09/26 09/28 09/30 10/02 10/04 10/06 10/08
Altitude [km]
Bird densities [bird/km2]
Bird densities [bird/km3]
Rain Ground Scattering
Figure A1.
Illustration of the cleaning procedure for the data of the Zaventem radar in Belgium
N, 4
E). (
) Raw reflectivity. (
) Raw data of vertical bird density profiles. (
) Manually
cleaned vertical bird profiles. (d) Final bird densities (integrated over all altitudes).
Appendix B. Model Parametrisation
Appendix Bpresents the method of model parametrisation and discusses the meaning of model
parameters in terms of bird migration. Table A1 summarizes the calibrated parameters.
Table A1. Calibrated parameters.
Power transformation p=0.133
Spatial trend w0=2.566 , wlat =0.024
Covariance of M C0=0.006, Cg=0.032, rt=1.24, rs=500, α=0.98, γ=0.71, β=0.95
Curve a=[0.04, 0.10, 0.07, 0.27, 1.29, 0.59, 2.86, 0.44, 1.92]
Curve variance b=[0.00, 0.00, 0.02, 0.04, 0.17, 0.17, 0.62, 0.26, 0.93, 0.12, 0.49]
Covariance of I C0=0.009, Cg=0.91, rt=0.07, rs=190, α=1, γ=0.4, β=1
Appendix B.1. Power Transform p
The value of power transformation
is inferred by minimizing the Kolmogorov–Smirnov criterion
of the p-transformed observations
. The Kolmogorov–Smirnov test (Massey, 1951) assesses the
hypothesis that the observations
are normally distributed. The optimal power transformation
parameter was found for
7.5 and the resulting
histogram is displayed in Figure A2 together
with the initial data B.
Remote Sens. 2019,11, 2233 15 of 24
100 200 300 400
Histogram of bird densities [bird/km2]
0.4 0.8 1.2 1.6 22.4
Histogram of transformed bird densities
Figure A2.
Histogram of the raw bird densities data
) and the power transformed bird densities
Bpfor the calibrated parameter p=1/7.5 (bottom).
The fitted distribution shows that the distribution of bird densities is highly skewed: the lowest
10% are below 1 bird/km
, the upper 10% are above 50 bird/km
, and the maximum density reaches
500 bird/km
. Consequently, the central value (mean of 19 bird/km
or median of 8 bird/km
) and
variance (753 bird
) do not adequately characterize the distribution. A power transformation on
such skewed data creates significant nonlinear effects in the back-transformation. For instance, the
symmetric uncertainty of an estimated value in the transformed space (quantified by the variance of
estimation) will become highly skewed in the original space. Consequently, the uncertainty of the
estimation of bird densities is highly dependent on the value of the power transform: low densities
estimations have a smaller uncertainty than high densities.
Such effects also have consequences from an ecological and conservation point of view. Indeed,
efficient protection of birds along the migration route (from artificial light or wind turbines) requires
particular attention to the peak densities, during which the majority of birds are moving in a few nights.
These peaks can only be accurately reproduced by accounting for the high tail of the distribution.
This is done here by using a full distribution to quantify the uncertainty of the estimation.
Appendix B.2. Spatial Trend µ
The parameters of the spatial trend (
) are calibrated on the nightly average of each
radar with an ordinary least square method. The resulting planar trend is shown in Figure A3a
together with the average transformed bird densities of each radar. The trend displays a strong
north–south gradient, which can be explained by the higher migration activity in southern Europe
during the study period. A 2-dimensional planar trend was initially tested in order to accommodate
the northeast–southwest flyway. However, this more complex model did not significantly improve
the fit with data, and was therefore discarded. The detrended values illustrated in Figure A3b are
more stationary, with the exception of Finland and Sweden. Reference [
] also noted this difference
between these countries, but excluded the fact that this difference is due to errors in the data since
the southern Swedish radar shows consistent amounts of migratory movements with a neighboring
German site. Figure A3b highlights the central European continental flyway as illustrated by the arrow.
Remote Sens. 2019,11, 2233 16 of 24
Transformed bird densities
Transformed bird densities
Figure A3.
Fitted trend with averaged observations at radar location (
) and detrended data (
Appendix B.3. Curve Trend ιand Variance σI
Figure A4 displays the intra-night scale component
of the weather radar data
together with the fitted polynomial curve trend
(Equation (6)) and polynomial variance function
(Equation (8)). The curve reveals that migration is mainly concentrated between 10% and 90% of the
night-time with slightly larger densities of birds in the first half of the night. The larger variance at the
beginning and end of the night is partly due to the power-transformation of the raw data.
-1 -0.5 0
Normalized Night Time (NNT)
Transformed bird densities
-0.5 1
Figure A4.
Intra-night scale component of the weather radar data (black dots) and fitted curve trend
(black line). The shaded gray areas each denote 1-, 2-, and 3-σIfitted.
Appendix B.4. Covariance Functions of M and I
The parameters of the space-time covariance functions (
, and
) of
(Equation (4)) and
(Equation (9)) are inferred by minimizing the misfit between the covariance
function and the empirical covariances of the weather radar data. The empirical covariances are
derived for several lag-distances and lag-times on an irregular grid.
The calibrated covariance functions provide some information about the degree of spatial and
temporal correlation of the bird migration process. Indeed, the absolute value of bird density covariance
should also include the effect of trend and power transformation. The fitted spatial covariance at the
multi-night scale
(Figure A5a) is fitted with a large spatial nugget (16%), suggesting a significant
variability of bird density uncorrelated in space. This can be explained by either local features of the
migration process or radar measurement errors. It is important to recall that since the weather radars
are relatively well-spread (Figure 1b), the spatial covariance of both
is poorly constrained for
small lag-distances (approximately less than 100 km), and consequently the uncertainty of the nugget
value is large. The temporal covariance of
has an asymptotic behavior with 20% covariance after
Remote Sens. 2019,11, 2233 17 of 24
6 days (Figure A5b) because no temporal trend was used. Note that as the covariance of
is evaluated
only on a discrete 1-day lag-time, the shape of the covariance between 0 and 1 is artificially created to
fit the Gneiting function.
account for most of the spatial covariance of bird density, but some
spatial correlation is still present at the intra-night scale as suggested by Figure 1c. The temporal
correlation of
is high for short lags (67% at 0.04 days, or 1 hr), indicating consistent measurements
from each weather radar.
Distance [km] Time [day]
200 400 600 8000
200 400 600 8000
0 0.04 0.08 0.12 0.16 0.2
0 2 4 6
1 3 5
(c) (d)
Figure A5.
