Time-Frequency Composition of Mosquito Ight Tones Obtained Using Hilbert Spectral Analysis

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 9

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/267043979

Time-frequency composition of mosquito flight tones obtained using Hilbert


spectral analysis

Article  in  The Journal of the Acoustical Society of America · October 2014


DOI: 10.1121/1.4895689 · Source: PubMed

CITATIONS READS

4 265

4 authors, including:

Andy Aldersley A. R. Champneys


Imperial College London University of Bristol
6 PUBLICATIONS   112 CITATIONS    226 PUBLICATIONS   7,600 CITATIONS   

SEE PROFILE SEE PROFILE

Daniel Robert
University of Bristol
148 PUBLICATIONS   3,695 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

The role of acoustic signals in determining male mosquito mating success View project

Onomarchus ear View project

All content following this page was uploaded by Daniel Robert on 23 June 2016.

The user has requested enhancement of the downloaded file.


Time-frequency composition of mosquito flight tones obtained
using Hilbert spectral analysis
Andrew Aldersley
Bristol Centre for Complexity Sciences, University of Bristol, Queen’s Building, University Walk,
Bristol BS8 1TR, United Kingdom
Alan Champneysa) and Martin Homer
Department of Engineering Mathematics, University of Bristol, Merchant Venturers Building,
Woodland Road, Bristol BS8 1UB, United Kingdom
Daniel Robert
School of Biological Sciences, University of Bristol, Life Sciences Building, 24 Tyndall Avenue,
Bristol BS8 1TQ, United Kingdom

(Received 20 February 2014; revised 23 July 2014; accepted 25 August 2014)


Techniques for estimating temporal variation in the frequency content of acoustic tones based on
short-time fast Fourier transforms are fundamentally limited by an inherent time-frequency trade-
off. This paper presents an alternative methodology, based on Hilbert spectral analysis, which is
not affected by this weakness, and applies it to the accurate estimation of mosquito wing beat fre-
quencies. Mosquitoes are known to communicate with one another via the sounds generated by
their flapping wings. Active frequency modulation between pairs of mosquitoes is thought to take
place as a precursor to courtship. Studying the acoustically-based interactions of mosquitoes there-
fore relies on an accurate representation of flight frequency as a time-evolving property, yet con-
ventional Fourier spectrograms are unable to capture the rapid modulations in frequency that
mosquito flight tones exhibit. The algorithms introduced in this paper are able to automatically
detect and extract fully temporally resolved frequency information from audio recordings.
Application of the technique to experimental recordings of single tethered mosquitoes in flight
reveals corroboration with previous reported findings. The advantages of the method for animal
communication studies are discussed, with particular attention given to its potential utility for
studying pairwise mosquito interactions. VC 2014 Acoustical Society of America.

[http://dx.doi.org/10.1121/1.4895689]
PACS number(s): 43.80.Ka, 43.60.Hj, 43.80.Ev [ANP] Pages: 1982–1989

I. INTRODUCTION at least several species of mosquito (Gibson and Russell,


2006; Cator et al., 2009, 2010; Cator and Harrington, 2011;
As a prominent vector of numerous viruses and para-
Warren et al., 2009; Pennetier et al., 2010). Audio record-
sites, mosquitoes pose a major health risk to millions of indi-
ings of tethered pairs of mosquitoes have revealed the exis-
viduals across the globe. Disease is transmitted when the
tence of a dynamic interaction between the two. Males and
female mosquito bites her target to obtain a blood meal, a
females tend to reduce the difference between their flight fre-
protein supply vital to egg development. While the relation-
quencies prior to contact on the wing, a phenomenon that
ship between the mosquito reproductive cycle and their pro-
can occur at harmonic components beyond the fundamental
pensity to spread disease is well established, there remain
flight tone.
large gaps in our knowledge of the mating ecology of these
Despite significant growth in the mosquito bioacoustics
insects. Understanding mosquito breeding behaviors is a
literature, there remain aspects of their frequency-based
vital component of any targeted population control strategy
interactions that have not been consistently described or
proposed to mitigate the effects of mosquito-borne illnesses
quantified with sufficient detail. This can partly be attributed
on humans (Howell and Knols, 2009; Alphey et al., 2010).
to a lack of numerical and statistical analysis, an issue that
It has long been known that male mosquitoes in the
motivates the research described in this paper.
search for a mate are attracted toward the sounds produced
Information transfer between mosquitoes appears to
by the beating wings of nearby females (Johnston, 1855;
take place both in the frequency and time domains (Jackson
Roth, 1948; Wishart and Riordan, 1959; Belton, 1994;
and Robert, 2006). Consequently, a tool that offers an accu-
G€opfert et al., 1999; Charlwood et al., 2002). Recent
rate time-frequency representation of wing beat content is
ground-breaking behavioral experiments have highlighted
required to accurately study mosquito interactions. Previous
the key role played by audition in not only sexual identifica-
studies have generally extracted frequency information from
tion, but also the courtship rituals and selection practices of
audio recordings of mosquito flight through the use of a
Fourier transform. Breaking up the original input into
a)
Author to whom correspondence should be addressed. Electronic mail: smaller overlapping intervals in the time domain generates a
a.r.champneys@bristol.ac.uk short-time Fourier transform (STFT), and hence the

