The core device of the mosquito trap has two chambers (collection boxes), one of which collects mosquitoes determined to be Ae. aegypti or Ae. albopictus and the other collects mosquitoes that are determined to be Culex (Cx.) quinquefasciatus (Figure 1). Above these two chambers, a circular capture plate is paired with a camera that captures images of mosquitoes entering the bottom two chambers. The capture plate is made of the following components: 1) the main plate, 2) a sliding guide and an attached bottom plate, and 3) a stepper motor.

The main plate is in the form of a cylinder on which a linear track is etched to place the sliding guide. The main plate provides an area (the capture area) for the mosquito to enter the trap and undergo detection (mosquito detection is described in detail in the image recognition system section). The main plate also rotates when needed when a mosquito enters the trap (Figure 1 and Figure 2). The sliding guide is placed on the main plate’s linear track, with a bottom plate attached to the part that slides freely on the linear track. When the sliding guide slides to its limit on one side, an area forms between the track and the top of the bottom plate; this area is called the capture area. Mosquitoes can fly into this area and undergo detection. When the sliding guide slides the bottom plate to its limit on the other side, it closes off the capture area and pushes the mosquito, if present, from the capture area to the top of the track.

A stepper motor is attached to the main plate through the center of the circle. The stepper motor rotates the base plate when the mosquito trap detects target mosquitoes that fly into the capture area. There is a ceiling with an opening in the middle above the capture plate. This ceiling comprises an area identical to that of the capture plate and leaves the mosquito nowhere to go except through the top opening. The ceiling helps prevent mosquitoes from escaping during capture.

The core device of the mosquito trap was placed into a BG trap (Biogents™, Regensburg, Germany). Mosquito attractants (Guangzhou Baikong Biological Co., Ltd., China) are placed on the BG trap cover or inside the BG trap itself to lure mosquitoes into the mosquito trap capture area (Figure 1). The original BG trap fan is replaced with an AC110V fan (17251A1-HBAPL-TC-N, Sunta, Taiwan), and it is protected by a casing to prevent foreign objects from contacting and breaking the fan blades and entanglement of the wire from the sensors. The mosquito trap was also equipped with a carbon dioxide sensor module (MG811, Sandbox Electronics, Finland) to detect CO₂ concentration and other sensors to detect humidity and temperature (DHT22, Aosong Electronics Co., Ltd., China), as shown in Supplementary Figure S1. A Raspberry Pi microcontroller (Module B, Raspberry Pi Foundation, England) is used to control all of the electrical components. Once a mosquito enters the entry port and is identified with the image recognition software, as described below, the mosquito species is then classified according to the phenotypic characteristics and collected in the respective chamber (collection box) (Supplementary Figure S2).

When the mosquito trap is not actively trapping mosquitoes, the empty rectangular track on the capture plate aligns with the ceiling opening. The sliding guide’s base plate is at its lowest limit, forming a space in the rectangular track between the top of the sliding guide and the ceiling opening. This area is called the capture area. When a mosquito is detected in the capture area, the capture plate is then rotated by the motor controlled by the Raspberry Pi, which immediately closes the opening, trapping the mosquito in the capture area. When the capture plate rotates more than 90°, the sliding guide’s base plate is lowered by gravity. When the bottom plate starts gliding along with the sliding guide, the capture area decreases and forces the mosquito to move in response to the capture plate (Supplementary Figures S2A,B). Finally, the bottom plate reaches its upper limit. There is no space left in the capture area, and thus the mosquito is forced inside the storage chamber (Supplementary Figure S2C).

After capturing the mosquito, the capturing plate rotates in the opposite direction back to its original position, as confirmed by a positioning sensor (TCST2103, Vishay Intertechnology, United States), as shown in Figures 2D, F. The bottom plate is then again lowered by gravity until it reaches its lowest limit. Finally, the capture area is restored, and the trap is ready to catch another incoming mosquito (Figure 2F). Since the capture mechanism requires rotation of the capture plate up to 180°, the capture plate can be controlled to rotate in different directions according to the detection of different types of mosquitoes, thus achieving the goal of effectively sorting different types of mosquitoes into different chambers.