Illustration of the calibrated covariance function of the multi-night scale
) and the
intra-night scale ˜
Appendix C. Kriging
Appendix Cprovides detailed explanations for the kriging estimation. Note that this procedure
is identical to Gaussian regression in the field of machine learning, where the kriging estimation is
equivalent to the mean of the posterior distribution (i.e., conditional to the known location).
Kriging provides an estimated value of a random variable
) at the target point
(s0,t0)based on a linear combination of known data points {X(sα,tα)}α=1,...,n0with
λαX(sα,tα). (A1)
The kriging weights
Λ=[λ1,· · · ,λn0]T
are determined by minimizing the variance of the
estimation under unbiased conditions which leads to the following linear system, commonly referred
to as the kriging system,
Cα,αΛ=Cα,0, (A2)
is the cross-covariance matrix of the observations and
is the covariance matrix between
the observations and the target point. These covariances are computed using the fitted covariance
function of Equation (4) or Equation (9). The kriging weights can be solved using
and are subsequently used in Equation (A1) to compute the kriging estimate. Figure A6 illustrates the
value of the kriging weights for the kriging estimation of the multi-night scale at an unknown location
(red dot).
Remote Sens. 2019,11, 2233 18 of 24
Latitude 15
45 5
Kriging Weight Value (log-scale)
Figure A6.
Illustration of the kriging weight values computed for the kriging estimation of
at an
unknown location (red dot). Here we illustrate only the neighbors within +/
1 day and a spatial
extent of 600 km.
In addition to the estimated value
, kriging also provides a measure of uncertainty with
the variance of estimation,
var X(s0,t0)=CX(0, 0)Λ>Cα,0 (A3)
Appendix D. Cross-Validation
Appendix Dprovides additional details on the cross-validation described in Section 2.5.1.
Figure A7 displays the histogram of the normalized error of estimation (Equation (15)) when all
data across all radars are considered. The mean of the distribution is close to zero which indicates that
the estimation is unbiased (i.e., on average, the estimation neither underestimates (mean below 1) nor
overestimates (mean above 1) bird density). Its variance is slightly above 1, which indicates a small
tendency to underestimate the uncertainty range (i.e., on average, the kriging variance is too small).
Overall, the cross-validation indicates a good performance of the model.
In order to test radar-specific performance, the normalized error of estimation is computed for
each radar. As displayed in Figure A8, their respective mean and variance do not reveal any clear
spatial pattern, thus suggesting no spatial biases in the estimation. However, the large absolute value of
the mean of the normalized error of estimation (color scale) indicates a strong variability in the model
performance for each radar. The model underestimates or overestimates bird density at some radar
locations with a mean of 1.5 times the variance of estimation. The time series comparing estimations of
the cross-validation with the observed data for each radar are provided in Supplementary Material
Remote Sens. 2019,11, 2233 19 of 24
-4 -2
Normalized error of estimation with mean: 0.01 and variance: 1.08
2 40
Figure A7.
Histogram of the normalized error of estimation (Equation (15)) over all radars. The red
curve is the standard normal distribution which should be matched by the histogram in case of ideal
Figure A8.
Illustration of the performance of the cross-validation with the mean and variance of the
normalized error of estimation (Equation (15)) at each radar location. Negative or positive mean values
respectively indicate an under- or overestimation of the model estimation. The reproduction of the
variance is illustrated by a black circle: a perfect variance would match the color circle while a smaller
circle indicates an underconfidence (uncertainty range too large).
Appendix E. Manual for Website Interface
Appendix Edescribes the web interface developed for the visualization and querying of the
interpolated data. The website is available on and the code
on The web interface is displayed in Figure A9
with annotations and further details below.
Remote Sens. 2019,11, 2233 20 of 24
Figure A9. Illustration of the website interface with annotations for each interactive component.
Appendix E.1. Block 1: Interactive Map
The main block of the website is a map with interactive visualization tools (e.g., zoom and pan).
On top of this map, three layers can be displayed:
The first layer illustrates bird densities in a log-color scale. This layer can display either the
estimation map or a single simulation map. Users can choose using the drop-down menu (1a).
The second layer displays the rain in light-blue. The layer can be hidden/displayed using the
checkbox (1b).
The third layer corresponds to bird flight speed and direction, visualized by black arrows.
The checkbox (1c) allows users to display/hide this layer. Finally, the menu (1d) provides a
link to (1) documentation, (2) model description, (3) Github repository, (4) MATLAB livescript,
and (5) Researchgate page.
Appendix E.2. Block 2: Time Series
The second block (hidden by default on the website) shows three time series, each in a different tab (2a):
Densities profile shows the bird densities [bird/km2] at a specific location.
Sum profile shows the total number of birds [bird] over an area.
MTR profile shows the mean traffic rate (MTR) [bird/km/h] perpendicular to a transect.
A dotted vertical line (2d) appears on each time series to show the current time frame displayed
on the map (Block 1). Basic interactive tools for the visualization of the time series include zooming
on a specific time period (day, week or all periods) (2b) and general zoom and auto-scale functions
(2e). Each time serie can be displayed or hidden by clicking on its legend (2c). The main feature of
this block is the ability to visualize bird densities at any location chosen on the map. For the densities
profile tab, the button with a marker icon (2f) allows users to plot a marker on the map, and displays
the bird densities profile with uncertainty (quantile 10 and 90) on the time series corresponding to this
Remote Sens. 2019,11, 2233 21 of 24
location. Users can plot several markers to compare the different locations (Figure A10). Similarly, for
the sum profile, the button with a polygon icon (2f) allows users to draw a polygon and returns the
time series of the total number of birds flying over this area (Figure A11). For the MTR tab, the flux of
birds is computed on a segment (line of two points) by multiplying the bird densities with the local
flight speed perpendicular to that segment.
Appendix E.3. Block 3: Time Control
The third block shows the time progression of the animated map with a draggable slider (3d).
Users can control the time with the buttons play/pause (3b), previous (3a) and next frame (3c).
The speed of animation can be changed with a slider (3e).
Appendix E.4. API
An API based on mangodb and NodeJS allows users to download any time serie described in
Block 2. Instructions can be found on
Appendix E.5. Examples
Figure A10.
A print-screen of the web interface developed to visualize the dataset. The map shows the
estimated bird densities for the 23 September 2016 at 21:30 with the rain mask appearing in light blue
on top. The domain extent is illustrated by a black box. The time series in the bottom show the bird
densities with quantile 10 and 90 at the two locations symbolized by the markers with corresponding
color on the map. The button with a marker symbol on the right side of the time series allows users to
query any location on the map and to display the corresponding time series.
Remote Sens. 2019,11, 2233 22 of 24
Figure A11.
Print-screen of the web interface with the simulation map for the 3 October 2016 at 23:00.