1982 J. Acoust. Soc. Am. 136 (4), October 2014 0001-4966/2014/136(4)/1982/8/$30.00 C 2014 Acoustical Society of America
V
frequency spectrum for local sections of the data. When acquisition device, for duration of 60 s following the onset of
implemented in a spectrogram (the squared modulus of the consistent flight.
STFT), this technique allows for the reconstruction of the Laboratory temperatures were held between 24  C and

time evolution of the signal’s frequency content. 28 C for the duration of recordings. Lighting and humidity
However, one significant limitation of the STFT is that conditions were also monitored and controlled. Mosquitoes
it requires a trade-off between frequency resolution (the were used at the beginning of their night period, running on
scale at which nearby frequencies can be distinguished) and a D16:N8 circadian cycle. Each recording was taken in a
time resolution (the interval over which frequency changes sound-proof booth to minimize noise interference from
can be seen). Using a smaller window to partition the signal external sources.
allows for better resolution in the time domain but results in
a poorer frequency resolution. Similarly, a wider frequency B. Data output
bandwidth necessarily reduces temporal accuracy.
Spectrogram representations of the time-frequency plane are One example of the raw waveform of flight acoustics
often “blurred” as a result, and rendered unsuitable for the recorded from a single mosquito, prior to any processing, is
detection and analysis of rapid frequency modulations in displayed in Fig. 1. Variation in the flight tone amplitude is
bioacoustic signals (DiCecco et al., 2013), including mos- low, indicative of consistent and stable flight.
quito flight tones.
Besides the STFT, there are several alternative techni- III. FREQUENCY ANALYSIS METHODS
ques able to characterize the evolution of frequency in time. Resolving the frequency content of mosquito flight
Hilbert spectral analysis (HSA, Huang et al., 1998), which tones requires the isolation of harmonic components fol-
yields the instantaneous frequency of a signal, is a particu- lowed by frequency extraction. The algorithm presented here
larly useful example of such a method. The application of is based fundamentally on the notion of instantaneous fre-
HSA is especially effective in the context of mosquito flight quency, a concept which provides excellent resolution in the
tones, due to its utility in non-stationary systems that display time-frequency plane.
a highly variable frequency (Boashash, 1992). In this paper
we present a quantitative method for the extraction of fre- A. Instantaneous frequency
quency content from mosquito wing beat recordings, based
on HSA. Key features of our approach include automated The instantaneous frequency, xi ðtÞ, of a signal, xðtÞ,
identification of fundamental flight frequencies and har- derives from its Hilbert transform, HðxðtÞÞ. Following
monic overtones, application of time-varying filters to Huang et al. (1998, 2009), the general principle of HSA is as
reduce noise interference, and an algorithm to detect flight follows. We first take the Hilbert transform
cessation. This methodology achieves a significant improve- ð
1 1 xðsÞ
ment over previous studies in both the time and frequency ð ð Þ Þ ð
H x t ¼y t ¼ PÞ ds; (1)
resolution of mosquito flight tones. p 1 t  s

where P indicates the Cauchy principal value of the integral.


II. DATA ACQUISITION
The Hilbert transform is known to be well defined provided
A. Procedure that the data is Lebesgue integrable, or more specifically that
the time series is a function of class Lp , with 1 < p < 1.
Experimental specifications were motivated by similar
Note that this mild requirement allows for the data to include
previous studies (Gibson and Russell, 2006; Cator et al.,
discontinuities, spikes, and random noise. This is a notably
2009; Warren et al., 2009; Pennetier et al., 2010). A colony
of the species Aedes aegypti was initiated using laboratory
eggs obtained from the London School of Hygiene and
Tropical Medicine. Individual adult mosquitoes were anes-
thetized and fixed to the rounded end of an insect pin on the
dorsal surface of the thorax (a procedure detailed in Cator
et al., 2009), such that the subject was able to beat its wings
and move its head unimpeded.
Mosquitoes were placed in an upright manner to mimic
as closely as possible the natural flight position. Flight tones
were recorded using an electret condenser microphone (FG-
23329-C05 microphone, Knowles Electronics, Itasca, IL)
placed directly underneath the individual at a distance of less
than 1.5 cm (as in Cator et al., 2009; Warren et al., 2009),
beyond the reach of physical contact. Flight was initiated by
gently blowing on the insect, or by stroking its legs. Data
FIG. 1. Amplitude profile of a single male mosquito’s flight. Inset depicts
were logged at a sample rate of 40 kHz with the LabVIEW cycle-by-cycle resolution of the waveform over 25 ms. The magnitude of
SignalExpress software package (National Instruments, the amplitude is roughly constant, with little modulation, indicative of stable
Austin, TX), using a National Instruments USB data flight.

