AI Takes on Mosquitoes: How Machine Learning Is Reshaping Disease Surveillance

Mosquito-borne diseases such as malaria, dengue, Zika, and chikungunya remain among the most stubborn threats to global health, and a new review argues that artificial intelligence is rapidly changing how scientists find, count, and outmaneuver the insects that carry them. Published in BMC Artificial Intelligence, the paper synthesizes recent advances across three broad domains: mosquito biology and identification, ecological and disease modeling, and mosquito and disease control, while also flagging the data and deployment gaps that still stand between laboratory promise and field reality.

The scale of the problem is considerable. According to World Health Organization figures cited in the review, mosquito-borne diseases kill more than 700,000 people annually and account for 17 percent of the global infectious disease burden, leaving roughly 80 percent of the world’s population at risk. Dengue alone reached an estimated 14 million cases in 2024, with mortality exceeding historical records. With no widely available vaccines for many of these diseases and traditional mosquito management remaining labor-intensive and costly, the authors argue that more efficient control strategies are urgently needed.

The most mature application is automated species identification. Distinguishing medically important vectors such as Aedes aegypti and Anopheles gambiae enables targeted interventions, and deep learning has now been applied to whole-body images, wing morphology, and wingbeat audio recordings across mosquito life stages. Convolutional neural networks remain foundational for feature extraction, while YOLO architectures have proven faster than standard CNNs at localizing mosquito features in cluttered backgrounds, enabling real-time detection. More recently, transformer-based models have surpassed both traditional machine learning and CNNs in accuracy: a Swin Transformer achieved a benchmark accuracy of 99.04 percent and an F1-score of 99.16 percent across 17 species, outperforming ResNet34, which reached 98 percent precision and a 96 percent ROC-AUC. Hybrid CNN-transformer designs and interpretability tools such as Grad-CAM, which visualizes the image regions driving a prediction, are pushing the field toward models that are both accurate and transparent.

Identification is no longer limited to adults. Single-shot detector networks have reached 94.19 percent accuracy for Aedes species larvae, and CNN-based genus-level models perform comparably, with recent approaches using smartphone cameras fitted with specialized macro lenses to capture the high-resolution larval images needed for classification. The review cautions, however, that performance comparisons across studies must be interpreted carefully, because task definition, class number, image quality, and whether specimens were laboratory-reared or field-collected all strongly influence reported results. Models trained on larger, field-collected datasets generally reflect more realistic deployment conditions but often show lower or more variable performance than those tested in controlled laboratory settings.

Beyond naming species, AI is being used to read the biology of the mosquitoes themselves. Using MALDI-TOF mass spectrometry of protein patterns from mosquito thoraxes, one research team applied artificial neural networks and multilayer perceptrons to predict the age, blood-feeding status, and malaria infection of field-caught mosquitoes. Another group trained machine learning and deep learning models on mid-infrared spectra of laboratory-reared mosquitoes and achieved 92 percent and 99 percent accuracy, respectively, in predicting the ages of field-caught individuals. A separate study showed that MALDI-TOF spectra paired with CNNs can predict Wolbachia infection in Aedes aegypti in the laboratory, offering a scalable framework for tracking the prevalence of this biological control agent, while mid-infrared spectroscopy combined with supervised machine learning accurately predicted the blood-meal source of wild-caught female Anopheles. These approaches are faster, cheaper, and less labor-intensive than traditional methods such as PCR, ELISA, and mosquito dissection.

Behavioral analysis is following a similar trajectory. Support vector machines applied to three-dimensional flight trajectories distinguished sexes within mating swarms with ROC-AUC scores of 0.701 for stereo 3D, 0.688 for telecentric, and 0.632 for 2D data. A Mask R-CNN system detected, classified, and counted both flying and resting mosquitoes with 96 to 100 percent accuracy, and a follow-up CNN-based system, FlightTrackAI, tracked flight paths with 99.9 percent accuracy even when mosquitoes were partially or fully occluded. Another CNN-based tool, biteOscope, uses AI-enhanced video analysis to monitor blood-feeding events at fine temporal scales, providing insight into host-mosquito interactions and feeding preferences.

