Flight behaviour of malaria mosquitoes around odour-baited traps: capture and escape dynamics

Host-seeking mosquitoes rely on a range of sensory cues to find and approach blood hosts, as well as to avoid host detection. By using odour blends and visual cues that attract anthropophilic mosquitoes, odour-baited traps have been developed to monitor and control human pathogen-transmitting vectors. Although long-range attraction of such traps has already been studied thoroughly, close-range response of mosquitoes to these traps has been largely ignored. Here, we studied the flight behaviour of female malaria mosquitoes (Anopheles coluzzii) in the immediate vicinity of a commercially available odour-baited trap, positioned in a hanging and standing orientation. By analysing more than 2500 three-dimensional flight tracks, we elucidated how mosquitoes reacted to the trap, and how this led to capture. The measured flight dynamics revealed two distinct stereotypical behaviours: (i) mosquitoes that approached a trap tended to simultaneously fly downward towards the ground; (ii) mosquitoes that came close to a trap changed their flight direction by rapidly accelerating upward. The combination of these behaviours led to strikingly different flight patterns and capture dynamics, resulting in contrasting short-range attractiveness and capture mechanism of the oppositely oriented traps. These new insights may help in improving odour-baited traps, and consequently their contribution in global vector control strategies.


Introduction
defence strategies and that are sometimes even their predators. Thus, mosquitoes need to minimize the risk induced by this interaction with their host by feeding quickly, stealthily and effectively [1,2]. Additionally, when approaching a host, mosquitoes must be aware of any cues announcing host defensive behaviours such as looming objects or air gusts. Surprisingly, the impact of such disruptive cues on the short-range attraction of anthropophagic mosquitoes towards human hosts has been studied very little [3,4]. Actually, mosquito flight dynamics in reaction to humans or objects imitating hosts also received little attention [5].
By contrast, the odour-mediated host-seeking behaviour of mosquitoes has been studied extensively, because it is relevant for mosquito population management (vector control). Anthropophilic female mosquitoes are attracted to their hosts by a species-specific cocktail of human odours and CO 2 [6][7][8]. When a host-seeking mosquito encounters these cues, it performs a stereotypical 'cast and surge' flight behaviour in order to locate the host [6,9,10]. Similar flight behaviour is observed in many species of insects, such as moths, fruit flies and mosquitoes to find potential mates, food or hosts, respectively [11][12][13].
Detailed studies on host-seeking behaviour showed that flying Aedes aegypti mosquitoes (vectors of e.g. dengue and Zika fever) use a combination of cues to find and approach hosts [6,14,15]. At long distances, they only use CO 2 and odours to find a potential host via 'cast and surge' flights. At intermediate distances (of the order of metres), they start to inspect visual cues (highly contrasting objects) near the ground. And at short ranges (less than 1 m), they use heat and moisture cues to find a landing spot on the host [7,9,16,17]. One could suppose that the relatively high importance of visual cues in Aedes aegypti arises from their diurnal behaviour, but recent studies on host-seeking Anopheles mosquitoes shows that these nocturnal mosquitoes have a similar reaction to visual cues under moonlit or starlit conditions [18,19].
Based on the developed knowledge of host-seeking behaviour in mosquitoes, a range of different mosquito traps has been developed [20][21][22]. One promising odour-baited trap type is the counterflow trap that uses a single fan to generate both an inward airflow for capturing mosquitoes as well as an outward directed airflow carrying attractive odour away from the trap (figure 1a). The odour bait consists of a combination of chemicals mimicking human skin odour. Trap models usually combine the odour bait with CO 2 and/or visual cues in contrasting black and white, such as the BG-Sentinel trap and the BG-Suna trap (Biogents AG, Regensburg, Germany) [23,24]. Although odour-baited traps were originally developed as research tools [25], a recent large-scale field study in Kenya showed that