J. Acoust. Soc. Am., Vol. 136, No. 4, October 2014 Aldersley et al.: Frequency analysis of mosquito flight tones 1983
useful trait when handling experimentally derived acoustic amplitude component in every case (Fig. 2). The prominence
signals displaying non-stationary behaviors, as with mos- of this harmonic in the flight tones that we have recorded
quito flight tones. Under this definition, HðxðtÞÞ yields the lends it well to application in the remainder of the algorithm.
imaginary part of the analytic pair, yðtÞ, of the real data, xðtÞ, Isolation of the fundamental and harmonic overtones
so that we have a unique analytic signal given by from the main body of the signal is facilitated by the applica-
tion of a bandpass filter, which is defined by a center-
zðtÞ ¼ xðtÞ þ jyðtÞ ¼ ai ðtÞejhi ðtÞ ; (2) frequency and bandwidth. We take an automated approach
to the specification and design of these filter parameters (see
where ai ðtÞ and hi ðtÞ describe, respectively, the instantane-
pffiffiffiffiffiffiffi Fig. 3). The first step is to approximate the distribution of
ous amplitude and phase associated with xðtÞ, and j ¼ 1. the Fourier spectrum about the chosen frequency component.
The instantaneous frequency is then defined as the time de- The spectrum is normalized by dividing through by its stand-
rivative of hi ðtÞ; ard deviation (this improves function fitting). We then fit ei-
ther a skew Gaussian or a skew Laplace distribution to the
dhi
x i ðt Þ ¼ : (3) data; the choice of function is made by considering the kur-
dt tosis of the spectrum about the given frequency peak. We
Note that because hi is a function of time, so too is xi . find that data with a kurtosis  10 are best approximated by
the skew Gaussian, whereas the skew Laplacian consistently
fits data with a kurtosis >10.
B. Isolating harmonic frequencies
The fitting procedure and subsequent filter parameter
The notion of an instantaneous frequency is only mean- estimation is equivalent whichever distribution is used. The
ingful when applied to mono-component functions scale and location parameters of the fitting function are opti-
(Boashash, 1992), that is, data with a narrow bandwidth mized to the data using a least squares approach, recursively
(Huang et al., 1998). For multi-component signals, such as tuning values to minimize error. The peak of the distribution
the acoustic tones generated by mosquito flight, we must identifies the center point of the transfer function for the fil-
therefore filter each of the harmonic frequencies prior to ter. The bandwidth (i.e., the upper and lower cutoff frequen-
HSA. We have developed an automated procedure based on cies) is obtained using the frequencies at full-width, fraction
peak-frequency estimation that allows us to identify and iso- p% maximum of the fitted distribution (we typically choose
late the various harmonic frequency components contained p ¼ 2% for a Gaussian fit, and p ¼ 5% for a Laplacian).
within the sound profile of mosquito flight. The approach con- This process is outlined graphically in Fig. 3. The resultant
sists of several stages, described below. linear filter is subsequently applied to the audio recording.
Given a raw acoustic recording, an initial estimation of The temporal dynamics of mosquito flight frequencies
its frequency content is performed using Discrete Fourier are inherently non-stationary. With this in mind, an addi-
analysis. The signal is high-pass filtered (using a 4 order tional layer of sophistication must be introduced to our filter
Butterworth filter) to remove low-frequency components process to ensure that it robustly captures wing beat fre-
(300 Hz), which can cause noise interference, prior to quency information. In particular, a simple static filter
Fourier transformation. From the resultant Fourier spectrum applied to an entire recording can lead to a sub-optimal fre-
we identify the peak corresponding to the mosquito’s funda- quency response and, for instance, an increase in the noisi-
mental wing beat frequency, which we find to be the largest ness of the resultant instantaneous frequency functions (see
Fig. 4). One way to address this problem is to instead

FIG. 2. Fourier spectrum of a lone male mosquito’s flight tone, depicting


frequencies up to 6.5 kHz. The fundamental wing beat frequency is the larg-
est peak, located here at approximately 600 Hz. Harmonic overtones are FIG. 3. An example of parameter estimation for the time-varying filter. In this
clearly resolved above the background noise threshold up to the 9th compo- instance, a Laplace distribution is fitted to the Fourier spectrum around the
nent, with each appearing progressively weaker in amplitude. These data fundamental flight frequency. The filter is defined by a center frequency xmax
indicate that each harmonic peak is an integer multiple of the fundamental and bandwidth xD . xmax is denoted by the peak of the Laplacian, while xD is
frequency. given by the frequencies at full-width, 5% maximum of the fitted curve.