The bottom pane shows the time series of the total number of birds within the color-coded polygon
drawn on the map. The button with the polygon symbol on the right side of the time series allows to
query the total number of birds flying within any polygon drawn on the map.
References and Note
Hahn, S.; Bauer, S.; Liechti, F. The natural link between Europe and Africa—2.1 billion birds on migration.
Oikos 2009,118, 624–626, doi:10.1111/j.1600-0706.2008.17309.x.
Dokter, A.M.; Farnsworth, A.; Fink, D.; Ruiz-Gutierrez, V.; Hochachka, W.M.; La Sorte, F.A.;
Robinson, O.J.
Rosenberg, K.V.; Kelling, S. Seasonal abundance and survival of North America’s migratory avifauna
determined by weather radar. Nat. Ecol. Evol. 2018,2, 1603, doi:10.1038/s41559-018-0666-4.
Bauer, S.; Hoye, B.J. Migratory animals couple biodiversity and ecosystem functioning worldwide. Science
2014,344, doi:10.1126/science.1242552.
Runge, C.A.; Martin, T.G.; Possingham, H.P.; Willis, S.G.; Fuller, R.A. Conserving mobile species.
Front. Ecol. Environ. 2014,12, 395–402, doi:10.1890/130237.
Sanderson, F.J.; Donald, P.F.; Pain, D.J.; Burfield, I.J.; van Bommel, F.P. Long-term population declines in
Afro-Palearctic migrant birds. Biol. Conserv. 2006,131, 93–105, doi:10.1016/j.biocon.2006.02.008.
Vickery, J.A.; Smith, K.W.; Pain, D.J.; Gregory, R.D.; Škorpilová, J.; Bairlein, F.; Ewing, S.R. The decline of
Afro-Palaearctic migrants and an assessment of potential causes. Ibis
,156, 1–22, doi:10.1111/ibi.12118.
7. Diehl, R.H. The airspace is habitat. Trends Ecol. Evol. 2013,28, 377–379, doi:10.1016/j.tree.2013.02.015.
La Sorte, F.A.; Horton, K.G.; Nilsson, C.; Dokter, A.M. Projected changes in wind assistance under
climate change for nocturnally migrating bird populations. Glob. Chang. Biol.
,25, 589–601,
Van Doren, B.M.; Horton, K.G.; Dokter, A.M.; Klinck, H.; Elbin, S.B.; Farnsworth, A. High-intensity
urban light installation dramatically alters nocturnal bird migration. Proc. Natl. Acad. Sci. USA
114, 11175–11180, doi:10.1073/pnas.1708574114.
Winger, B.M.; Weeks, B.C.; Farnsworth, A.; Jones, A.W.; Hennen, M.; Willard, D.E. Nocturnal flight-calling
behaviour predicts vulnerability to artificial light in migratory birds. Proc. Biol. Sci.
,286, 20190364,
Remote Sens. 2019,11, 2233 23 of 24
Aschwanden, J.; Stark, H.; Peter, D.; Steuri, T.; Schmid, B.; Liechti, F. Bird collisions at wind turbines in a
mountainous area related to bird movement intensities measured by radar. Biol. Conserv.
,220, 228–236,
van Gasteren, H.; Krijgsveld, K.L.; Klauke, N.; Leshem, Y.; Metz, I.C.; Skakuj, M.; Sorbi, S.; Schekler, I.;
Shamoun-Baranes, J. Aeroecology meets aviation safety: early warning systems in Europe and the Middle
East prevent collisions between birds and aircraft. Ecography 2019,42, 899–911, doi:10.1111/ecog.04125.
Loss, S.R.; Will, T.; Marra, P.P. Direct Mortality of Birds from Anthropogenic Causes. Annu. Rev. Ecol.
Evol. Syst. 2015,46, 99–120, doi:10.1146/annurev-ecolsys-112414-054133.
14. Winkler, R. Avifaune de Suisse; Nos Oiseaux; 1999; Volume 3.
Drake, A.V.; Bruderer, B. Aeroecological Observation Methods. In Aeroecology; Chilson, P.B., Frick, W.F.,
Kelly, J.F., Liechti, F., Eds.; Springer International Publishing: Cham, Switzerland, 2017; pp. 201–237,
Gauthreaux, S.A.; Belser, C.G.; van Blaricom, D. Using a Network of WSR-88D Weather Surveillance Radars
to Define Patterns of Bird Migration at Large Spatial Scales. In Avian Migration; Springer: Berlin/Heidelberg,
Germany, 2003; pp. 335–346, doi:10.1007/978-3-662-05957-9_23.
Shamoun-Baranes, J.; Alves, J.A.; Bauer, S.; Dokter, A.M.; Hüppop, O.; Koistinen, J.; Leijnse, H.; Liechti, F.;
van Gasteren, H.; Chapman, J.W. Continental-scale radar monitoring of the aerial movements of animals.
Mov. Ecol. 2014,2, 4–9, doi:10.1186/2051-3933-2-9.
Bauer, S.; Chapman, J.W.; Reynolds, D.R.; Alves, J.A.; Dokter, A.M.; Menz, M.M.; Sapir, N.; Ciach, M.;
Pettersson, L.B.; Kelly, J.F.; et al. From Agricultural Benefits to Aviation Safety: Realizing the Potential of
Continent-Wide Radar Networks. BioScience 2017,67, 912–918, doi:10.1093/biosci/bix074.
Van Doren, B.M.; Horton, K.G. A continental system for forecasting bird migration. Science
361, 1115–1118, doi:10.1126/science.aat7526.
Huuskonen, A.; Saltikoff, E.; Holleman, I. The operational weather radar network in Europe. Bull. Am.
Meteorol. Soc. 2014,95, 897–907, doi:10.1175/BAMS-D-12-00216.1.
Nilsson, C.; Dokter, A.M.; Verlinden, L.; Shamoun-Baranes, J.; Schmid, B.; Desmet, P.; Bauer, S.; Chapman, J.;
Alves, J.A.; Stepanian, P.M.; et al. Revealing patterns of nocturnal migration using the European weather
radar network. Ecography 2019,42, 876–886, doi:10.1111/ecog.04003.
Copernicus Climate Change Service (C3S). ERA5: Fifth Generation of ECMWF Atmospheric Reanalyses
of the Global Climate. Available online: (accessed on 20
February 2018)..
Chilès, J.P.; Delfiner, P. Geostatistics;Wiley Series in Probability and Statistics; John Wiley & Sons, Inc.: Hoboken,
NJ, USA, 1999; Volume 497, doi:10.1002/9780470316993.
Lantuéjoul, C. Geostatistical Simulation; Springer: Berlin/Heidelberg, Germany, 2002;
Williams, C.K.I.; Rasmussen, C.E. Gaussian Processes for Machine Learning; MIT Press: Cambridge, MA, USA,
2006; Volume 2.