in combination with pre-existing bed nets, odour-baited traps reduced the number of human malaria cases by 30% [26]. This study indicated that the use of such insecticide-free traps could now be considered as an effective supplement to conventional vector control systems. By combining multiple host cues such as odours, CO 2 and visual contrast, odour-baited traps aim to trigger the natural host-seeking behaviour of mosquitoes. But surprisingly little attention has been given to the optimization of the capture mechanism of the trap. This might, at least partly, be due to a lack of detailed knowledge on the flight dynamics of mosquitoes in the vicinity of traps. In addition, until now, mosquito traps have been primarily optimized via an iterative design process, where the number of mosquitoes caught is compared between different trap designs [25,27,28]. Thus, obtaining detailed knowledge of mosquito flight dynamics around traps should open new paths for trap improvement and could be crucial in identifying the short-range approach and capture mechanisms involved [29].
To the best of our knowledge, Cooperband & Cardé [5] are the only ones to have studied mosquito flight behaviours near traps. They analysed three-dimensional tracks of Culex quinquefasciatus and Culex tarsalis approaching four models of CO 2 -baited traps in a large field wind tunnel, and showed that the traps had very different capture efficiency (4-26% of upstream released mosquitoes were captured). In addition, they highlighted that, when approaching the trap, mosquitoes decelerated and adopted tortuous flights that varied in dynamics between the traps. These differences in dynamics, especially in the close vicinity of the traps, might help explain the differences in capture rate among the traps, but the spatial and temporal resolution of their flight tracking system did not allow for a detailed analysis of these dynamics [5].
Here, we will zoom in on these tortuous flight manoeuvres close to the trap to elucidate how the close-range response of the mosquitoes affects their capture dynamics. These close-range dynamics complement the long-distance approach dynamics as previously studied by Cooperband & Cardé [5]. We specifically address the tortuous flight paths close to the trap, and how these flight manoeuvres eventually lead to capture or escape from the trap.
For this study, we used female malaria mosquitoes Anopheles coluzzii, and the odour-baited BG-Suna trap (Biogents AG), which was developed and is used for malaria vector control in Africa [26]. We tested the trap in two orientations, the original hanging orientation and an upside-down standing orientation. The trap in its original hanging orientation has an upward-directed airflow for capturing mosquitoes (figure 2a), whereas the upside-down standing trap has a downward-directed airflow for mosquito capture, and thus simulated the widely used and similar BG-Sentinel trap (figure 2b) [23,24]. The use of these two trap orientations allowed us to investigate whether mosquitoes exhibit distinct behaviours in response to odour cues and opposite airflow orientations, and how this affects the capture dynamics.
We filmed a total of 530 mosquitoes flying around the trap using a stereoscopic high-speed videography system, from which we reconstructed more than 2500 three-dimensional flight paths. Based on these results, we found that the tortuous flight behaviour of mosquitoes near odour-baited traps was the result of two distinct and stereotypic behavioural responses of the mosquitoes to the trap: (i) when a mosquito flew towards the trap, it would tend to simultaneously fly downwards towards the floor, possibly in order to host-seek near the ground; (ii) when a mosquito came close to the trap, it responded to the strong air currents induced by the trap or to a lack of short-range host cues, by performing an upward-directed manoeuvre, leading to high vertical accelerations in the flight path. Similar behaviours have been described previously in the literature [9,18,30 -32]. Here we showed that the combination of these two stereotypical behaviours can lead to strikingly different flight dynamics depending only on trap orientation, which consequently led to similar differences in capture dynamics and capture efficiencies between the traps. These new insights into the close-range interaction between mosquito and trap may help in improving future trap designs.