1984 J. Acoust. Soc. Am., Vol. 136, No. 4, October 2014 Aldersley et al.: Frequency analysis of mosquito flight tones
comprised of the harmonic frequencies up to and including
the male second and female third overtones; convergent
behaviors are observed to take place at or below these fre-
quencies (Cator et al., 2009).

C. Extracting instantaneous frequency


After filtering the original data, we obtain the analytic
signal via the Hilbert transform. Computation of xi ðtÞ fol-
lows from the subsequently derived instantaneous phase
function. Finally, we smooth the output using a Wiener filter
with a window of size 200 data points; this removes low-
amplitude high-frequency instabilities generated by the
time-varying filter while preserving the temporal qualities of
the instantaneous frequency signal. The instantaneous fre-
FIG. 4. Applying a static filter to a non-stationary signal can lead to instabil-
ities in the Hilbert transform, resulting in a noisy response of xi ðtÞ. Using a quency data obtained for each mosquito provide a rich pool
time-varying filter we are able to significantly reduce this noise. of information that is unbounded by the trade-off between
Instantaneous frequency is plotted as a time-evolving function under two time and frequency resolution necessarily imposed by
different filter regimes. A static linear Butterworth bandpass filter, with pa-
rameters estimated in situ, has been applied to the audio signal, and the in-
Fourier-based approaches. Our determination of xi ðtÞ is con-
stantaneous frequency, extracted via HSA, is plotted in gray. This is tinuous in frequency space and limited temporally only by
compared with the output from the time-varying filter approach outlined in the sampling rate at which the signals were recorded. Such
Sec. III (black trace). high-resolution data sets have previously been inaccessible
and largely unused in mosquito bioacoustics (exceptions are
Jackson and Robert, 2006; Jackson et al., 2009).
implement a time-varying filter, enabling us to enforce a
locally optimal bandwidth. In this instance, timescale selec-
D. Detecting flight cessation
tion is of primary importance.
In our procedure, we iterate over a given recording at A frequent observation during experimentation was that
increments of time duration sw , and extract local filter pa- the tethered mosquito would cease flight for a duration last-
rameters for each segment using the process outlined above. ing anything up to several seconds. While such periods
We generally choose sw ¼ 1 s, a condition that satisfies two formed a relatively minor part of our total recorded time,
requirements for our data: that the frequency content of the there was nevertheless a strong motivation to develop an
subset is well preserved, and that frequency over the subset automated method for the removal of such dormant sections
is unimodally distributed with one dominant component. of the data. The main reason for this was to preserve as
Large frequency modulations in mosquito flight tones tend much useable data from each recording as possible; Hilbert
to occur slowly, especially when the insects are flying alone transformation of background noise contained in a recording
(Arthur et al., 2014). (i.e., when there is no mosquito flight) can lead to distortions
A time-varying filter is specified using a set of one- in future analysis.
dimensional vectors to prescribe the center frequency and Flight cessation is typified by a rapid—and sometimes
bandwidth, respectively, at each corresponding time instant erratic—decrease in wing beat frequency and amplitude,
in the recorded signal. Thus for each increment sw ¼ 1 s, while the first moments of flight, often referred to as the
there are 4  104 filter points, the sample rate of our data “startle” period (Gibson and Russell, 2006; Cator et al.,
being 40 kHz. Linear interpolation of the filter parameters is 2010), constitute high frequency and amplitude. Therefore,
performed at each boundary sw to ensure a smooth parameter to detect when a mosquito has stopped flying, we use the am-
transition between time steps, and to reduce the occurrence of plitude of the waveform of the recorded flight.
instabilities in the frequency response of the filter. The time- Similar to the process for estimating filter parameters,
varying filter can be repeatedly applied n times to achieve the the audio signal from a given experimental trial is bandpass
desired order (we choose n ¼ 3 as a suitable compromise filtered to remove frequencies outside the expected range of
between computational time and noisiness of output). the tethered male/female fundamental flight tone (typically
Our method thus far has primarily focused on isolation those below 300 Hz and above 1100 Hz for A. aegypti) [Fig.
of the fundamental flight frequency. The process is easily 5(a)]. We then twice smooth the modulus of the resultant
extended, however, to yield higher harmonic components. signal using a Gaussian-weighted moving-average (effec-
Analysis of wing beat frequencies indicates that harmonic tively applying a low-pass finite impulse response filter)
overtones are located at integer multiples of the fundamental [Fig. 5(b)]. Taking the gradient of this smoothed function
wing beat tone, even during frequency modulation (Arthur allows us to identify points of rapid amplitude decline or
et al., 2014). This allows us to approximate and filter the sig- increase using a peak-finding algorithm [Fig. 5(c)]. Iterating
nal in the general region of any given harmonic, based on over maxima pairwise, we extract those punctuated by peri-
analysis of the fundamental tone. Once the mono-component ods with a median smoothed magnitude less than Ac . A value
function is resolved, we apply the time-varying filter routine. Ac ¼ 0:01 V consistently indicated flight cessation through-
For A. aegypti mosquitoes, a minimal useful dataset may be out our data. This mechanism enabled us to remove silent

