MOSKITA: Multi-Domain Research Systems

• MOSKITA is a research label designating independent systems for mosquito surveillance and collider-based low-energy particle detection.
• The mosquito systems employ computer vision and acoustic ML to classify specimens from smartphone images and wingbeat sounds, achieving accuracies up to 86–100%.
• The collider instrument uses a skipper-CCD sensor with sub-electron readout to probe low-energy particles, setting new constraints in particle physics.

MOSKITA is a research name used for multiple, technically unrelated systems spanning mosquito surveillance and collider instrumentation. In the mosquito-technology literature, it designates both a computer-vision pipeline for classifying trapped vector specimens from smartphone images and a smartphone-oriented acoustic architecture for detecting Aedes aegypti from wingbeat sounds. In particle physics, it denotes a skipper-CCD detector installed near the CMS interaction point to search for low-energy particles produced in high-energy collisions (Minakshi et al., 2020, Paim et al., 2023, Cervantes-Vergara et al., 8 Aug 2025).

1. Terminological scope

Within the literature surveyed here, MOSKITA is not a single platform but a reused label attached to distinct research programs. The meanings diverge in scientific objective, sensing modality, and deployment environment.

MOSKITA usage	Domain	Core modality
Automated surveillance of trapped mosquito vectors	Computer vision	Smartphone images of specimens
Real-time Ae. aegypti detection and tracking	Acoustic ML on smartphones	Wing-beat sound spectrograms
Low-energy particle search near CMS	Collider instrumentation	Skipper-CCD ionization readout

The trapped-specimen system is presented as a method for genus and species classification of mosquitoes collected in field traps, using smartphone-acquired images processed by a CNN (Minakshi et al., 2020). The acoustic system is specified as a smartphone application for live mosquito detection from wing-beat sound, with explicit requirements for low-latency execution, noise robustness, and mobile deployment (Paim et al., 2023). The collider detector is a stainless-steel T-shaped vacuum vessel housing a skipper-CCD sensor, deployed about $33$ m from the CMS collision point and shielded by rock, with the purpose of probing beam-related low-energy ionization events (Cervantes-Vergara et al., 8 Aug 2025).

A recurring source of confusion is therefore lexical rather than technical: identical nomenclature does not imply shared hardware, software, or scientific lineage. Context is essential for interpretation.

2. Image-based MOSKITA for trapped-specimen surveillance

The computer-vision MOSKITA system addresses taxonomic surveillance of trapped disease vectors. Its dataset was assembled from Hillsborough County, Florida, where expert taxonomists selected 250 female specimens spanning nine vector species across the three genera Aedes, Anopheles, and Culex. Each specimen was imaged under normal indoor lighting on flat white, pink, or cream tile surfaces using a smartphone camera mounted about $1$ ft above the insect, with three camera orientations. This produced 6,807 raw images across 10 phone models; augmentation by zoom and brightness/contrast variation yielded a training corpus of 25,867 images. Before training, each image underwent non-local means denoising with a $7\times7$ patch and $h=10$, was resized to $299\times299$ pixels, and normalized by dividing each channel by 255 (Minakshi et al., 2020).

The classifier is built around a fine-tuned Inception-ResNet-V2 backbone pretrained on ImageNet. The architecture was adapted for three tasks: genus classification, species-within-genus classification, and direct nine-way species classification. For the genus and species-only heads, the IRV2 base was cut at its 433rd layer, followed by Global Average Pooling, four fully connected ReLU layers, concatenation of those outputs, and a final softmax layer. Training proceeded in two phases: an initial frozen-backbone stage with randomly initialized dense layers using Glorot uniform, followed by end-to-end fine-tuning. The optimizer was Adam with $\beta_1=0.89$ and $\beta_2=0.999$, the loss was categorical cross-entropy, the learning-rate schedule was cyclic triangular between $2\times10^{-7}$ and $2\times10^{-5}$, the batch size was 32, and regularization combined dropout, batch normalization, and early stopping (Minakshi et al., 2020).