Experimental animals
For this study, we used a colony of Anopheles coluzzii (from the Anopheles gambiae complex [33]) that originated from Suakoko, Liberia in 1987. Since then, the colony has been reared at the Laboratory of rsos.royalsocietypublishing.org R. Soc. open sci. 5: 180246 Entomology (Wageningen University & Research, The Netherlands) at a temperature of 278C, a relative humidity of 70%, and with a clock shifted 12 L : 12 D cycle. Adults were kept in BugDorm cages (30 Â 30 Â 30 cm) in which they were fed on sugar-water with 6% glucose solution and offered daily blood meals from a blood bank (Sanquin, Nijmegen, The Netherlands) via a membrane feeding system (Hemotek, Discovery Workshop, UK). Thereafter, they were allowed to lay eggs on wet filter paper, which were then moved to a plastic tray filled with water. Liquifry No.1 fish food and TetraMin Baby were provided for larvae feeding. Finally, pupae were placed in new BugDorm cages to emerge. During our experiments, we used non-blood fed female malaria mosquitoes, which had most likely mated (5-10 days post-emergence with males and females housed together). They were collected between 12 and 16 h before experiments and were not blood fed in order to increase host-seeking responses.

Experimental set-up
Experiments were conducted in a 3.01 Â 4.92 Â 3.25 m (width Â length Â height) climate room at the Laboratory of Entomology (Wageningen University & Research, The Netherlands). The room was maintained at 278C and 70% relative humidity and had a continuously running air filter system preventing accumulation of odours or CO 2 (see Spitzen et al. [10] for more details). The experimental set-up consisted of a stereoscopic high-speed camera system that filmed the vicinity of a BG-Suna trap (Biogents AG) positioned in front of two perpendicular walls, with the trap centre at 35 cm from each wall (figure 1). A net covered a volume of approximately 8 m 3 around the set-up in order to prevent mosquitoes from escaping.
The BG-Suna trap is an odour-baited trap with a diameter of 52 cm and height of 39 cm that was developed for malaria vector control in Africa [26]. For each separate experimental trial, the trap was randomly positioned in its original hanging orientation (figure 2a), or in a standing orientation in which it resembled a BG-Sentinel trap (figure 2b) [23,24]. In its original hanging orientation, the trap produced an upward-directed airflow for capture, whereas the upside-down standing trap produced a downwarddirected airflow for mosquito capture. The capture entrance of the hanging and standing trap had a respective height of 81 cm and 54.5 cm above ground level (figure 2a,b, respectively). These heights were chosen to keep the camera heights equal to 62 cm in order to avoid repetitive realignment and recalibration of the camera system. Note that because of this reason, the hanging trap was positioned higher than the 30 cm height that has been found optimal for capturing mosquitoes [25].  We used the MB5 blend of five attractants [34] inside each trap to simulate human odour. The CO 2 release pipe of the trap was connected to a pressurized gas canister containing a mixture of 5% CO 2 þ 95% air and with a flow rate of 200 ml min -1 . To minimize blind spots in the camera views, this CO 2 pipe was shortened by 3.25 cm relative to the original length for the Suna trap. All handling of the materials and mosquitoes was done wearing nitrile gloves to minimize the risk of skin odour contamination of the traps.
Because Anopheles mosquitoes are night active, we performed the experiments in dimmed light conditions, using a single spotlight (15 W incandescent bulb) directed towards the ceiling, and experiments were carried out during the period of the day which corresponded to the dark phase of the mosquito rearing; the period when the An. coluzzii were expected to exhibit the greatest degree of host-seeking behaviour.
Mosquito tracks around the trap were recorded using two synchronized high-speed cameras (PROMON 501 camera head with NIR sensors and 45 mm lenses), filming at 90 frames per second with a resolution of 1240 Â 1080 pixels. Because mosquitoes cannot see infrared light [35], we used two infrared light emitting lamps (Bosch Aegis SuperLed, 850 nm, 108 beam pattern (SLED10-8BD)) as illumination for the camera system. The camera system was calibrated at the start of each experimental day, with the use of a modified direct linear transformation (DLT) algorithm [36], by comparing known fx, y, zg coordinates of 25 suspended lead beads to their detected pixel coordinates in each camera. In addition, lens corrections were applied using pictures of a chequerboard pattern and using a Matlab script [37].
For each trial, 10 female mosquitoes were released from a holding container within the flight arena, after which the experimenter immediately left the room. From outside the room, the experimenter then started the 15 min video recording, and thus all experiments were performed without a potentially disturbing human present in the room. Five minutes after each trial, captured and non-captured mosquitoes were collected and killed. Out of the 61 successfully performed trials, a total of 53 trials were analysed (32 for the hanging trap and 21 for the standing trap). Five trials were discarded because male mosquitoes were found in the arena or trap and three other trials were discarded due to illumination or calibration errors.

Simultaneous tracking of multiple flying mosquitoes
To compute the three-dimensional tracks (flight segments) from the stereoscopic recordings of flying mosquitoes, we first determined the two-dimensional positions of the mosquitoes within each image using the image processing toolbox of Matlab (MathWorks). Then, for each camera recording, the two-dimensional flight tracks were constructed using a 'Hungarian linker' algorithm [38]. This algorithm reconstructs the tracks by finding the minimum distance between the detected positions of the mosquitoes in subsequent frames. Missed detections were taken into account by keeping tracks alive for 10 frames before deciding that they had ended. In the rare event that two tracks merged and then split up, two new tracks were started.
Next, the two-dimensional tracks within each camera view were combined into three-dimensional tracks using a DLT method [36]. A DLT error was calculated as the root mean square of the error (RMSE) between the original time-overlapping two-dimensional tracks and the two-dimensional back projections of the three-dimensional reconstructed tracks. To find the correct matches between twodimensional tracks, we used an RMSE threshold that separated RMSE distributions for matching and non-matching tracks.
Finally, pieces of three-dimensional tracks were stitched together whenever a single two-dimensional track matched multiple two-dimensional tracks in the other image. A Hampel filter was added to remove positional outliers on the three-dimensional tracks. In this way, the complete flight track of a mosquito could be reconstructed. Individuals were, however, not identified because mosquitoes could enter and exit the filmed volume multiple times during one experiment. Throughout each resulting three-dimensional flight track, we calculated the mosquito's linear and angular flight speeds, and its linear acceleration (electronic supplementary material, figure S1). The angular flight speed was calculated as v ¼ Du=Dt, where Dt ¼ t n À t nþ1 is the time elapsed between two consecutive video frames n and n þ 1, and Du is the turn angle defined as where v n and v nþ1 are the three-dimensional velocity vectors of the mosquito at video frames n and n þ 1.