J. Acoust. Soc. Am., Vol. 136, No. 4, October 2014 Aldersley et al.: Frequency analysis of mosquito flight tones 1985
FIG. 6. Comparing a spectrogram-based frequency representation (gray) of
an individual male mosquito’s flight tone to that of the instantaneous fre-
quency (black) demonstrates superior resolution in the time and frequency
domains for the latter approach. Rapid frequency modulations are captured
in excellent detail, providing a much clearer visualization of local mosquito
wing beat frequency behaviors. Note that the time-axis is extended either
side of the instantaneous frequency data merely to enable adequate presenta-
tion and comparison to the spectrogram image.

sample of which are shown in Fig. 6—were made to confirm


accuracy of the resultant time-frequency representation.
During recordings some mosquitoes were observed to
adopt an erratic flight position, possibly in an attempt to free
themselves from the tether. This behavior produced a “saw-
tooth” pattern in the wing beat frequency trace, as noted by
Warren et al. (2009). Data sets in which this trait was
observed were pruned such that only periods of consistent,
“stable” flight, with a minimum duration of 20 s, were
included in the analysis.

B. Properties of mosquito flight


Our approach to frequency extraction yields a one-
dimensional time series—xi ðtÞ—for each harmonic compo-
nent with a high temporal resolution, limited only by the
FIG. 5. Automated detection and removal of intra-recording flight cessation. sample rate of the original audio recordings. These data lend
(a) Raw (gray) and low-pass filtered (black) waveform. (b) Modulus (gray)
and twice-smoothed modulus (black) of the filtered signal. (c) Identification
themselves well to a thorough analysis of the properties of
of silent components, based on peak detection in the first-time derivative of male and female flight tone characteristics, an example of
the smoothed function in (b). (d) Removal of the identified silent elements which we now present.
from the xi ðtÞ trace. As has been documented for A. aegypti (Clements, 1999;
G€opfert et al., 1999; Cator et al., 2009; Arthur et al., 2014),
components of the time-frequency representations for each we observed tethered males to fly at a higher fundamental fre-
mosquito’s flight [Fig. 5(d)]. It avoids excessive pruning of quency than females. Average flight frequency ranged from
the audio signal and preserves the maximum amount of use- 492.1 to 880.3 Hz across the different males recorded, with an
able data. overall mean and standard deviation of 691.1 6 90.4 Hz. For
females, the average individual flight frequency was in the
interval 415.9 to 532.6 Hz, with a group-wide mean and stand-
IV. RESULTS ard deviation of 480.6 6 32.5 Hz. As in Arthur et al. (2014),
the variance in wing beat frequency was much lower for indi-
A. Frequency extraction of mosquito flight tones
vidual mosquitoes than across the totality of data collected.
We recorded individual flight tones of 27 male and 19 For males, the mean intra-recording standard deviation was
female A. aegypti. Frequency extraction using the combined 11.2 Hz, while for females it was 8.9 Hz. The smaller spread in
time-variant filter and HSA approaches presented here was lone female flight tones is a result of their naturally lower wing
carried out for each insect, up to the second and third har- beat frequency.
monic frequencies for males and females, respectively. The non-stationary and time-variant properties of mos-
Visual comparisons with spectrogram-based analyses—a quito flight tones can cause the usefulness of applied