Performance is reported at several levels. The species-only model achieved about 80% accuracy on the full augmented dataset and 81%–92% accuracy on unseen test sets of 600 three-image specimen sets. In the nine-way species head, validation accuracy reached 86% for Aedes aegypti and 100% for Anopheles stephensi, while other classes ranged from 63% to 93%. On an independent test of 50 new specimens, set-level species-only accuracy was about 81% for Aedes, about 77% for Anopheles, and about 92% for Culex. Feature-map visualization indicated that the network focuses on the thorax, scutum, wings, abdomen, and legs, aligning the learned representation with anatomical landmarks used by human taxonomists. The paper also describes a practical field rig based on a smartphone over a matte tile, with server-side inference and dashboard reporting; a single NVIDIA T4 is stated to process about 100 images/s, with per-mosquito decisions in about 100 ms when images are grouped into specimen triples (Minakshi et al., 2020).

The principal limitations are explicitly stated: the dataset is restricted to nine female vector species from one U.S. county, requires controlled flat backgrounds and still specimens, and excludes larvae and males. Future work includes broader geography, additional taxa and life stages, in-trap imaging, and integration of acoustic wingbeat data (Minakshi et al., 2020).

3. Acoustic MOSKITA for smartphone detection of Aedes aegypti

The acoustic MOSKITA design is a smartphone application for real-time detection of Ae. aegypti from wing-beat sound. Its requirements are formulated as performance and efficiency on off-the-shelf smartphones, robustness to uncontrolled background noise, and resilience against malicious or spurious noise. The signal chain begins with capture on a built-in mono smartphone microphone, low-pass filtering at $4$ kHz, down-sampling to $1$0 kHz, and segmentation into overlapping windows of $1$1 s with 50% overlap. Feature extraction uses an STFT with $1$2 and hop length $1$3, followed by decibel conversion, clipping to $1$4, and normalization to $1$5 (Paim et al., 2023).

The classifier is a lightweight residual CNN. Its input is a single-channel spectrogram image. The main body contains $1$6 arrays, each formed by a residual block with two $1$7 convolutions using $1$8 filters, batch normalization, ReLU activation, a skip connection of the form $1$9, then max-pooling and dropout at rate $7\times7$0. The classification block flattens the representation, applies a dense layer of size 256 with ReLU, and ends with a two-unit softmax output for $7\times7$1. Training uses binary cross-entropy, Adam with learning rate about $7\times7$2, batch size 32, and about 50–100 epochs with early stopping (Paim et al., 2023).

The main benchmark configuration, dataset D5, combines 537 positive clips of Ae. aegypti from female and male sources with 2,158 negative clips from ambient noise and other insects. After down-sampling and segmentation, this yields about 2,695 spectrograms. Under stratified 10-fold cross-validation, D5 produced about 87% $7\times7$3 accuracy, about 88% $7\times7$4 precision, about 88% $7\times7$5 recall, and about 88% $7\times7$6. Easier settings exceed 93%. The model contains 927,458 parameters; after dynamic-range quantization it becomes about twice as fast, the model file is about 4 MB, inference latency is about 320 ms per 60 s frame on proof-of-concept devices, and battery consumption is reported as less than 5% per hour of continuous detection (Paim et al., 2023).

The acoustic MOSKITA work sits within a broader mosquito-acoustics ecosystem. HumBugDB provides 20 hours of finely labeled mosquito flight sounds, 18 hours of which contain annotations from 36 species, together with metadata in PostgreSQL and Bayesian CNN baselines for mosquito event detection and species classification (Kiskin et al., 2021). Complementarily, 3D audio-visual recordings of Culex quinquefasciatus have combined simultaneous slow-motion video at 20 kfps with rotating 12-microphone-array recordings, using vein-tracked wing deformation, immersed-boundary CFD, the Ffowcs-Williams–Hawkings formulation, and physics-based ICA to reconstruct and simulate wing-tone directivity (Feugère et al., 2023). This suggests a link between mobile acoustic classification and mechanistic aeroacoustic modeling, although the papers treat those problems separately.