Analysing three-dimensional flight tracks
Owing to the high number of reconstructed tracks, a statistical approach was needed to visualize the flight dynamics of the mosquitoes around the two traps. For this purpose, we assumed the average flight behaviour of the mosquitoes around the trap to be axially symmetric, despite the presence of the trap's CO 2 pipe and the slope of the trap's entry tube (visible in figure 3a). We divided the filmed volume into multiple threedimensional rings (figure 3a), which were centred around the trap's axis of symmetry. We computed statistical metrics from the mosquito's position over time in each of the rings. To allow metric comparisons across the rings, the volume of all rings was the same. We then projected each three-dimensional ring onto a two-dimensional parametric space with radial distance and vertical position as the key dimensions (figure 3b). Similarly, top-down view projections were reconstructed by dividing the flight volume into three-dimensional vertical rods projected onto a two-dimensional horizontal plane. Using this method, each flight dynamics metric was visualized as a set of two-dimensional heat maps, one on the radial-vertical plane and one on the horizontal plane. In addition to the translational speed, angular speed and acceleration, we also visualized the distribution of positional likelihood and capture probability of mosquitoes around the trap. The estimated capture probability was expressed by the percentage of mosquito tracks in each sub-volume that ended in capture. We defined the positional likelihood as the normalized probability of a mosquito to fly within a certain sub-volume (e.g. a specific three-dimensional ring) of the field of view. It was calculated as where P i is the likelihood that a mosquito is present in cell i, whereby cell i represents a threedimensional ring projected onto the previously described two-dimensional parametric space. n i is the number of video frames that a mosquito was present in cell i throughout all sets of recordings, N is the total number of frames recorded, and I is the total number of cells within the recording volume. Thus, a random flight behaviour would result in a uniform probability throughout the region of interest, with P i equal to one for all cells. Furthermore, we visualized the mean flight dynamics as velocity vectors within the axisymmetric radialvertical plane. Based on the resulting velocity fields, we visualized the average flight paths of mosquitoes by computing streamlines based on their velocity fields using the linear integral convolution algorithm [39].
For all calculated parameters, we performed sensitivity analyses to test the independence of our results to cell size, to the total number of analysed tracks and experimental duration (see electronic supplementary material, figures S2 -S4). Additional data as well as 95% confidence intervals of the metrics presented in figures 5 -8 are shown in the electronic supplementary material, figures S5-S10. For the statistical tests, we used a one-sample Kolmogorov-Smirnov test to determine whether data were normally

Analysing the airflow dynamics around the mosquito trap
The path of a mosquito flying around a trap results from a combination of the mosquito's manoeuvre dynamics and the air movements induced by the trap. Thus, to determine the effect of air movements on mosquito flight dynamics, we measured the vertical speed component of the airflow using a onedimensional hotwire anemometer (tetso 405i). This device had a 1 Hz sample rate, and thus did not allow us to quantify turbulence levels. For these measurements, we again assumed the axial symmetry of the airflow, and thus ignored possible local differences in airflow around the CO 2 pipe or due to the asymmetric design of the Suna tube entrance (see positional likelihood around the CO 2 tube in electronic supplementary material, figure S11). Horizontal air velocities were not measured, as we expect that these were much lower than the vertical velocities induced by the vertically oriented fan. With a custom-built set-up, we used the hotwire anemometer to measure the vertical airspeed at 76 locations within a two-dimensional vertical plane oriented perpendicularly to the open trapdoor in the     Of all the 2570 recorded flight tracks, only 87 tracks resulted in capture of the mosquito by the traps, divided into 25 captures by the hanging trap and 62 by the standing trap. Because of the low number of tracks that led to capture, we were unable to compare the flight dynamics of the captured and non-captured tracks using our two-dimensional parametric space. Instead, to determine what causes the differences in number of captures between the traps, we calculated three different capture ratios: (i) the percentage of released mosquitoes that were captured (R mosquitoes ¼ N captures /N released . 100%); (ii) the percentage of recorded tracks that led to capture (R tracks ¼ N captures /N tracks . 100%) and (iii) the number of captures per minute flight duration (capture frequency f captures ¼ N captures /T tracks ).
On average, only 8% of the mosquitoes released were captured by the hanging trap, whereas 30% of the mosquitoes were captured by the standing trap (electronic supplementary material, table S1). Therefore, the standing trap captured almost four times the percentage of released mosquitoes captured by the hanging trap. However, because there were many more flight tracks around the standing trap, the average percentage of flight tracks that led to capture by the standing trap was only 1.4 times higher than that