1986 J. Acoust. Soc. Am., Vol. 136, No. 4, October 2014 Aldersley et al.: Frequency analysis of mosquito flight tones
“global” statistics to be somewhat diminished. To illustrate
any potential differences between male and female flight
characteristics it may be more informative to partition the
data into smaller, approximately-stationary subsets. Using
the fundamental flight component, we determine the array
( )
j xi  l  j ðsÞ
yi ¼ (4)
 j ðsÞ
l
i¼1;…;n

for the mosquito j, where xi represents the instantaneous fre-


quency at a given time instant, and l  j ðsÞ is the piecewise
moving mean taken at time intervals of s, evaluated over n
consecutive data points. We select s ¼ 1 second, giving
n ¼ 4  104 .
FIG. 8. Cumulative distribution functions (CDFs) for the aggregated male
For each mosquito, yji represents the fractional deviation and female wing beat frequency deviations presented in Fig. 7. Male mos-
from the local mean flight tone, standardized to account for quitoes exhibit a broader range of flight tone deviations, as evidenced by the
differences between individuals and the non-stationary fea- shallower slope of the CDF, when compared to females.
tures of wing beat frequency. It enables us to aggregate flight
information from each recording across the male and female
their local mean flight frequency than their female counter-
groups. Gender specific “meta” time series are generated
parts (Fig. 8).
separately, by combining yji for j ¼ 1; …; 15 mosquitoes,
selected at random from within our experimental database.
Figure 7 shows the probability density functions (PDFs) of V. DISCUSSION
the aggregate yji for males and females in solo flight. The dis-
We have developed a novel method for extracting flight
tributions are well approximated by Gaussian functions.
frequencies from acoustic recordings of tethered mosquitoes.
While broadly similar, there is a notable difference between
Our technique involves several steps, including frequency
the sexes: male mosquitoes generally deviate more from
estimation using Fourier transformation, the application of a
time-varying bandpass filter to isolate harmonic overtones,
and computation of instantaneous frequency content utilizing
the concepts of HSA. We also report on the design of an
algorithm able to detect and remove sections of data in
which mosquitoes cease to fly, which can be problematic
from an instantaneous frequency perspective.
The signal processing tools described are fully auto-
mated and have several advantages over Fourier-based meth-
ods that have typically been used in the study of mosquito
interactions in the frequency domain (Gibson and Russell,
2006; Cator et al., 2009; Warren et al., 2009; Pennetier
et al., 2010). The primary benefit is the absence of a trade-
off between temporal and frequency resolution, which ulti-
mately leads to a higher quantity and quality of wing beat
frequency data. Our technique offers significant improve-
ments on both these of fronts (Fig. 6), and, crucially, is able
to capture the rapid frequency modulations during mosquito
flight that are unclear in previous visualizations.
Spectrogram representations of frequency yield a data
matrix of frequency “power” as a time-evolving function.
Obtaining an accurate estimation of the largest amplitude
frequency components within the matrix at any given time
(the fundamental frequency and harmonic overtones of a
mosquito’s flight) requires the use of feature extraction tech-
niques, which may be limited by resolution in the time and/
or frequency domain. The procedure presented here, how-
ever, returns a readily available set of one-dimensional time
series (one for each harmonic component), negating the need
FIG. 7. PDFs showing the deviation from the mean fundamental flight tone
for a random selection (N ¼ 15) of single tethered (a) male and (b) female
for further manipulation of the time-frequency plane. These
mosquitoes, illustrated using gray bars. Black lines indicate subsequently fit- data facilitate subsequent analyses that go beyond simple
ted Gaussian distributions. time series plots (for example, see Figs. 7 and 8) and allow

