Infrared light sensors permit rapid recording of wingbeat frequency and bioacoustic species identification of mosquitoes

Recognition and classification of mosquitoes is a critical component of vector-borne disease management. Vector surveillance, based on wingbeat frequency and other parameters, is becoming increasingly important in the development of automated identification systems, but inconsistent data quality and results frequently emerge from different techniques and data processing methods which have not been standardized on wingbeat collection of numerous species. We developed a simple method to detect and record mosquito wingbeat by multi-dimensional optical sensors and collected 21,825 wingbeat files from 29 North American mosquito species. In pairwise comparisons, wingbeat frequency of twenty six species overlapped with at least one other species. No significant differences were observed in wingbeat frequencies between and within individuals of Culex quinquefasciatus over time. This work demonstrates the potential utility of quantifying mosquito wingbeat frequency by infrared light sensors as a component of an automated mosquito identification system. Due to species overlap, wingbeat frequency will need to integrate with other parameters to accurately delineate species in support of efficient mosquito surveillance, an important component of disease intervention.

Mosquitoes are vectors of causative agents for numerous diseases, including malaria, filariasis, dengue, Zika, chikungunya, and encephalitis, ultimately resulting in more than one million deaths annually 1 and enormous economic losses through the costs of vaccinations, vector controls, and trade embargoes 2 . The recent Zika virus outbreak in Latin America cost approximately USD 18 billion from 2015 to 2017 3 . A number of control strategies have been developed and deployed to combat vector mosquitoes 4 , however, exotic mosquito-borne pathogens continue to be introduced into new areas and the frequency of epidemics is effectively increasing. Mosquito population data are crucial for assessing the local risk of mosquito-borne disease 5 .
Mosquito surveillance is the identification and enumeration of mosquitoes to determine the species composition, abundance, survival, and presence of disease agents prevalent in a given area and time to justify and implement timely data for risk estimate 6 . Important components of efficient and effective mosquito surveillance for making a scientifically-based decision (e.g., mosquito control, infection, insecticide resistance status) rely on consistent specimen collection and precise vector identification by trained staff. However, insufficient and unreliable surveillance data result from labor-intensive/time-consuming tasks, time lags, incorrect classification, and spatial constraints (e.g., remote areas), which undermine surveillance efforts and the corresponding capacity to detect, anticipate, and respond to vector-borne disease 7 . Therefore, advances in automated mosquito capture and reliable identification could provide critical tools effectively to monitor mosquito populations in real-time; a public health necessity in a state that frequently experiences outbreaks of medically important mosquito-vectored pathogens.
Recognition and classification of animals that use flapping flight 8 , based on wingbeat frequency and other parameters, have become increasingly important in the automated identification of these groups 9-11 using radar and mathematical models or artificial intelligence (AI) [12][13][14] . Similarly, radar can be used to differentiate flying mammals (bats) into groups (small, medium, or large) according to wingbeat frequency 15 . Much work has been done for revealing aerodynamic properties and kinematic focusing on insect wingbeats of relatively large size such as locusts, hawkmoths, and dragonflies 16,17 . However, mosquitoes display unusual wing kinematics in comparison with other insect groups 18  www.nature.com/scientificreports/ through an atypically small stroke amplitude arc, resulting in extremely high wing loading (body mass/wing surface area), low wingbeat amplitude of approximately 40° (150° in fruit fly), and high wingbeat frequency of approximately up to 1000 beats per second 19,20 . The powered and aerodynamic flight strategy in the mosquito facilitates is used for all general activities (e.g., dispersal, foraging, oviposition, evading predators), as well as acoustic communication 21 . The exceptional feature of mosquito wingbeat has frequently been suggested as a physical signature for species-level identification of mosquitoes [22][23][24] . To date, however, none of the automated mosquito identification systems using a wingbeat has been ready for real-world deployment. Numerous factors have limited progress toward an automated mosquito identification system based on wingbeat frequency. Importantly, the vast majority of experimental methods to collect mosquito wingbeats have used an acoustic microphone [25][26][27][28] . There are some drawbacks to this method. For example, following the inverse squared law, the distance between a mosquito and the acoustic microphone is negatively correlated with a sound's intensity which is a limiting factor for wingbeat data quality. In the case of using a condenser microphone (extremely sensitive), air current and ambient noise are overamplified 23 . Particularly, recording wingbeat in field conditions is challenging due to background "noise" from natural and man-made sources. Also, the difficulty of obtaining and maintaining a variety of mosquito species to collect representative wingbeat frequencies results in poor sample size and low diversity 23 . Consequently, inadequate sampling results in insufficient power to detect a wingbeat distinction for detailed analysis 29 and even causes contradictory findings (i.e., extremum). Inconsistent data quality and results frequently emerge from different techniques and data processing methods which have not been standardized by comparison to wingbeat collections of numerous species.