Positional likelihood of mosquitoes
Heat maps of the positional likelihood of mosquitoes flying around the traps allow the identification of regions with increased activity, which are quite different between the two trap orientations ( figure 4). Above the standing trap, a cone-shaped region of increased activity is visible, in which mosquitoes were up to five times more likely to be present compared to the average. By contrast, around the hanging trap no such clearly defined region of increased activity was found, except for a smaller, cylindrical region directly underneath the entry tube edge within which mosquitoes are up to four times more likely to be found. The top-down views ( figure 4b,d) show that flight activity is approximately axially symmetric, except for a small reduction in activity near the CO 2 pipe (see also electronic supplementary material, figure S11), and for the hanging trap an interesting concentration of mosquitoes near the bottom corner of the back walls of our experimental set-up was present (figure 1).

Airflow dynamics of the traps
We visualized the airflow dynamics around the standing and hanging trap as the two-dimensional distribution of vertical speeds ( figure 5). The distribution of air speed around the traps was consistent with previous findings [25], where vertical airflow speeds were very high at the mouth of the entry tube (up to 3 m s 21 ), but these speeds rapidly decreased when moving away from the trap entrance. As a result, the region in front of the tube where the airflow is more than 1.5 m s 21 is only 3 cm high and 10 cm wide. This suggests that for a mosquito to be inevitably captured by the inward-directed airflow, it needs to pass very close to the trap entrance. A comparison of the vertical airspeeds around the standing and hanging trap (figure 5 and electronic supplementary material, S5) shows that trap orientation has a relatively small effect on airflow dynamics, suggesting that a possible ground effect below the hanging trap is mostly negligible.

Mosquito flight dynamics
Based on all flight trajectories, we determined the translational flight speeds, angular speeds and accelerations of mosquitoes around the hanging and standing trap (electronic supplementary material, figure S1). The mosquitoes flying around the standing trap flew faster, had higher turn rates as expressed by higher angular speeds, and they produced higher body accelerations than those around the hanging trap ( p , 0.001; see electronic supplementary material, table S1).
The two-dimensional distributions of the translational speeds of mosquitoes flying around the standing and hanging traps were similar (figure 6a-d). Mosquitoes had their highest mean ground speeds (up to 0.5 m s 21 ) near the trap entrance, whereas in both trap orientations they had their lowest mean speeds at a radial distance of 10-20 cm from the entry tube. At both these regions with increased and reduced translational speeds, the angular speeds were high (figure 6e-h).
In contrast with the relatively high similarity in absolute translational and angular flight speeds around the traps, the velocity fields around the traps were strikingly different (figure 7). The pattern of mosquito streamlines near the hanging trap suggests that two regions can be identified, separated by a diagonal that runs from the top right to the bottom left of the velocity field. On average, mosquitoes that flew below this diagonal continued to fly downwards and away from the trap, whereas those flying above this border would initially fly downward, but when they got close to the trap entrance, they turned towards the trap entrance and got caught. Example flight tracks of mosquitoes within these two regions are in figure 2a and electronic supplementary material, movie S1.
Around the standing trap, mosquitoes followed a more complex circulating flight dynamic. On average mosquitoes entered the filmed volume by flying down towards the trap platform, after which they turned towards the black entry tube. When they were above the tube, mosquitoes either got caught by entering the trap, or they escaped capture by accelerating upwards. After flying upwards for approximately 20 cm, on average they turned around and started to fly again downwards towards the trap platform, thus completing the loop. Two typical examples of mosquitoes performing such a circulating flight manoeuvre can be seen in figure 2b and electronic supplementary material, movie S2.
As for the hanging trap, we also identified the capture and escape regions for the standing trap based on the distributions of mosquito streamlines. The area for which the streamlines end at the trap entrance is much smaller for the standing trap than for the hanging trap, suggesting that mosquitoes need to approach the standing trap entry tube more closely before being captured than for the hanging trap.
As expected, accelerations of mosquitoes flying around the trap were highest near the trap entrance ( figure 8a,b and electronic supplementary material, figure S6). Around both traps, the mosquitoes flying close to the tube entrance tended to accelerate upwards, despite the oppositely oriented airflow for capture of the different traps. This means that the mosquitoes flying near the entry tube of the standing trap, rsos.royalsocietypublishing.org R. Soc. open sci. 5: 180246 accelerated in the direction opposite to the airflow direction, and thus avoided capture. On the other hand, the mosquitoes flying near the hanging trap also accelerated upwards, but in this case, it was in the same direction as the airflow, and thus these mosquitoes flew straight into the mouth of the trap.