4. MOSKITA as a skipper-CCD detector at the LHC

In high-energy physics, MOSKITA is a skipper-CCD detector deployed near CMS for the study of low-energy particle remnants from LHC collisions. The apparatus is a stainless-steel T-shaped vacuum vessel at $7\times7$7 Torr inside a 2-inch lead shield, mounted on a 1.2 m$7\times7$8 aluminum pallet in the CMS drainage gallery about 70 m underground, at azimuthal angle $7\times7$9 and pseudorapidity $h=10$0. The sensor sits about 33 m from the interaction point, behind about 17 m of rock, and is oriented to intercept low-energy byproducts of collisions (Cervantes-Vergara et al., 8 Aug 2025).

The sensing element is a 6.29 MPix p-channel skipper CCD with active array $h=10$1, pixel pitch $h=10$2, thickness $h=10$3, and active mass about 2.2 g. Readout uses four floating-gate amplifiers, although only two quadrants were usable because of a charge-transfer issue in one serial register. The detector operated with $h=10$4 or 256 non-destructive measurements per pixel. For runs 4–6, the readout noise was $h=10$5–0.23 $h=10$6, the gain was about 226 ADU/$h=10$7, and the single-electron rate was about $h=10$8–$h=10$9. A 1 $299\times299$0 detection threshold was selected per image by modeling the $299\times299$1 and $299\times299$2 pixel distributions as Gaussians and maximizing the $299\times299$3 score; typical thresholds were about $299\times299$4–$299\times299$5 (Cervantes-Vergara et al., 8 Aug 2025).

Data were taken from March 15 to December 31, 2024. During the proton-proton period at $299\times299$6 TeV, MOSKITA images corresponded to 113.3 fb$299\times299$7; during Pb–Pb running at $299\times299$8 TeV, MOSKITA used 1.54 nb$299\times299$9. Images are classified by whether the integrated luminosity assigned over their readout time is nonzero. The low-energy region is defined as total event charge $\beta_1=0.89$0, and the high-energy region as $\beta_1=0.89$1. Events are reconstructed as contiguous pixel clusters above threshold after baseline subtraction and gain calibration, with multiple masks applied image-by-image or per run, including border, hot-column, transfer-gate-trap, serial-register-event, crosstalk, bleeding-zone, halo, and hot-zone masks. For low-energy analysis, single-row events are additionally rejected, and efficiencies for $\beta_1=0.89$2–$\beta_1=0.89$3 are obtained from Monte Carlo simulations of 50k images per setting (Cervantes-Vergara et al., 8 Aug 2025).

The low-energy statistical analysis uses a Poisson likelihood with signal term $\beta_1=0.89$4 and background term $\beta_1=0.89$5, profiled over $\beta_1=0.89$6. In proton-proton data, the $\beta_1=0.89$7–$\beta_1=0.89$8 combined region gives $\beta_1=0.89$9, and the 95% C.L. upper limit in that interval is 7.8 events. In Pb–Pb data, the 7 $\beta_2=0.999$0 bin shows $\beta_2=0.999$1 locally, but the look-elsewhere probability is about 0.29; the combined $\beta_2=0.999$2–$\beta_2=0.999$3 interval gives $\beta_2=0.999$4. The combined proton-proton plus Pb–Pb analysis yields $\beta_2=0.999$5 in $\beta_2=0.999$6–$\beta_2=0.999$7, which approaches $\beta_2=0.999$8 but remains compatible with background. In the high-energy region, spectra show Cu $\beta_2=0.999$9 and $2\times10^{-7}$0 peaks and a $2\times10^{-7}$1-bump at about 220 keV; a low-energy excess after Pb–Pb coincides with a rise in single-electron rate and is suggested to have a common origin, perhaps neutron-induced lattice defects (Cervantes-Vergara et al., 8 Aug 2025).

Using the low-energy proton-proton result, MOSKITA sets 95% C.L. constraints on the mass-millicharge parameter space for millicharged particles. The paper characterizes these limits as weaker than milliQan and reactor or beam-dump constraints because of the small detector mass, but also as the first collider-based skipper-CCD limit in this mass range. It recommends more massive CCD arrays, improved passive shielding, coincidence or veto layers, and further study of post-Pb–Pb neutron activation and charge-trap effects (Cervantes-Vergara et al., 8 Aug 2025).