Gneiting, T. Nonseparable, stationary covariance functions for space time data. J. Am. Stat. Assoc.
97, 590–600.
Bruderer, B.; Liechti, F. Variation in density and height distribution of nocturnal migration in the south of
israel. Isr. J. Zool. 1995,41, 477–487, doi:10.1080/00212210.1995.10688815.
Goovaerts, P. Geostatistics for Natural Resources Evaluation; Oxford University Press: Oxford, UK, 1997; p. 483.
Gómez-Hernández, J.J.; Cassiraga, E.F. Theory and Practice of Sequential Simulation; Kluwer Academic
Publishers: Dordrecht, The Netherlands, 1994; pp. 111–124, doi:10.1007/978-94-015-8267-4_10.
30. Deutsch, C.V.; Journel, A.G. GSLIB: Geostatistical Software Library and User’s Guide; Oxford University Press:
New York, NY, USA, 1992; p. 147.
Schmid, B.; Zaugg, S.; Votier, S.C.; Chapman, J.W.; Boos, M.; Liechti, F. Size matters in quantitative radar
monitoring of animal migration: Estimating monitored volume from wingbeat frequency. Ecography
42, 931–941, doi:10.1111/ecog.04025.
Nilsson, C.; Dokter, A.M.; Schmid, B.; Scacco, M.; Verlinden, L.; Bäckman, J.; Haase, G.; Dell’Omo, G.;
Chapman, J.W.; Leijnse, H.; et al. Field validation of radar systems for monitoring bird migration. J. Appl. Ecol.
2018,55, 2552–2564, doi:10.1111/1365-2664.13174.
Remote Sens. 2019,11, 2233 24 of 24
Erni, B.; Liechti, F.; Underhill, L.G.; Bruderer, B. Wind and rain govern the intensity of nocturnal bird
migration in central Europe-a log-linear regression analysis. Ardea 2002,90, 155–166.
Van Belle, J.; Shamoun-Baranes, J.; Van Loon, E.; Bouten, W. An operational model predicting autumn bird
migration intensities for flight safety. J. Appl. Ecol. 2007,44, 864–874, doi:10.1111/j.1365-2664.2007.01322.x.
Lin, T.; Winner, K.; Bernstein, G.; Mittal, A.; Dokter, A.M.; Horton, K.G.; Nilsson, C.;
Van Doren, B.M.
Farnsworth, A.; La Sorte, F.A.; et al. MISTNET: Measuring historical bird migration in the US
using archived weather radar data and convolutional neural networks. Methods Ecol. Evol.
Angell, R.; Sheldon, D.R. Inferring Latent Velocities from Weather Radar Data Using Gaussian Processes; Number
Nips; Neural Information Processing Systems (NIPS); Bengio, S., Wallach, H., Larochelle, H., Grauman, K.,
Cesa-Bianchi, N., Garnett, R., Eds.; Curran Associates, Inc.: Montréal, QC, Canada, 2018; pp. 8984–8993.
Nussbaumer, R.; Benoit, L.; Mariethoz, G.; Liechti, F.; Bauer, S.; Schmid, B. Interpolated Maps of Bird Density
and Flight Vector over Europe [Data Set]. Zenodo. 2019. Available online:
3243465 (accessed on 17 June 2019).
Nussbaumer, R.; Benoit, L.; Schmid, B. Vertical Profiles and Integrated Time Series of Bird Density and Flight
Speed Vector (19–10 October 2016) [Data Set]. Zenodo. 2019. Available online:
3243396 (accessed on 17 June 2019).
Bruderer, B.; Peter, D.; Korner-Nievergelt, F. Vertical distribution of bird migration between the Baltic Sea
and the Sahara. J. Ornithol. 2018,159, 315–336, doi:10.1007/s10336-017-1506-z.
2019 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (
... However, most studies so far have focused on specific stages of the migration journey: migratory flights (e.g. [2,[6][7][8][9]) or stopovers (e.g. [10][11][12]). ...
... [26]). However, we employed a complete dataset (in space and time) of bird density and velocity interpolated from weather radar measurements (figure 1) [8], allowing us to apply directly the advection equation without assuming an underlying process of movement (e.g. movement proportional to the gradient of a quantity). ...
... It consists of vertical profiles of bird density [birds km −3 ], flight speed . First, we interpolate vertical profile time series of bird density and velocity field measured by weather radar data into continuous spatio-temporal maps following [8]. 2. Flow model ( §2.3). ...
Full-text available
To understand the influence of biomass flows on ecosystems, we need to characterize and quantify migrations at various spatial and temporal scales. Representing the movements of migrating birds as a fluid, we applied a flow model to bird density and velocity maps retrieved from the European weather radar network, covering almost a year. We quantified how many birds take-off, fly, and land across Western Europe to (1) track bird migration waves between nights, (2) cumulate the number of birds on the ground and (3) quantify the seasonal flow into and out of the study area through several regional transects. Our results identified several migration waves that crossed the study area in 4 days only and included up to 188 million (M) birds that took-off in a single night. In spring, we estimated that 494 M birds entered the study area, 251 M left it, and 243 M birds remained within the study area. In autumn, 314 M birds entered the study area while 858 M left it. In addition to identifying fundamental quantities, our study highlights the potential of combining interdisciplinary data and methods to elucidate the dynamics of avian migration from nightly to yearly time scales and from regional to continental spatial scales.
... Continental networks of weather radars are increasingly becoming essential tools for monitoring 7 large-scale migratory movements . However, most studies so far have focused on specific 8 stages of the migration journey: migratory flights (e.g., Dokter et al., 2018;Horton et al., 2020;Nilsson et al., 9 2019; Nussbaumer et al., 2019;Van Doren & Horton, 2018), or stop-overs (e.g., Buler et al., 2017;Cohen et 10 al., 2020;McLaren et al., 2018). Yet, none have explicitly considered and differentiated between the three 11 successive stages of take-off, flight and landing, and we therefore lack a comprehensive model of the entire 12 migratory journey. ...
... 1. Interpolation and Simulation (section 2.2). First, we interpolate vertical profile time series of bird density and velocity field measured by weather radar data into continuous spatio-temporal maps following Nussbaumer et al. (2019). 2. Flow model (section 2.3) Then, using the interpolated data in a flow model allows us to estimate the number of birds entering, leaving, taking off from and landing in each grid cell at each time step. ...
... It is made (Dokter et al., 2011) and are 33 available on the ENRAM repository (ENRAM, 2020) at a 5 min x 200 m (0-5000m a.s.l.) resolution. Similar 34 to previous studies Nussbaumer et al., 2019), the vertical profiles were cleaned as 35 follows (supplementary material 1.2). First, we eliminated high-reflectivity contamination (e.g. ...