Distribution of the capture probability
The distribution of the capture probability is remarkably different for mosquitoes flying around the two oppositely oriented traps ( figure 8c,d). Mosquitoes that approached the entry tube of the hanging trap within a 10 cm radius have a 75% chance of being caught, whereas for the standing trap mosquitoes need to enter the single cell directly above the entry tube for the likelihood of being caught to reach 75%. We reconstructed the three-dimensional volume within which more than 75% of detected mosquito tracks resulted in captures by revolving the two-dimensional heat map around the axis of symmetry. For the hanging trap, this volume was 17.5 times larger than for the standing trap (electronic supplementary material, table S1).

Discussion
We reconstructed three-dimensional flight tracks from stereoscopic high-speed videos of mosquitoes flying around an odour-baited trap set in two different orientations. The resulting high number of detected tracks was essential to dissect the flight behaviour of mosquitoes around the traps as it allowed the computation of the average flight dynamics throughout the complete region of interest. To study how these dynamics vary in space and as a function of trap orientation, we visualized these three-dimensional flight dynamics using two-dimensional heat maps and vector fields. For this, we assumed that the average flight behaviour of mosquitoes around the trap was axisymmetric with respect to the axis of symmetry of the trap, which was confirmed as almost all top-down heat maps showed exclusively axisymmetric patterns.
The only deviation from this symmetry was an off-centred volume under the hanging trap within which mosquitoes had a high probability of being present (figure 4b,d). Mosquitoes might have been attracted to this region because of an accumulation of odours and CO 2 in this corner, or because a shadow was cast here by the hanging trap. But because top-down views of all other metrics presented in this paper did not show a bias towards this area, we cannot conclusively determine the cause of this accumulation. Additionally, we observed a small reduction in activity near the CO 2 pipe of both traps (figure 4b,d and electronic supplementary material, figure S11a), suggesting that mosquitoes were avoiding this region, possibly due to increased CO 2 concentrations or airflow anomalies near the shortened CO 2 pipe [7,9,42,43]. Note that both the increased activity of mosquitoes at the off-centred volume under the hanging trap and the reduced activity near the CO 2 pipe had a relatively small effect on the flight pattern of mosquitoes in the near vicinity of the trap (figure 6 and electronic supplementary material, figure S11b, respectively), suggesting that the assumption of axisymmetric flight behaviour around the trap is still valid.
Beside these small anomalies, the heat maps of positional likelihood of the mosquitoes (figure 4), and of mean flight speed and mean angular speed (figure 6) contribute some new elements concerning the short-range attractiveness of the studied odour-baited traps. Indeed, with the exception of the previously described regions, the volumes where mosquito activity was the highest are also where their mean flight speeds and mean angular speeds were the highest, suggesting that these mosquitoes were performing casting behaviour and might thus have been host-seeking. This is consistent with findings of a previous study whereby host-seeking mosquitoes increased their flight tortuosity in the proximity of a host [10]. Given that the used odour-baited trap has been shown to be successful in attracting host-seeking mosquitoes [25,26], we hypothesize that the majority of mosquitoes flying close to the traps were attracted by the trap, despite that random encounters with the traps likely also occurred.
The question remains as to what particular sensory cues trigger the search behaviour within the highly unsteady airflow conditions around the trap. For example, mosquitoes have been found to surge upwind more easily in airflow well mixed with odours and in turbulent CO 2 plumes [44,45]. To answer this question, an extensive study of the airflow turbulence levels and three-dimensional distributions of CO 2 and odour would be required [46].