In this study, we describe a simple method to detect and record mosquito wingbeat by multi-dimensional optical sensors capturing a reflection of the transient waveform, which consequently avoids the issue of background acoustic noise. Secondly, we describe a wave file library created from the collection of twenty nine mosquito species of vector and nuisance significance that allowed us to generate a robust dataset from a more diverse suite of mosquito species than that used in previous research efforts. Finally, we include variations in the frequency of wingbeats between and within individuals of a single species to investigate whether a wingbeat frequencies differ among individuals of a species or within an individual over time. This work integrates field and laboratory entomology and engineering, with the future goal of developing an automated mosquito identification system that can increase mosquito surveillance capacity.

Results
Using our semi-automated device, we were able to produce substantial wingbeat datasets from 29 North American mosquito species (Table 1). In total, 21,825 wingbeat files were recorded for building the wave file library which included important vector and nuisance species. Alternating UV-LED lights at either end of the flight tube, controlled by a timer, successfully elicited back-and-forth flight of mosquitoes, increasing the probability of mosquitoes passing through the sensor array and the overall sample size. However, the phototactic response to UV-LEDs differed between mosquito species. The majority of vector and nuisance mosquito species were positively phototactic (e.g., species of Aedes and Psorophora), while amphibian biting (Uranotaenia lowii) and non-nuisance (Wyeomyia) were far less responsive to the UV-LED light source. As such, greater numbers of wingbeat files were recorded for vector species (Aedes aegypti n = 1161; Aedes taeniorhynchus n = 1196; Anopheles albimanus n = 3775; Anopheles quadrimaculatus n = 2241; Culex quinquefasciatus n = 2738). We also observed mosquito flight activity varied according to natural circadian rhythms. Diurnal mosquito species (Ae. aegypti and Ae. albopictus) were relatively more attracted to UV-LEDs during daylight hours, while nocturnal (active at night) or crepuscular (active at dawn/dusk) species (e.g., Anopheles spp.) were more responsive in natural darkness during the night time.
No significant differences (F 4 = 1.3788; P ≤ 0.2398) were observed in wingbeat frequencies between individual females of Cx. quinquefasciatus (Fig. 2). The mean (standard deviation) wingbeat frequencies of five individuals were 448.  (Table 2). Also, wingbeat frequencies of female individuals did not differ significantly over time (r 2 = 0.001; P = 0.8303) (Fig. 2). Wingbeat frequencies were relatively steady over a period lasting 20 min. The mean wingbeat frequency of mosquitoes was 453.3 Hz accounted for 22.95% of the wingbeat variation.