Full-text available
The movements of migratory birds constitute huge biomass flows that influence ecosystems and human economy, agriculture and health through the transport of energy, nutrients, seeds, and parasites. To better understand the influence on ecosystems and the corresponding services and disservices, we need to characterize and quantify the migratory movements at various spatial and temporal scales. Representing the flow of birds in the air as a fluid, we applied a flow model to interpolated maps of bird density and velocity retrieved from the European weather radar network, covering almost a full year. Using this model, we quantified how many birds take-off, fly, and land across Western Europe, (1) to track waves of bird migration between nights, (2) cumulate the number of bird on the ground and (3) quantify the seasonal flow into and out of the study area through several regional transects. Our results show that up to 188 million (M) birds take-off over a single night. Exemplarily, we tracked a migration wave in spring, in which birds crossed the study area in 4 days with nocturnal flights of approximately 300 km. Over the course of a season, we estimated that 494 million (M) birds entered through the southern transects and, at the same time, 251 M left in the northern transects, creating a surplus of 243 M birds within the study area. Similarly, in autumn, 544 M more birds departed than arrived: 314 M birds entered through the northern transects while 858 M left through the southern transects. Our study show-cases the potential of combining interdisciplinary data and methods to elucidate the dynamics of avian migration from nightly to seasonal and yearly time-scales and from regional to continental spatial scales.
... To date, insects are usually filtered using a threshold on the standard deviation of radial velocity and/or an additional threshold on absolute [7,11,[19][20][21][22][23][24][25][26]. Here, we combine these features within a Gaussian mixture model to estimate the proportions of birds and insects and their density and speed within a single measurement. ...
... The vertical profiles of reflectivity were manually cleaned using a MATLAB graphical user interface program, which consisted of removing vertical profiles contaminated by precipitation and eliminating altitude bins contaminated by ground scattering (detailed in Appendix A of Nussbaumer et al. [26]). The ground speed values were only kept where reflectivity values were available, thus removing erroneous speeds. ...
Full-text available
Recent and archived data from weather radar networks are extensively used for the quantification of continent-wide bird migration patterns. While the process of discriminating birds from weather signals is well established, insect contamination is still a problem. We present a simple method combining two Doppler radar products within a Gaussian mixture model to estimate the proportions of birds and insects within a single measurement volume, as well as the density and speed of birds and insects. This method can be applied to any existing archives of vertical bird profiles, such as the European Network for the Radar surveillance of Animal Movement repository, with no need to recalculate the huge amount of original polar volume data, which often are not available.
... Future research using these radars will provide more insight into the spatial variability of the vertical distribution, by providing more insight into the effects of elevation or habitat on the vertical distribution. This work paves the way toward developing methods that make more accurate assumptions on the local vertical profile compared to the nearest neighbor interpolation (taking the vertical profile of the closest radar) that is currently used; for example, by taking into account the full auto-correlation structure between profiles [44]. Combining weather radar with local bird radars is a powerful combination in which weather radar can provide the context for bird radars. ...
... These provide a powerful tool for experts to gain insight into a range of migratory phenomena. Future improvements are approaches that can account for regional variations in altitudinal distributions [44,45], as well as tools to exclude non-biological signals to ease the interpretation and work toward fully quantitative analysis. Lin et al. [46] already took important initial steps by segmenting out rain from the projected images using machine learning; alternatively, dual-polarization metrics can be used [e.g., 47]. ...
Full-text available
Weather radars provide detailed information on aerial movements of organisms. However, interpreting fine-scale radar imagery remains challenging because of changes in aerial sampling altitude with distance from the radar. Fine-scale radar imagery has primarily been used to assess mass exodus at sunset to study stopover habitat associations. Here, we present a method that enables a more intuitive integration of information across elevation scans projected in a two-dimensional spatial image of fine-scale radar reflectivity. We applied this method on nights of intense bird migration to demonstrate how the spatial distribution of migrants can be explored at finer spatial scales and across multiple radars during the higher flying en-route phase of migration. The resulting reflectivity maps enable explorative analysis of factors influencing their regional and fine-scale distribution. We illustrate the method’s application by generating time-series of composites of up to 20 radars, achieving a nearly complete spatial coverage of a large part of Northwest Europe. These visualizations are highly useful in interpreting regional-scale migration patterns and provide detailed information on bird movements in the landscape and aerial environment.
... Gauthreaux and Diehl [9] explore the efficacy of polarimetric weather radar observations for bioscatter classification and delineation. Nussbaumer et al. [10] and Kranstuaber et al. [11] both examine bird migration across Europe and are examples of burgeoning use of the OPERA weather radar network to estimate both density and distribution of migration, respectively. Finally, all of these advances in data processing allow a greatly expanded capacity to ask fundamental questions about the ecology of animal movement. ...
Full-text available
Aeroecology is an emerging discipline founded by Tom Kunz and colleagues in the early 2000s to address the challenges of studying animal flight in the lower atmosphere [...]
... To this end, citizen science data (e.g., offer great potential to study the timing of bird migration across Europe, and the continent-wide weather radar network can provide the rate of migration (Nussbaumer et al., 2019), whereas individual-based tracking still presents the best opportunity to link breeding, stopover and non-breeding sites at the individual and population levels. ...
Full-text available
Aim: Knowledge of broad-scale biogeographical patterns of animal migration is important for understanding ecological drivers of migratory behaviours. Here, we present a flyway-scale assessment of the spatial structure and seasonal dynamics of the Afro-Palaearctic bird migration system and explore how phenology of the environment guides long-distance migration. Location: Europe and Africa. Time period: 2009-2017. Major taxa studied: Birds. Methods: We compiled an individual-based dataset comprising 23 passerine and near-passerine species of 55 European breeding populations, in which a total of 564 individuals were tracked during migration between Europe and sub-Saharan Africa. In addition, we used remotely sensed primary productivity data (the normalized difference vegetation index) to estimate the timing of vegetation green-up in spring and senescence in autumn across Europe. First, we described how individual breeding and non-breeding sites and the migratory flyways link geographically. Second, we examined how the timing of migration along the two major Afro-Palaearctic flyways is tuned with vegetation phenology at the breeding sites. Results: We found the longitudes of individual breeding and non-breeding sites to be related in a strongly positive manner, whereas the latitudes of breeding and non-breeding sites were related negatively. In autumn, migration commenced ahead of vegetation senescence, and the timing of migration was 5-7 days earlier along the Western flyway compared with the Eastern flyway. In spring, the time of arrival at breeding sites was c. 1.5 days later for each degree northwards and 6-7 days later along the Eastern compared with the Western flyway, reflecting the later spring green-up at higher latitudes and more eastern longitudes. Main conclusions: Migration of the Afro-Palaearctic landbirds follows a longitudinally parallel leapfrog migration pattern, whereby migrants track vegetation green-up in spring but depart before vegetation senescence in autumn. The degree of continen-tality along migration routes and at the breeding sites of the birds influences the timing of migration on a broad scale.