J. Acoust. Soc. Am., Vol. 136, No. 4, October 2014 Aldersley et al.: Frequency analysis of mosquito flight tones 1987
us to explore further the nature of frequency as a property of
flight.
We recorded the flight tones of tethered lone male and
female A. aegypti mosquitoes for a duration up to 60 s. Male
and female wing beat frequencies obtained via HSA are stat-
istically consistent with those reported elsewhere for this
species, both in terms of global mean flight frequency and its
spread across individuals (G€opfert et al., 1999; Cator et al.,
2009; Arthur et al., 2014), indicating that our method does
indeed accurately capture the true frequency content of the
mosquito flight trace. Furthermore, by partitioning the data
into smaller short-time subsets, we have been able to account
for the inherent non-stationarity of mosquito flight. Our find-
ings reveal that wing beat frequencies are normally distrib-
uted about a local mean value. Male mosquitoes, however, FIG. 9. (Color online) When positioned close together (around 2 cm apart),
tend to deviate more from this mean than their female male and female Aedes aegypti display convergence of their flight tones.
This behavior commonly takes place at the male 2 and female 3 harmonic
conspecifics.
frequency components. When viewed using a spectrogram, resolution is
The significance of this qualitative difference between coarse and it is difficult to accurately prescribe signal ownership. Using the
male and female flight tone characteristics is unclear, yet instantaneous frequency representation of flight tone, resolution in the time-
analyses of pairwise mosquito interactions point to wing beat frequency plane is significantly enhanced. Recording mosquitoes on sepa-
rate channels also allows us to assign the male (M) and female (F) flight
frequency as a critical factor in species and gender identifica- tones with precision. As in Fig. 6, the time-axis displays both the spectro-
tion (Gibson and Russell, 2006; Cator et al., 2009; Warren gram alone (0–30 s), and the spectrogram overlaid with the instantaneous
et al., 2009; Pennetier et al., 2010), as well as in sexual selec- frequencies of each mosquito (30–60 s). This allows thorough inspection
tion (Cator et al., 2010; Cator and Harrington, 2011). Despite and clear comparison between the two methods.
this, there is a lack of quantitative investigation in this field,
and consequently the ubiquity and accuracy of the described assigned to particular individuals using spectrogram analy-
phenomena remain uncertain. For instance, during harmonic sis. Our method is able to unambiguously ascribe a given
convergence (Cator et al., 2009) between males and females, xi ðtÞ function to its producer (Fig. 9).
how close do their flight tones have to be to result in a suc- The study of acoustic communication between mosqui-
cessful duet? And what information do the more detailed and toes, and its broader role in reproductive processes, is central
dynamic properties of wing beat signals, such as high- to our understanding of their behavioral ecology. It is hoped
frequency tonal modulations and local variance, convey to that the analysis presented here will stimulate further investi-
other individuals? In order to answer these questions, and gations into the nature of frequency-based interactions
thus more comprehensively investigate interactions between between these insects.
mosquitoes, a detailed realization of frequency as a time- The characterization of nonlinear and non-stationary
dependent property is required. bioacoustic signals using HSA has arguably been underutil-
Arthur et al. (2014) recently presented two tools to sepa- ized in the literature. Exceptions include the study of whale
rate simultaneous flight tones during convergence behavior. clicks (Adam, 2006) and bat echolocation calls (DiCecco
The first of these is based on the phase-derivative of spectral et al., 2013). One possible reason for this under-utilization is
maxima in the spectrogram, while the second uses polyspec- that one has to first isolate a narrow band around the fre-
tra to enhance time-frequency resolution. Despite offering quency of interest in the signal before applying the Hilbert
greater accuracy than the coarsely represented spectrogram, transform. The central frequency and width of the band is
these approaches are still inherently limited by similar fac- likely to vary from species to species and between individu-
tors. By taking an instantaneous frequency approach to the als of the same species. Thus, unlike the STFT there has,
extraction of mosquito flight tone data, we have demon- until now, been no black-box method that can be easily
strated a far superior clarity in time and frequency compared applied to different biological systems. In this work, we
to any method used to date in the mosquito bioacoustics lit- have shown how to construct a highly optimized tool for the
erature. As illustrated in Fig. 9, convergence of flight tones extraction of frequency content contained within mosquito
now has the potential to be studied in much greater quantita- flight tones. In principle this same methodology can be
tive depth, owing to the quality of data available. adapted to work for many other bioacoustic communication
Experimentally, resolution of nearby harmonic frequencies studies involving non-stationary harmonic functions.
from different mosquitoes can be achieved by recording
each individual on a separate digital channel. This is in con- ACKNOWLEDGMENTS
trast to previous research, which has tended to log the duet This work was funded by the Engineering and Physical
on a single microphone positioned equidistant from each Sciences Research Council (EPSRC), Grant No. EP/
mosquito (Gibson and Russell, 2006; Cator et al., 2009; E501214/1. D.R. is supported by the Royal Society London.
Warren et al., 2009; Pennetier et al., 2010). In these studies
it is somewhat unclear how overlapping signals from unique Adam, O. (2006). “Advantages of the Hilbert Huang transform for marine
sources (observed during harmonic convergence) were mammals signals analysis,” J. Acoust. Soc. Am. 120(5), 2965–2973.