The image recognition tool is essential for the proper functioning of the proposed mosquito trap. The image recognition tool determines whether a mosquito is present in the entry box and, if so, whether it is an Ae. aegypti or Cx. quinquefasciatus mosquitoes. The algorithm is initiated when motion is detected. To detect the motion, the difference between images from successive timeframes is analyzed to determine whether the movement has occurred; then, the contours of the different images are detected, and for each contour form, a subimage is fed into the neural network. This whole process is implemented with OpenCV. Next, a neural network, that is, sufficiently fast and accurate in detecting the types of mosquitoes is developed on the Raspberry Pi device. We developed two imaging platforms (i.e., static and living imaging platforms) to effectively collect mosquito image data to train our neural network and adapted the SqueezeNet model (Iandola et al., 2016) as our neural network model, where we utilize the method of transfer learning to train our recognition system. The deployed algorithm incorporates OpenCV for motion detection and our trained neural network to recognize mosquitoes in real-time using the Raspberry Pi.

A neural network trained to differentiate between different mosquitoes is developed to complete the image recognition process (Figure 3). The development of the neural network consists of three significant steps. First, an imaging system is employed to obtain training data by imaging different mosquitoes. Such a system is used to obtain sufficient training data by imaging different static or live mosquito specimens. Finally, the obtained data are fed into an appropriate neural network model for transfer learning.

Due to its success in the ImageNet competition, AlexNet has become one of the most popular research tools since 2012. We use the SqueezeNet neural network to develop the mosquito image recognition system. It is a lightweight and efficient convolutional neural network (CNN) model with 50 times fewer parameters than AlexNet, but the model performance (such as accuracy) is close to that of AlexNet. Small models have many advantages, such as reducing the number of calculations during model training and predictive identification. The speed of a single step is faster, and there are fewer parameters than in a larger model; therefore, the video memory occupied by the network is smaller. The smaller model is beneficial in field-programmable gate array (FPGA) and embedding type device deployment. The proposed mosquito trap uses edge computing to directly perform image recognition of mosquitoes with the system embedded in the mosquito trap. No additional computer or server is needed. Therefore, the model and calculation volume should not be too large or compressed. The model method mainly comprises network pruning, quantization, and knowledge distillation. To reduce the amount of calculation (3 × 3 × number of input channels x number of filters), in the CNN, 3 × 3 convolution kernels (filters) are mostly replaced with 1 × 1 convolution kernels (filters). This concept effectively reduces the parameters by nine times while retaining the accuracy of model identification and then reduces the input of 3 × 3 convolution kernels. The number of input channels dramatically reduces the number of convolution operations. The first layer is a standard convolution base layer (Conv) in terms of structure, with nine fire modules in the middle. The fire module divides the original convolutional layer into two layers, including the squeeze layer of 1 × 1 convolution filters and 1 × 1. It comprises an expanding layer of 3 × 3 convolution filters interspersed with max pooling layer compression parameters, retains features, and prevents interference. After connecting a convolutional layer to control the size of the input and output, global average pooling is used instead of fully connected layers to reduce parameters and reduce overfitting. According to the literature, SqueezeNet can reduce the parameters by 50 times while retaining the same accuracy as AlexNet.

The first imaging system we developed allowed us to take pictures of mosquitoes in two directions at different angles (Supplementary Figure S3) and use a MATLAB-based graphical user interface (GUI) (Supplementary Figure S3A and Supplementary Movie S1) to set parameters and perform imaging operations. The system consisted of a platform that could be rotated along the z-axis using a stepper motor (17hs1910-p4170, JIALIBAO Electronics, China) and rotated 360° (Figure 3B). Mosquitoes were attached to a platform by insect pins. A camera bracket fixed the camera and rotated it along the x-axis (Figure 3B). However, due to platform limitations, the rotation angle was limited to 160° (Figure 3B).

On both sides of the camera, a controllable light-emitting diode (LED) (61-238/LK2C-B50638F6GB2/ET, Everlight Electronics Co. Ltd., Taiwan) was designed and placed to achieve necessary lighting conditions. The entire system was placed in a studio with a white background to ensure that the captured images contained only the target mosquito. Some image samples obtained from this imaging system are shown in Figures 3C, D. Different types of mosquito samples were placed on the platform for photographing, and these images formed the basis of the dataset used to train the network (Supplementary Figure S4).

In addition to obtaining fixed specimen images for neural network training, we also developed a system to obtain live mosquito images to improve the database used for neural network training. The setup resembled the capture area of our mosquito trap (Figure 4A) and included a small, 3D-printed (3030K, Kingssel, Taiwan and Delta, FLUX, Taiwan) box, a camera (Raspberry Pi camera V2, Raspberry Pi Foundation, United Kingdom) and an LED (61-238/LK2C-B50638F6GB2/ET, Everlight Electronics Co., Ltd., Taiwan). The small box, camera, and LED were affixed to a 3D printed base to prevent shifting when moving the setup or putting live mosquitoes into the box. One side of the box was transparent to allow the camera to monitor the inside of the box, and the top of the box was sealed by another transparent plate. With this design, the light from the LED could penetrate the inside of the box. This was to eliminate shadows in the images captured by the camera and to prevent the characteristics of shadows from affecting neural network training.