... Erni et al. 2002;Van Belle et al. 2007). Recently, a model to forecast bird migration across the United States using a network of weather radars has been developed (Van Doren and Horton 2018), and a Europe-wide system is within reach Nussbaumer et al. 2019). These systems offer the unique possibility to monitor and forecast bird migration at the continental scale, but they are not adequate to provide migration forecasts over a complex terrain, such as the mountainous Swiss landscape, because of the inadequate coverage of low flight altitudes (H€ uppop et al. 2019). ...
Full-text available
Each year, billions of birds migrate across the continents by day and night through airspaces increasingly altered by human activity, resulting in the deaths of millions of birds every year through collisions with man‐made structures. To reduce these negative impacts on wildlife, forecasts of high migration intensities are needed to apply mitigation actions. While existing weather radar networks offer a unique possibility to monitor and forecast bird migration at large spatial scales, forecasts at the fine spatial scale within a complex terrain, such as the mountainous Swiss landscape, require a small‐scale network of ornithological radars. Before attempting to build such a network, it is crucial to first investigate the consistency of the migratory flow across space and time. In this study, we simultaneously operated three ornithological radar systems across the Swiss lowlands to assess the spatio‐temporal consistency of diurnal and nocturnal bird movements during the spring and autumn migration season. The relative temporal course of migration intensities was generally consistent between sites during peak migration, in particular for nocturnal movements in autumn, but absolute intensities differed greatly between sites. Outside peak migration, bird movement patterns were much less consistent and, unexpectedly, some presumably non‐migratory bird activity achieved intensities close to peak migration intensities, but without spatial correlations. Only nocturnal migration intensity in autumn could be predicted with consistently high accuracy, but including parameters of atmospheric conditions in the model improved predictability of diurnal movements considerably. Predictions for spring were less reliable, probably because we missed an important part of the migration season. Our results show that reliable forecasts of bird movements within a complex terrain call for a network of year‐round bird monitoring systems, whereas accurate information of atmospheric conditions can help to limit the number of measurement points. We assessed the spatio‐temporal consistency of diurnal and nocturnal bird movement patterns during the spring and autumn migration season across the Swiss lowlands by operating multiple ornithological radars. Our study showed that bird movement patterns were very consistent during peak migration in terms of relative intensities, but much less in terms of absolute intensities. Surprisingly, some presumably non‐migratory bird activity achieved intensities close to peak migration intensities, but without spatial correlations.
A high level of flight safety is a priority for all aviation organizations in the world. The negative impact of the natural environment on air operations is one of the most common factors causing the situations exerting a negative influence on flight safety (in short-aviation occurrence) in the Polish Army. The most common reason of the aviation occurrence connected with the environment area is the bird strike. The lack of a method for real-time monitoring and forecasting bird strike risk level is a significant gap in the proactive approach to flight safety within the Polish Air Force. The article presents a review of the methods for monitoring and forecasting the intensity of bird movements, which were used to build advisory systems used by air forces of the United States, Germany, the Netherlands, Belgium and Israel. The article consists of summary of most important properties of the methods and analysis in terms of their applicability in the Polish conditions. Particular attention was paid to the complicated network of routes of bird species occurring on Polish territory. Finally, after proving the incomplete usefulness of the above-mentioned methods in Polish conditions, the need to develop a more adequate method was justified.
The results of the analysis of factors influencing the emergence of dangerous situations in the air (in short, aviation incidents) in the Polish Army show a significant negative impact of the environment on the flight operations performed. The most common cause of an aviation incident in the area of the environment is the collision of the aircraft with birds. The lack of methods for the continuous monitoring and forecasting of the level of risk of collision between aircraft and birds makes a significant gap in a proactive approach to the safety of flights in the Air Force of the Republic of Poland. This paper presents an overview of the most important methods for detection and forecasting of bird flight intensity, which have been used for construction of systems aiming to prevent collisions with birds and which are employed by the air forces of the United States, the Netherlands, Belgium and Israel. An accurate analysis of the models and algorithms used in the selected methods shows contemporary trends in research on the negative impact of the environment on flight safety. These methods show incomplete usefulness in Poland’s conditions, which justifies the need to develop a more appropriate method.
Full-text available
Essentially all hydrogeological processes are strongly influenced by the subsurface spatial heterogeneity and the temporal variation of environmental conditions, hydraulic properties, and solute concentrations. This spatial and temporal variability needs to be considered when studying hydrogeological processes in order to employ adequate mechanistic models or perform upscaling. The scale at which a hydrogeological system should be characterized in terms of its spatial heterogeneity and temporal dynamics depends on the studied process and it is not always necessary to consider the full complexity. In this paper, we identify a series of hydrogeological processes for which an approach coupling the monitoring of spatial and temporal variability, including 4D imaging, is often necessary: (1) groundwater fluxes that control (2) solute transport, mixing and reaction processes, (3) vadose zone dynamics, and (4) surface-subsurface water interaction occurring at the interface between different subsurface compartments. We first identify the main challenges related to the coupling of spatial and temporal fluctuations for these processes. Then, we highlight some recent innovations that have led to significant breakthroughs in this domain. We finally discuss how spatial and temporal fluctuations affect our ability to accurately model them and predict their behavior. We thus advocate a more systematic characterization of the dynamic nature of subsurface processes, and the harmonization of open databases to store hydrogeological data sets in their four-dimensional components, for answering emerging scientific question and addressing key societal issues.
Full-text available
Understanding interactions between biota and the built environment is increasingly important as human modification of the landscape expands in extent and intensity. For migratory birds, collisions with lighted structures are a major cause of mortality, but the mechanisms behind these collisions are poorly understood. Using 40 years of collision records of passerine birds, we investigated the importance of species' behavioural ecologies in predicting rates of building collisions during nocturnal migration through Chicago, IL and Cleveland, OH, USA. We found that the use of nocturnal flight calls is an important predictor of collision risk in nocturnally migrating passerine birds. Species that produce flight calls during nocturnal migration tended to collide with buildings more than expected given their local abundance, whereas those that do not use such communication collided much less frequently. Our results suggest that a stronger attraction response to artificial light at night in species that produce flight calls may mediate these differences in collision rates. Nocturnal flight calls probably evolved to facilitate collective decision-making during navigation, but this same social behaviour may now exacerbate vulnerability to a widespread anthropogenic disturbance. Our results also suggest that social behaviour during migration may reflect poorly understood differences in navigational mechanisms across lineages of birds.