Stereotypical mosquito flight dynamics
The flight dynamics of mosquitoes around the two oppositely oriented traps were strikingly different but also highly stereotypical and repeatable ( figure 7). Near the hanging trap set-up, mosquitoes flew on rsos.royalsocietypublishing.org R. Soc. open sci. 5: 180246 average down towards the ground. But if they flew within a range of approximately 10 cm from the trap entry tube during this downward flight, they would turn and start to fly upwards, and as a consequence likely be caught by the trap. Mosquitoes flying around the standing trap showed a very different, circular flight pattern. These mosquitoes would initially also fly downwards when entering the filmed volume, but in this case towards the trap platform instead of the ground. When they came close to the entry tube, they would also turn and start to fly upwards, which resulted in most cases in a successful escape from the trap. After this manoeuvre, they would turn around and again start to fly downwards towards the trap platform, completing the circular flight pattern. For the standing trap, mosquitoes had to approach the entry tube much closer in order to be captured than for the hanging trap ( figure 8c,d).
Because the trap-induced airspeeds were very similar between the traps (figure 5; electronic supplementary material, S5), the large differences in flight dynamics around the two traps must be the result of a difference in behavioural response towards the two oppositely oriented traps. Here, we hypothesize that these flight patterns are the result of two stereotypic behaviours in host-seeking mosquitoes: (i) mosquitoes that approached a trap tended to also fly downwards to the ground and (ii) mosquitoes that came close to the traps changed their flight direction by rapidly accelerating upwards, possibly reacting to adverse high-velocity airflow cues or to a lack of short-range host cues. This set of behaviours can explain the complex flight patterns observed around the traps, as well as the differences in flight patterns around the oppositely oriented traps.

Upwind-and downward-directed approach flights
Upon entering the observed volume and despite inverse trap airflow, mosquitoes flew on average downwards, as well as towards the axis of symmetry of the trap, for both traps (figure 7). While doing so, they had low angular speeds and low positional likelihood ( figure 4). Thus, mosquitoes did not remain long in the flight volume for this approach phase. As a result of the downward orientation, mosquitoes naturally flew toward the standing trap, whereas they tended to fly away from the hanging trap.
This flight behaviour is very similar to that described in female mosquitoes seeking a human host. Female Anopheles mosquitoes preferred to land on human body parts that were closest to the ground, which they found by flying downwards and upwind by tracking odours and convective air currents produced by that host [30,47]. When approaching an odour source, Aedes aegypti mosquitoes flew towards the ground to inspect visually intriguing objects [9,18]. The similarity in flight pattern of our mosquitoes flying near the trap to those described in the literature suggests that our mosquitoes performed a similar host-seeking behaviour. In our case, the upwind and downward flight behaviour might have been triggered by the detection of increased concentrations of odours, CO 2 and airflow turbulence, but the importance of visual cues cannot be excluded as possible light conditions differences have not been investigated.

Fast upward-directed flight manoeuvres
The second phase of mosquito flight dynamics takes place in the near vicinity of the trap entry tube. It starts at the boundary between the high-speed airflow volume a few centimetres from the visually contrasting entry tube and the volume where attracted mosquitoes are host-seeking. Here, on average, mosquitoes were found to quickly turn by accelerating and flying upward. For the mosquitoes that flew below the hanging trap, these rapid upward flight manoeuvres often led to capture, whereas the mosquitoes that flew above the standing trap, accelerated away from the trap and thus mostly escaped successfully.
Similar fast upward-directed flight behaviours have been described previously for other host-seeking insects [18,31,32]. Host-seeking Anopheles mosquitoes tended to fly quickly upwards after inspecting black tiles on the ground [18], and horse and deer flies tend to fly upwards after having inspected potentially interesting visual cues [31,32]. In fact, the Malaise trap and the Manitoba fly trap have both been designed to trap such upward flying insects [31,32].
The question remains about what sensory cues trigger these rapid upward flight manoeuvres near the trap. This could be the absence of short-range host cues such as heat or moisture [17,18], or the flight manoeuvre could be the result of positive phototaxis, reaction to visual cues, or to avoid/evade the high-speed airflow regions. The upward flying mosquitoes in the shadow of the hanging trap flew towards its black entry tube, whereas mosquitoes flew away from the black entry tube of the standing trap. Therefore, positive phototaxis as well as reaction to visual cues most likely do not explain these manoeuvres. Thus, these upward manoeuvres were probably either the result of the absence of rsos.royalsocietypublishing.org R. Soc. open sci. 5: 180246 short-range host cues, or they might have been executed to avoid or evade the high-speed airflow regions induced by the traps.
Should these upward-directed flights be evasive manoeuvres, then mosquitoes did not use information about the direction of the airflow to steer away from a potential threat, as all manoeuvres were directed upwards regardless of the direction of the airflow. This makes these manoeuvres strikingly different from the evasive manoeuvres described in flying fruit flies, hawkmoths and hummingbirds, that are directed away from the danger [48,49], or downwards towards the ground [50]. Horse flies, on the other hand, have also shown to fly upwards after inspecting potential hosts [31,32]. Because mosquitoes and horse flies both feed on terrestrial animals, these upward-directed evasive or avoidance manoeuvres might be particularly successful for such animals.
The stereotypical upward accelerating manoeuvres of our mosquitoes also explain the large difference in capture dynamics between the oppositely oriented traps. The air volume around the trap entrance within which 75% or more of the mosquitoes were caught, V 75% , was 17.5 times larger for the hanging trap than for the standing trap ( figure 8c,d). For the standing trap, this V 75% volume overlays well with the region within which the airspeeds directed into the entry tube were more than 1.5 m s 21 (figure 5), but for the hanging trap the V 75% volume extended well outside this high airspeed region. In flow tunnel experiments, it has been demonstrated that mosquitoes are able to fly against airflow with speeds of up to 1.5 m s 21 [51], and thus mosquitoes caught by the standing trap were most likely sucked downwards into the trap, as they were unable to accelerate upward fast enough to avoid capture. Because the V 75% volume underneath the hanging trap extended well beyond the region with high suction airspeeds, many of the mosquitoes captured by the hanging trap must have actively flown into the entry tube when performing an upward-directed flight manoeuvre.