Discussion
Our finding that most mosquito species (26/29) overlapped with at least one other species in the distribution of wingbeat frequency has important implications for using this metric as a predictive variable for discriminating mosquito species using an automated identification system. Other variables or parameters will need to be incorporated into the field capture data to produce a robust classification of mosquito species. Potential variables include mosquito size, UV-fluorescence, location (distribution), habitat, and hour of activity. Mosquito species compositions and abundance are known to change along spatial gradients including land cover, altitude, and human/livestock population density 30 . Therefore, detailed information on habitat associations and geographic distributions of individual species could be used to guide algorithms for classifying captured wingbeats to particular species. For example, two species, Ae. aegypti and Wy. smithii are among 80 mosquito species currently known to occur in Florida 31 that did not significantly differ in wingbeat frequency (Fig. 1), yet these two species differ in habitat association and geographic distribution. Aedes aegypti is considered a domestic (urban/ periurban) species that is common in south Florida but rare in northern Florida. Wyeomyia smithii, in contrast, is restricted to pitcher plant bogs and is not found in southern Florida. Therefore, a mosquito wingbeat captured in southern Florida is far more likely to be an Ae. aegypti than Wy. smithii. Extensive research showed that mosquitoes have specific circadian rhythms (hours of activity), which are dependent on changes in ambient light intensity 32 . Data on circadian rhythms of individual species could  34 . A possible explanation for this variability could have resulted from different recording devices or mosquitoes modulating their flight tone for atmospheric conditions such as temperature. For example, wingbeats in the study 34 were measured at 32.6 ± 0.5 °C while ours were conducted at 23.0 ± 0.5 °C. These differences will be important for compensating wingbeat value in the case of a changing temperature which could be automatically adjustable in a range of wingbeat frequency collected from the future system to identify mosquito species. Species variation in trapping outcome based on trap type would also be an additional predictive variable parameterizing the automated identification system. Mosquito trapping methods vary in their efficacy to capture specific species and life stages which may cause species-specific trap bias 35,36 . For example, the dominant species attracted with light-baited traps may differ from the BioGents Sentinel trap (BG) in a given area where both species co-exist. Our result showed that wingbeat frequencies of Ae. aegypti were not significantly different from that of the An. quadrimaculatus mosquitoes. However, a mosquito collected from BG trap during daylight hours is far more likely to be Ae. aegypti than An. quadrimaculatus, which is more responsive to light traps including the New Jersey light trap or the Centers of Disease Control and Prevention (CDC) light trap 37 . In a field deployment, mosquito wingbeat frequencies may incorporate mosquito trap type to increase accuracy and reliability. Surprisingly, wingbeat did not vary between individuals of Cx. quinquefasciatus and even individual mosquitoes exhibited at relatively stable wingbeat frequencies over time (Fig. 2). This low dispersion of wingbeat frequencies found between individuals transitorily could have resulted from laboratory colony being relatively more homogeneous with respect to size comparing to wild strains. However, size variations in both mass and linear dimensions (individual length, width, height) occur in mosquito populations in nature. Physical characteristics including individual mass, wing area, and wingspan are known to impact wingbeat frequencies in   38 . For example, flight-associated morphometrics across insect orders suggest that larger insects may not afford lower wing loadings for their body mass, consequently, leading to increased wingbeat frequency comparing to the group with a similar wing length 39 . Also, individual variation in wing length and area is a function of numerous allometric traits including life history (e.g., the availability of nutritional resources during the larval stage), genetics, and physiology [40][41][42] . For example, the wing length from a wild population of Cx. quinquefasciatus females ranging from 2.81 to 4.25 mm was found 43,44 . It is also possible that the wingbeat may be related to wing morphometrics, such as a total wing area as well as the unidimensional wing length 39 . The relationship between wing size and wingbeat frequency is supported by the aerodynamic theory that smaller wings produce less force per beat than larger wings; and therefore, more beats are needed per unit time 39 . This trade-off is likely to explain how energetic costs and pressure change mosquito morphometrics which leads to wingbeat frequency change.
As we didn't cover individual variation in physical appearance (wing length and area) and mass, it is possible that if we manipulate female size by abiotic and biotic conditions, then we may observe a clear relationship between the small and large cohorts in single species. This information would have significant implications for understanding evolutionary trade-off and biomechanical aspects, and for the automated identification system based on a wingbeat may incorporate specific external morphological traits (i.e., wing size and length) to increase accuracy and reliability. An image sensor can be relatively easily added to the sensor array so that insect size can be estimated and used as an additional variable in algorithms to predict mosquito species. Our novel device permitted rapid measurement of wingbeat frequency data of a large diversity of mosquito species from laboratory colonies and wild strains. Wingbeat frequencies did not cluster phylogenetically (by genus) but were distributed across higher-order taxa. Few species were significantly different than all others  www.nature.com/scientificreports/ examined but wingbeat frequencies of the same species, Cx. quinquefasciatus, were not highly variable between and within individuals. However, there are many outstanding questions. For example, variation in the wingbeat frequency from all physiological states including gravid and blood-fed females as well as sexual dimorphism must be resolved about the female wingbeat for parameterizing the automated identification system to successfully increase the sensitivity (number) and specificity (species) for species identification because overlapped wingbeat frequencies preclude the use of species classification alone as a method. Finally, our simple and reliable method opens the way for bringing an automated mosquito identification system into the real world and exploring mosquito flight behavior, which is a key component in biodiversity, ecology, and pathogen dynamics.