Full-text available
Quantitative radar studies are an important component of studying the movements of birds. Whether a bird, at a certain distance from the radar, is detected or not depends on its size. The volume monitored by the radar is therefore different for birds of different sizes. Consequently, an accurate quantification of bird movements recorded by small‐scale radar requires an accurate determination of the monitored volume for the objects in question, although this has tended to be ignored. Here, we demonstrate the importance of sensitivity settings for echo detection on the estimated movement intensities of birds of different sizes. The amount of energy reflected from a bird and detected by the radar receiver (echo power) depends not only on the bird's size and on the distance from the radar antenna, but also on the beam shape and the bird's position within this beam. We propose a method to estimate the size of a bird based on the wingbeat frequency, retrieved from the echo‐signal, independent of the absolute echo power. The estimated bird‐size allows calculation of size‐specific monitored volumes, allowing accurate quantification of movement intensities. We further investigate the importance of applying size‐specific monitored volumes to quantify avian movements instead of using echo counts. We also highlight the importance of accounting for size‐specific monitored volume of small scale radar systems, and the necessity of reporting technical information on radar parameters. Applying this framework will increase the quality and validity of quantitative radar monitoring. This article is protected by copyright. All rights reserved.
Full-text available
The aerosphere is utilized by billions of birds, moving for different reasons and from short to great distances spanning tens of thousands of kilometres. The aerosphere, however, is also utilized by aviation which leads to increasing conflicts in and around airfields as well as en‐route. Collisions between birds and aircraft cost billions of euros annually and, in some cases, result in the loss of human lives. Simultaneously, aviation has diverse negative impacts on wildlife. During avian migration, due to the sheer numbers of birds in the air, the risk of bird strikes becomes particularly acute for low‐flying aircraft, especially during military training flights. Over the last few decades, air forces across Europe and the Middle East have been developing solutions that integrate ecological research and aviation policy to reduce mutual negative interactions between birds and aircraft. In this paper we (1) provide a brief overview of the systems currently used in military aviation to monitor bird migration movements in the aerosphere, (2) provide a brief overview of the impact of bird strikes on military low‐level operations, and (3) estimate the effectiveness of migration monitoring systems in bird strike avoidance. We compare systems from the Netherlands, Belgium, Germany, Poland and Israel, which are all areas that Palearctic migrants cross twice a year in huge numbers. We show that the en‐route bird strikes have decreased considerably in countries where avoidance systems have been implemented, and that consequently bird strikes are on average 45% less frequent in countries with implemented avoidance systems in place. We conclude by showing the roles of operational weather radar networks, forecast models and international and interdisciplinary collaboration to create safer skies for aviation and birds. This article is protected by copyright. All rights reserved.
Full-text available
Avian migration is one of Earth's largest processes of biomass transport, involving billions of birds. We estimated continental biomass flows of nocturnal avian migrants across the contiguous United States using a network of 143 weather radars. We show that, relative to biomass leaving in autumn, proportionally more biomass returned in spring across the southern United States than across the northern United States. Neotropical migrants apparently achieved higher survival during the combined migration and non-breeding period, despite an average three- to fourfold longer migration distance, compared with a more northern assemblage of mostly temperate-wintering migrants. Additional mortality expected with longer migration distances was probably offset by high survival in the (sub)tropics. Nearctic-Neotropical migrants relying on a 'higher survivorship' life-history strategy may be particularly sensitive to variations in survival on the overwintering grounds, highlighting the need to identify and conserve important non-breeding habitats.
Full-text available
Nocturnal avian migration flyways remain an elusive concept, as we have largely lacked methods to map their full extent. We used the network of European weather radars to investigate nocturnal bird movements at the scale of the European flyway. We mapped the main migration directions and showed the intensity of movement across part of Europe by extracting biological information from 70 weather radar stations from northern Scandinavia to Portugal, during the autumn migration season of 2016. On average, over the 20 nights and all sites, 389 birds passed per 1 km transect per hour. The night with highest migration intensity showed an average of 1 621 birds per km per hour passing the radar stations, but there was considerable geographical and temporal variation in migration intensity. The highest intensity of migration was seen in central France. The overall migration directions showed strong southwest components. Migration dynamics were strongly related to synoptic wind conditions. A wind‐related mass migration event occurred immediately after a change in wind conditions, but quickly diminished even when supporting winds continued to prevail. This first continental‐scale study using the European network of weather radars demonstrates the wealth of information available and its potential for investigating large‐scale bird movements, with consequences for ecosystem function, nutrient transfer, human and livestock health, and civil and military aviation. This article is protected by copyright. All rights reserved.