We used a Raspberry Pi (Raspberry Pi three Model B, Raspberry Pi Foundation, United Kingdom) to control a DC motor (DC Gearbox Motor, TT, China) and the camera used to capture images (Figure 4A). The captured images were stored in a secure digital (SD) memory card in the Raspberry Pi. The Raspberry Pi and DC motor were housed in a 3D-printed protective shell to allow attachment and protect the circuits of the Raspberry Pi and DC motor (Figure 4B).

To obtain images of moving mosquitoes, we confined live mosquitoes to the box and used a rotating rod to disturb the entire system at a constant rate. The rotating rod produced seven knocks on the system for approximately 10 s, rested for 10 s, and repeated the cycle (Figure 4B; Supplementary Figure S5A and Supplementary Movie S2). This disturbance ensured that a mosquito would fly continuously inside the 3D printed box, as shown in Figure 5. The camera then captured an image similar to that captured by a mosquito trap when a live mosquito flies over the capture area (Figure 5B, Supplementary Movie S3). When the interference stopped, the mosquito landed on the box wall, and the captured image reflected the state at which the mosquito had landed after flying (Supplementary Figures S5C,D, Supplementary Movie S4, and Supplementary Movie S5).

We then fed the video feed into the previously mentioned motion detection algorithm to obtain subimages. Since the subimages sometimes included unusable images, they were manually classified based on mosquito presence or absence. Accordingly, we were able to quickly obtain valuable data.

To identify different kinds of mosquitoes, SqueezeNet was selected as our neural network model (Figure 3), and we further improved it by applying the transfer learning technique based on the mosquito datasets for both static specimens (Supplementary Figure S4) and live mosquitoes (Figures 4C, D), including Ae. aegypti and Cx. quinquefasciatus, that we obtained. The training, validation, and testing set splits are listed in Supplementary Table S1 and Supplementary Table S2. The confusion matrix of the testing set is listed in Table 1. We also tested the core device of the mosquito trap to identify and classify Cx. quinquefasciatus and Ae. aegypti mosquitoes (Supplementary Figure S6 and Supplementary Movie S6). Each test consisted of a manual feeding of 10 Cx. quinquefasciatus and 10 Ae. aegypti mosquitoes into the detection zone of the core device to observe how the mosquito trap behaved. Three of the six tests test used dead mosquitoes (Table 2, test 1–3), while the remaining three tests used live mosquitoes (Table 2, test 4–6).

After confirming that the recognition result was correct, the core device was placed in an insect cage (60 cm × 60 cm × 120 cm; BD6S620, MegaView Science Co. Ltd., Taiwan), and the mosquito lure was used to test live mosquito trapping efficiency. An air-conditioning system kept the environment at a constant temperature (27°C), and the mosquito trap and two cameras (C920E, Logitech Co., Ltd., Taiwan) were placed in the center of the insect cage, as shown in Figure 5A. The two cameras were placed on the left and right sides of the mosquito trap (Figures 5B, C). Then, the zipper of the insect cage was closed, and the camera transmission line and power cord were fed through. Before releasing the mosquitoes, the mosquito trap was activated and the cameras were turned on. We confirmed that the recording screen could be visualized properly and that the recorded images could be stored on the computer’s hard disk (Figure 5A). After turning on the recording device, 50 mosquitoes were released into the insect cage through open cuffs, which were then closed once the mosquitoes entered the cage. Mosquitoes were able to fly freely in the narrow space and were caught in the mosquito trap. The camera recorded the mosquitoes’ flight status and mosquito trap status and classified the images as right video images (Figure 5D, Supplementary Movie S7) and left video images (Figure 5E, Supplementary Movie S8). After each capture, capture data were recorded, and the capture rate and recognition rate were calculated based on statistical data. In addition, a smart mosquito trap integrated with a BG trap was placed in the insect cage (with an air conditioning system room, Supplementary Figure S7) and the small living room of a simulated house (3.1 m × 2.3 m × 3 m, without furniture and air conditioning, Supplementary Figure S8) to perform the abovementioned capture rate and recognition rate tests.