Materials and methods
Instrument to capture wingbeat frequency. The novel wingbeat recording device included a sensor array with two pairs of infrared emitters (Lite-on, Inc., Taipei, Taiwan) and corresponding receivers (Lite-on, Inc., Taipei, Taiwan) covering horizontal and vertical planes, which captured wing light disturbance when mosquitoes flew through the array (Fig. 3). The sensor array also contained a cellular modem (Simcom, Shanghai, China), a secure digital card (SanDisk, CA, USA), with supporting electronics powered by a 12 V DC source (Fig. 3) provided by TrakitNow, Inc., SC, USA. A mosquito flight apparatus, passing through the sensor array, was made with a section of 18.0 cm × 1.5 cm × 1.5 cm transparent polypropylene tube (ClearBags, CA, USA).
The tube was close-ended and had a 1.5 cm diameter of hole as an access point on the top for mosquito release and withdrawal. Sections of the flight tube within the sensor array were covered with 5.0 cm × 2.0 cm × 1.0 cm of Near-Infra-Red filters (Astra Products, Inc., NY, USA); eliminating unwanted visible spectrum. A pair of 1.5 V ultra-violet light-emitting diode (UV-LED) bulbs in parallel connection was positioned at both ends of the flight apparatus that was operated from 6.0 V by DC adaptor. Voltage was applied to the infrared emitters to maintain a 1.5 vdc ± 50 mv output from the receivers prior to sensing a rapid drop in voltage of at least 100 mv and 200 ms following the sensing of the rapid drop indicating a wingbeat, using an STM32F540G microcontroller. An ESP32-Room microcontroller (Espressif Systems, Shanghai, China) was added to control the on/off interval of the UV-LED such that the UV-LEDs alternated on and off in order to attract mosquitoes flying back and forth through the flight tube and sensor array, continually capturing wingbeats. Sample size varied from availability for mosquito species (Table 1). Maximums for sample size were twenty mated female mosquitoes (3-5 d old) that had never received a blood meal, were collected using an aspirator were released into the flight apparatus. It allowed acclimating to the environment for 5 min. Constant environmental conditions (23.0 ± 0.5 °C, 60.0 ± 5.0% RH) were used. Especially, the wingbeat files from species that medically important (Ae. aegypti, Ae. albopictus, An. albimanus, and Cx. quinquefasciatus) were collected with a minimum of 2000 valid files. When a mosquito passed through the sensor array, the mosquito movement partially occluded the infrared light, resulting in tiny light disturbances (fluctuations). Wingbeat was captured by infrared receivers intermediate in the opposite site as changes in currents which were connected to an electronic board to filter and amplify the wingbeat signals; sensed by the triggering of the STM32's a rapid drop off in voltage as sensed by the microcontroller and ended after a 200 ms recording interval. The voltage was adjusted by the STM32 after the interval (the receivers' output was more than 50 mv out of calibration). The data during the 200 ms interval was recorded, stored, and sent to AWS via a SIMCOMM 5320 cellular modem programmed by Trakitnow, Inc. In order to determine whether the mosquito wingbeat variation was consistent between and within an individual in the same species, individual mosquito wave files in Cx. quinquefasciatus species were collected. We introduced an individual mosquito into the flight apparatus with the sensor array in succession and was removed over valid data points (minimums for sample size). The experiment was replicated five times.   Statistical analysis. Mosquito wingbeat data was collected only in frequency spectrum range from 100 to 2000 and selected based on the validated files to reduce the wingbeat error rate. Extracted wing-beat fragments in each mosquito spp. were altered from mbn to wav (Waveform Audio File Format) using Wave Converter with python 3.7 (Python Software Foundation, OR, USA). Fundamental wingbeat frequencies were analyzed by one-