Combining the two stereotypical flight behaviours
The flight dynamics observed near the two traps can be interpreted as a combination of the upwind-and downward-directed approach flights and the fast upward-directed evasive manoeuvres. Although alternative behavioural explanations of the observed flight dynamics are possible, the here-described flight behaviours have previously been identified in host-seeking insects [9,18,30 -32], and the combination of these two behaviours explains well the striking differences in flight dynamics around the oppositely oriented traps (figure 9). The upwind-and downward-directed flight behaviours illustrate why mosquitoes tended to fly away from the hanging trap, and towards the horizontal platform of the standing trap; the fast upward-directed flight manoeuvres caused mosquitoes to fly towards the capture entrance of the hanging trap and away from the standing trap, and thus also explain the larger capture region around the hanging trap (figure 8). After the upward flight movement away from the standing trap, these mosquitoes switched back to the downward-directed flight pattern, explaining the advent of the circular flight path, and why mosquitoes remained near the standing trap for a longer time.

The flight behaviour of mosquitoes explains trap efficiency
Our results suggest that the standing trap has a higher short-range attractiveness, as expressed by both a higher number of detected flight tracks (N track /N trial ) and larger flight duration per trial (T track /T trial ) for mosquitoes flying around the standing trap. By contrast, the hanging trap has a better capture mechanism because the air volume around the trap entrance within which 75% or more of the mosquitoes were caught, V 75% , was more than 17 times larger for the hanging trap compared with the standing trap. However, both the percentage of released mosquitoes caught and the percentage of flight tracks that led to capture (R mosquitoes and R tracks , respectively) were higher for the standing trap. Although R mosquitoes and R tracks were not significantly different between the traps, this suggests that the less efficient capture mechanism of the standing trap was more than compensated by its superior short-range attractiveness. In this way, the greater number of mosquitoes captured by the standing trap seems to be not only because it attracts more mosquitoes to its vicinity, as was previously suggested [5], but also because mosquitoes remained near the standing trap for a longer period of time.
These results, and especially the differences in capture efficiency, would need to be verified in field experiments, where the wind and light conditions would likely impact trap finding by the mosquitoes as well as their use of visual cues. In addition, the fact that our experiments were performed in an enclosed environment, where the proportion of random encounters might have been relatively high, may have impacted our results. However, such random encounters would probably only increase the background noise on the observed flight dynamics and should not result in distinct flight patterns. Finally, height differences might have affected how easily mosquito were finding the trap, and hence the trapping efficacy [25]. This subject would deserve to be studied in dedicated experiments.
Besides contributing to the expansion of general knowledge on mosquito flight behaviour, our results help to understand the short-range attractiveness and the capture mechanism of odour-baited traps. This new insight could be used to develop novel trap designs with improved trapping efficiency. Thus, in our opinion, such trap design process would greatly benefit from the use of iterative testing of traps in studies similar to this one. Because the Suna trap is already part of a successful vector control system [26], it is likely that the resulting trap improvements would make a valuable contribution to the fight against vector-borne disease.