1988 J. Acoust. Soc. Am., Vol. 136, No. 4, October 2014 Aldersley et al.: Frequency analysis of mosquito flight tones
Alphey, L., Benedict, M., Bellini, R., Clark, G. G., Dame, D. A., Service, M. W., G€opfert, M. C., Briegel, H., and Robert, D. (1999). “Mosquito hearing:
and Dobson, S. L. (2010). “Sterile-insect methods for control of mosquito- Sound-induced antennal vibrations in male and female Aedes aegypti,”
borne diseases: An analysis,” Vector-borne Zoonot. 10(3), 295–311. J. Exp. Biol. 202 (part 20), 2727–2738.
Arthur, B. J., Emr, K. S., Wyttenbach, R. A., and Hoy, R. R. (2014). Howell, P., and Knols, B. G. J. (2009). “Male mating behaviour,” Malaria J.
“Mosquito (Aedes aegypti) flight tones: Frequency, harmonicity, spherical 8(Suppl. 2), S8.
spreading, and phase relationships,” J. Acoust. Soc. Am. 135(2), 933–941. Huang, N. E., Shen, Z., Long, S. R., Wu, M. C., Shih, H. H., Zheng, Q.,
Belton, P. (1994). “Attraction of male mosquitoes to sound,” J. Am. Mosq. Yen, N.-C., Tung, C. C., and Liu, H. H. (1998). “The empirical mode
Control Assoc. 10(2), 297–301. decomposition and the Hilbert spectrum for nonlinear and non-stationary
Boashash, B. (1992). “Estimating and interpreting the instantaneous fre- time series analysis,” Proc. R. Soc. London, Ser. A 454(1971), 903–995.
quency of a signal. I. Fundamentals,” Proc. IEEE 80(4), 520–538. Huang, N. E., Wu, Z., Long, S. R., Arnold, K. C., Chen, X., and Blank, K.
Cator, L. J., Arthur, B. J., Harrington, L. C., and Hoy, R. R. (2009). (2009). “On instantaneous frequency,” Adv. Adapt. Data Anal. 1(2),
“Harmonic convergence in the love songs of the dengue vector mosquito,” 177–229.
Science 323(5917), 1077–1079. Jackson, J. C., and Robert, D. (2006). “Nonlinear auditory mechanism
Cator, L. J., and Harrington, L. C. (2011). “The harmonic convergence of enhances female sounds for male mosquitoes,” Proc. Natl. Acad. Sci.
fathers predicts the mating success of sons in Aedes aegypti,” Anim. U.S.A. 103(45), 16734–16739.
Behav. 82, 627–633. Jackson, J. C., Windmill, J. F. C., Pook, V. G., and Robert, D. (2009).
Cator, L. J., Ng’Habi, K. R., Hoy, R. R., and Harrington, L. C. (2010). “Synchrony through twice-frequency forcing for sensitive and selective au-
“Sizing up a mate: Variation in production and response to acoustic sig- ditory processing,” Proc. Natl. Acad. Sci. U.S.A. 106(25), 10177–10182.
nals in Anopheles gambiae,” Behav. Ecol. 21(5), 1033–1039. Johnston, C. (1855). “Original communications: Auditory apparatus of the
Charlwood, J. D., Pinto, J., Sousa, C. A., Madsen, H., Ferreira, C., and Do Culex mosquito,” Q. J. Microsc. Sci. 1(10), 97–102.
Rosario, V. E. (2002). “The swarming and mating behaviour of Anopheles Pennetier, C., Warren, B., Dabire, K. R., Russell, I. J., and Gibson, G.
gambiae s.s. (Diptera: Culicidae) from S~ao Tome Island,” J. Vector Ecol. (2010). “ ‘Singing on the wing’ as a mechanism for species recognition in
27(2), 178–183. the malarial mosquito Anopheles gambiae,” Curr. Biol. 20(2), 131–136.
Clements, A. N. (1999). “Flight performance in the laboratory,” in The Roth, L. M. (1948). “A study of mosquito behavior. An experimental labora-
Biology of Mosquitoes, Volume 2: Sensory Reception and Behaviour tory study of the sexual behavior of Aedes aegypti (Linnaeus),” Am. Midl.
(Chapman & Hall, London), Chap. 33, pp. 292–298. Nat. 40(2), 265–352.
DiCecco, J., Gaudette, J. E., and Simmons, J. A. (2013). “Multi-component Warren, B., Gibson, G., and Russell, I. R. (2009). “Sex recognition through
separation and analysis of bat echolocation calls,” J. Acoust. Soc. Am. midflight mating duets in Culex mosquitoes is mediated by acoustic dis-
133(1), 538–546. tortion,” Curr. Biol. 19(9), 485–491.
Gibson, G., and Russell, I. R. (2006). “Flying in tune: Sexual recognition in Wishart, G., and Riordan, C. F. (1959). “Flight responses to various sounds
mosquitoes,” Curr. Biol. 16(13), 1311–1316. by adult males of Aedes aegypti,” Can. Entomol. 91(3), 181–191.

J. Acoust. Soc. Am., Vol. 136, No. 4, October 2014 Aldersley et al.: Frequency analysis of mosquito flight tones 1989

View publication stats

You might also like