Large networks of weather radars are comprehensive instruments for studying bird migration. For example, the US WSR‐88D network covers the entire continental US and has archived data since the 1990s. The data can quantify both broad and fine‐scale bird movements to address a range of migration ecology questions. However, the problem of automatically discriminating precipitation from biology has significantly limited the ability to conduct biological analyses with historical radar data. We develop MistNet, a deep convolutional neural network to discriminate precipitation from biology in radar scans. Unlike prior machine learning approaches, MistNet makes fine‐scaled predictions and can collect biological information from radar scans that also contain precipitation. MistNet is based on neural networks for images, and includes several architecture components tailored to the unique characteristics of radar data. To avoid a massive human labelling effort, we train MistNet using abundant noisy labels obtained from dual polarization radar data. In historical and contemporary WSR‐88D data, MistNet identifies at least 95.9% of all biomass with a false discovery rate of 1.3%. Dual polarization training data and our radar‐specific architecture components are effective. By retaining biomass that co‐occurs with precipitation in a single radar scan, MistNet retains 15% more biomass than traditional whole‐scan approaches to screening. MistNet is fully automated and can be applied to datasets of millions of radar scans to produce fine‐grained predictions that enable a range of applications, from continent‐scale mapping to local analysis of airspace usage. Radar ornithology is advancing rapidly and leading to significant discoveries about continent‐scale patterns of bird movements. General‐purpose and empirically validated methods to quantify biological signals in radar data are essential to the future development of this field. MistNet can enable large‐scale, long‐term, and reproducible measurements of whole migration systems. ⼀、⼤規模的氣象雷達網路是研究⿃類遷徙的全⽅位⼯具。US WSR‐88D 氣象雷達網路 覆蓋美國⼤陸,存擋從 1990 年代⾄今的雷達數據。這些數據可被⽤來量化⼤尺度到細尺 度上的⿃類活動,從⽽回答⼀系列⿃類遷徙的⽣態問題。然⽽,過去雷達數據分析仰賴⼤ 量的⼈⼒以區分雷達影像上的降⾬及⽣態活動。這個缺點侷限了利⽤歷史數據以分析⽣ 態活動的可能性。 ⼆、我們開發了 MISTNET,⼀個基於深度卷積神經網路的機器學習模型來分辨雷達數據 中的降⾬和⽣物訊號。不同於傳統的機器學習模型,MISTNET 可以做到細尺度的辨識, 並能從同時存在降⾬和⽣物訊號的雷達掃瞄中收集⽣物資訊。MISTNET 的設計基於影 像辨識的深度神經網路,並包含了針對雷達數據的特性所設計的架構。為了避免標記影 像所需的⼤量⼈⼒,我們使⽤從雙極化的雷達數據中獲得的帶有雜訊的影像標籤來訓練 MISTNET。 三、在歷史和近期的 WSR‐88D 雷達數據中,MISTNET 能辨識出⾄少 95.9% 以上的⽣ 物量,並且僅有 1.3% 假陽性錯誤率。透過保留和降⾬同時存在的⽣物訊號,MISTNET ⽐傳統過濾整張影像的⽅法多保留 15% 的⽣物量。MISTNET 實現了全⾃動化,能夠處 理多達百萬幅的雷達掃瞄數據集來產⽣空間上細尺度的辨識。這些特性成就 MISTNET 可廣泛⽤於各種應⽤,包含從⼤陸尺度到區域性的空域分析。 四、雷達⿃類學的發展迅速,並已經獲得了⼤陸尺度上⿃類活動的重⼤認識。通⽤於⼀般 ⽤途且可量化驗證的雷達⽣物訊號量化⽅法是這個領域未來發展的關鍵。MISTNET 可 ⽤於⼤規模且⾧時間的量測整體遷徙系統,且測量的結果是可以重複實現的。
Current climate models and observations indicate that atmospheric circulation is being affected by global climate change. To assess how these changes may affect nocturnally migrating bird populations, we need to determine how current patterns of wind assistance at migration altitudes will be enhanced or reduced under future atmospheric conditions. Here, we use information compiled from 143 weather surveillance radars stations within the contiguous United States to estimate the daily altitude, density, and direction of nocturnal migration during the spring and autumn. We intersected this information with wind projections to estimate how wind assistance is expected to change during this century at current migration altitudes. The prevailing westerlies at midlatitudes are projected to increase in strength during spring migration and decrease in strength to a lesser degree during autumn migration. Southerly winds will increase in strength across the continent during both spring and autumn migration, with the strongest gains occurring in the center of the continent. Wind assistance is projected to increase across the central (0.44 m/s; 10.1%) and eastern portions of the continent (0.32 m/s; 9.6%) during spring migration, and wind assistance is projected to decrease within the central (0.32 m/s; 19.3%) and eastern portions of the continent (0.17 m/s; 6.6%) during autumn migration. Thus, across a broad portion of the continent where migration intensity is greatest, the efficiency of nocturnal migration is projected to increase in the spring and decrease in the autumn, potentially affecting time and energy expenditures for many migratory bird species. These findings highlight the importance of placing climate change projections within a relevant ecological context informed through empirical observations, and the need to consider the possibility that climate change may generate both positive and negative implications for natural systems.
Billions of animals cross the globe each year during seasonal migrations, but efforts to monitor them are hampered by the unpredictability of their movements. We developed a bird migration forecast system at a continental scale by leveraging 23 years of spring observations to identify associations between atmospheric conditions and bird migration intensity. Our models explained up to 81% of variation in migration intensity across the United States at altitudes of 0 to 3000 meters, and performance remained high in forecasting events 1 to 7 days in advance (62 to 76% of variation was explained). Avian migratory movements across the United States likely exceed 500 million individuals per night during peak passage. Bird migration forecasts will reduce collisions with buildings, airplanes, and wind turbines; inform a variety of monitoring efforts; and engage the public.
1.Advances in information technology are increasing the use of radar as a tool to investigate and monitor bird migration movements. We set up a field campaign to compare and validate outputs from different radar systems. 2.Here we compare the pattern of nocturnal bird migration movements recorded by four different radar systems at a site in southern Sweden. Within the range of the weather radar (WR) Ängelholm, we operated a “BirdScan” (BS) dedicated bird radar, a standard marine radar (MR), and a tracking radar (TR). 3.The measures of nightly migration intensities, provided by three of the radars (WR, BS, MR), corresponded well with respect to the relative seasonal course of migration, while absolute migration intensity agreed reasonably only between WR and BS. Flight directions derived from WR, BS and TR corresponded very well, despite very different sample sizes. Estimated mean ground speeds differed among all four systems. The correspondence among systems was highest under clear sky conditions and at high altitudes. 4.Synthesis and applications. While different radar systems can provide useful information on nocturnal bird migration, they have distinct strengths and weaknesses, and all require supporting data to allow for species level inference. Weather radars continuously detect avian biomass flows across a wide altitude band, making them a useful tool for monitoring and predictive applications at regional to continental scales that do not rely on resolving individuals. BirdScan and marine radar's strengths are in local and low altitude applications, such as collision risks with man‐made structures and airport safety, although marine radars should not be trusted for absolute intensities of movement. In quantifying flight behaviour of individuals, tracking radars are the most informative. This article is protected by copyright. All rights reserved.
Bird collisions at wind turbines are perceived to be an important conservation issue. To determine mitigation actions such as temporary shutdown of wind turbines when bird movement intensities are high, knowledge of the relationship between the number of birds crossing an area and the number of collisions is essential. Our aim was to combine radar data on bird movement intensities with collision data from a systematic carcass search. We used a dedicated bird radar, located near a wind farm in a mountainous area, to continuously record bird movement intensities from February to mid-November 2015. In addition, we searched the ground below three wind turbines (Enercon E-82) for carcasses on 85 dates and considered three established correction factors to extrapolate the number of collisions. The extrapolated number of collisions was 20.7 birds/wind turbine (CI-95%: 14.3–29.6) for 8.5 months. Nocturnally migrating passerines, especially kinglets (Regulus sp.), represented 55% of the fatalities. 2.1% of the birds theoretically exposed to a collision (measured by radar at the height of the wind turbines) were effectively colliding. Collisions mainly occurred during migration and affected primarily nocturnal migrants. It was not possible to assign the fatalities doubtlessly to events with strong migration. Fresh-looking carcasses were found after nights with both strong and weak bird movement intensities, indicating fatalities are not restricted to mass movement events (onshore). Rather, it is likely that an important factor influencing collision risk is limited visibility due to weather conditions. Local and regional visibility should be considered in future studies and when fine-tuning shutdown systems for wind turbines.