5. Methodological commonalities across the MOSKITA literature

Although the MOSKITA systems are domain-specific and not architecturally unified, the literature suggests a common methodological pattern centered on extracting weak signals under deployment constraints. In the image-based surveillance system, weak inter-species visual differences are stabilized by denoising, augmentation, transfer learning, and multi-image decision rules (Minakshi et al., 2020). In the acoustic system, classification depends on low-pass filtering, spectrogram normalization, explicit inclusion of environmental-noise negatives, and compact residual blocks that remain executable on commodity smartphones (Paim et al., 2023). In the collider detector, the same general problem appears in another guise: low-energy ionization candidates are separated from readout artifacts and ambient backgrounds by threshold optimization, beam-on versus beam-off comparison, masking, Monte Carlo efficiency estimation, and profile-likelihood inference (Cervantes-Vergara et al., 8 Aug 2025).

A second commonality is the use of physically or behaviorally structured acquisition. Related mosquito-behavior platforms such as biteOscope employ a transparent $2\times10^{-7}$2 cm cage, controlled heat and CO$2\times10^{-7}$3, and DeepLabCut tracking of 38 keypoints to derive high-resolution feeding metrics, including a mean of $2\times10^{-7}$4 biting events per mosquito in 5 min and mean bite duration of 4.8 s (Murray et al., 2021). Likewise, the aeroacoustic wing-tone study reconstructs 4D wing kinematics from 11 landmarks per wing and combines immersed-boundary CFD with Ffowcs-Williams–Hawkings acoustics and physics-based ICA to match measured directivity patterns (Feugère et al., 2023). A plausible implication is that MOSKITA-labelled systems belong to a wider technical current in which acquisition geometry, preprocessing, and inference are tightly co-designed.

A third commonality is scale-aware engineering. The trapped-specimen MOSKITA assumes smartphone cameras and cloud or edge inference (Minakshi et al., 2020); the acoustic MOSKITA is explicitly optimized for Android-class hardware, quantized deployment, and low battery draw (Paim et al., 2023); the collider MOSKITA demonstrates that sub-electron skipper-CCD readout can operate stably in a high-luminosity environment with a detector mass of only about 2.2 g (Cervantes-Vergara et al., 8 Aug 2025). This suggests that, across highly different disciplines, MOSKITA has come to denote systems in which sensitivity is pursued without abandoning compact or fieldable hardware.

6. Limitations, ambiguity, and prospective development

The most immediate limitation is semantic ambiguity. Because MOSKITA names unrelated systems, citation context must specify whether the topic is image-based entomological surveillance, acoustic smartphone detection, or skipper-CCD collider instrumentation. The literature does not present these as versions of a common platform (Minakshi et al., 2020, Paim et al., 2023, Cervantes-Vergara et al., 8 Aug 2025).

The technical limitations are system-specific. The trapped-specimen vision system is restricted to nine female vector species from one county, requires controlled backgrounds and still adults, and therefore does not yet generalize to larvae, males, or unconstrained field imagery (Minakshi et al., 2020). The acoustic app identifies noise robustness and adversarial resilience as explicit requirements, and its performance degrades when classifying male wing-beat only, when ambient noise overlaps the 400–1,200 Hz bands, and on very low-end devices without optimized inference (Paim et al., 2023). The collider detector found no significant beam correlation in proton-proton or Pb–Pb running, used only two usable quadrants of the sensor, and remains limited by small mass and background systematics, including the post-Pb–Pb rise in single-electron rate (Cervantes-Vergara et al., 8 Aug 2025).

The forward trajectories are likewise distinct. For mosquito surveillance, future work includes extension to wider geographies, more taxa, in-trap imaging, and integration of acoustic wingbeat data (Minakshi et al., 2020). For acoustic detection, the stated development path includes on-device optimization, broader benchmarking, and smartphone deployment under real-world noise (Paim et al., 2023). For the LHC detector, the proposed next steps are O(100 g) or 200 g skipper-CCD deployments, improved shielding, active vetoes, and higher integrated luminosity (Cervantes-Vergara et al., 8 Aug 2025). Taken together, these directions indicate that MOSKITA research is presently constrained less by proof of principle than by scale, background control, and domain transfer.