Statistical analysis was performed using GraphPad Prism software (version 5.0). Use standard errors for statistical validation when testing catch and recognition rate. We analyzed the lure or temperature for capture and recognition rates using Mann‒Whitney-Wilcoxon paired t tests. A p-value less than 0.05 was considered statistically significant.

Supplementary Figure S6 and the associated recorded video from the experiment (Supplementary Movie S6) show the mosquito trap core device used to identify and classify Cx. quinquefasciatus (left container) and Ae. aegypti (right container) mosquitoes. The mosquito is manually herded to the detection zone of the core device directly from the top, as indicated by the red arrow in Supplementary Figure S6A. A mosquito flying into the detection zone activates the system; then, the capture plate rotates and drives the mosquito into the proper storage container. For example, in this study, Cx. quinquefasciatus and Ae. aegypti were captured in the left and right collection containers, respectively (Figure 6). According to our experimental data, mosquitoes constantly approached the trap during its operation, and the trap captured the mosquitoes successfully.

Regarding the recognition accuracy, SqueezeNet was applied to 1) train a model that can recognize mosquito types (Supplementary Figure S3 and Supplementary Movie S1) and 2) test the function of mosquito recognition under static and live mosquito conditions (Supplementary Table S1 and Supplementary Table S2). Our experimental data indicated that the real-time recognition accuracy rates were more than 91.57% and 89.29% in the test set images (Table 1) and live Cx. quinquefasciatus and Ae. aegypti system (Supplementary Figures S5C,D), respectively.

The intermediate deep neural networks required for the procedure introduced in the above section were developed and trained. First, the AlexNet classifier network for the mosquito contour classification shown in Figure 3 was trained and examined. The data reported that the developed neural network achieved approximately 91.57% accuracy in predicting the test set images (Table 1). Our system performed real-time classification of live Cx. quinquefasciatus or Ae. aegypti mosquitoes that were manually fed into the detection zone of the core device (Supplementary Figure S6 and Supplementary Movie S6). When integrated into a mosquito trap, our capture system can be used to successfully distinguish Cx. quinquefasciatus (Supplementary Figures S6A–C) and Ae. aegypti (Supplementary Figures S6D–F) mosquitoes in vivo and sorted into the proper chamber after identification.

To further test our mosquito trap system, we conduct the tests shown in Table 2. In these tests, the overall accuracy of mosquito recognition was 95 ± 7.6% (57/60), and the capture mechanism had a success rate of 90 ± 8.2% (27/30) in the tests with live mosquitoes. The capture and identification tests involving the core device of the prototype trap used real-time capture and classification video data to determine the effectiveness of the mosquito trap. The camera fully recorded the statuses of the mosquitoes and mosquito traps. Approximately ten experiments revealed that the accuracy of mosquito classification was more than 92 ± 3.4%. Although the capture rate was not very good, it was still nearly 44 ± 9.6% (Figure 5; Table 3). We also found that the number of lures seemed to have a slight effect on the average catch number, but the difference was not statistically significant (3 vs. 1 lure, 48% vs. 39%, p = 2.79, Mann-Whitney test).

Based on the good catch rate of the BG trap (Johnson et al., 2012; Barrera et al., 2013; Li et al., 2016), we integrated the core device of the mosquito trap into a commercial BG trap to form a new type of smart mosquito trap and then conducted tests related to mosquito trapping (Supplementary Figure S7). In a laboratory where the temperature was controlled at 27°C ± 1°C, the newly integrated smart mosquito trap was placed in a small insect cage for testing. The results showed that the average catch rate and recognition rate increased by 67 ± 25.0% and 97 ± 2.9%, respectively (Table 4). Afterward, we expanded the space in which the mosquitoes could fly by 50-fold by placing the trap in a simulated living room without air conditioning (fluctuations of 27–33°C) and conducted additional mosquito trap tests (Supplementary Figure S8). Similar results were obtained, with an average capture rate of 49 ± 17.7%; the recognition rate reached 90 ± 8.3% (Supplementary Table S3).

The user connects to the project website (Figure 6), and all the data are transmitted to the user on the backend, including the area where the mosquitoes are caught (Figures 6A, C), the mosquito trap number, address, GPS coordinates (Figure 6B), date and time, type of mosquito vector, temperature, humidity, etc. The user accesses the data via a client web page; weather data are constantly pulled from the Central Weather Bureau (CWB) via an open application programming interface (API) connection.