At the landscape scale, AI is merging remote sensing, ensemble learning, and aerial robotics into surveillance networks. In Greece, a multilayer perceptron trained on Landsat 8-derived vegetation indices, land surface temperature, and water indices forecasted peak wild mosquito populations up to 13 days in advance. In Morocco, XGBoost, Gradient Boosting, and Random Forest models trained on Anopheles and Culex presence-absence data alongside 225 environmental predictors reached accuracies of 74.5 to 86.2 percent, with XGBoost performing best. Convolutional neural networks applied to UAV imagery have accurately detected Anopheles larval habitats in Burkina Faso and Côte d’Ivoire, and in Campinas, Brazil, deep learning models combined with aerial imagery identified oviposition sites such as rooftop water tanks and swimming pools while assessing neighborhood socio-economic indices. Robotic platforms are joining the effort: a system called Dragonfly pairs a mosquito-trapping module with a YOLOv4 classifier, achieving 88 percent accuracy on public digital photos and 82 percent in field trials with live video, while an UAV equipped with a multispectral sensor and an AlexNet architecture identified mosquito larvae in the surface waters of California tidal marshes. Faster R-CNN applied to ovitrap imagery has automated egg detection with precision, recall, and F1 scores near 0.98, and Sentinel-1 SAR surface-water mapping has reached accuracies up to 0.948 in identifying potential larval habitats.

On the disease side, the review traces an evolution from rule-based expert systems to data-driven deep learning. Early expert systems, which replicated expert diagnostic reasoning using predefined rules, achieved diagnostic accuracies exceeding 90 percent for mosquito-borne diseases and usability scores as high as 83, though their dependence on domain experts limited scalability. Support vector machines and Bayesian classifiers subsequently delivered consistently high detection performance even with limited training data, and CNNs have now achieved accuracies exceeding 90 to 100 percent in detecting malaria-infected cells in microscopic blood images, with hybrid CNN and Random Forest architectures improving diagnostics further. For outbreak prediction, Random Forests and XGBoost have forecast West Nile virus, dengue, and malaria outbreaks while identifying temperature, rainfall, relative humidity, and vegetation indices as key drivers, and Long Short-Term Memory networks, which capture sequential dependencies in time-series data, often outperform Random Forests and support vector machines in dengue forecasting, particularly when combined with attention mechanisms. Explainable AI tools such as SHAP are increasingly used to open the black box: in West Nile virus prediction, SHAP revealed that winter climate and the normalized difference water index were dominant predictors, while in dengue risk mapping it highlighted the influence of landscape and demographic factors on spatial case distribution.

These capabilities are beginning to feed directly into Integrated Mosquito Management programs. The Greek early warning model enables timely insecticide applications, an app built on real-time habitat monitoring generates heat maps of Anopheles larval density for immediate larvicide targeting, and a risk mapping effort along India’s Buckingham Canal converted model outputs into a standardized scale that let local government prioritize aquatic zones for chemical treatment. An ontology-based system called AMAS, deployed across 20 localities in Cameroon, combines environmental data, behavioral data, and expert input through interacting organizational and analytical agents to support continuous mosquito surveillance and contextualized malaria-risk assessment. Deep learning is also extending into citizen science, with CNNs tracking mosquito species, sex, and abdominal status from images submitted by community members, including photographs of mosquitoes on human skin that offer real-time insight into human-mosquito exposure risk.

Substantial obstacles remain. Most deep learning models still rely on restricted datasets dominated by dorsal-only imaging, a handful of species, and adult stages, and models trained on laboratory-reared mosquitoes suffer sharp performance drops when confronted with morphologically diverse wild populations, varied terrain, unpredictable lighting, and clutter. Audio- and video-based methods demand high-quality recordings and noise management, spectral approaches require costly specialized instruments, and UAV and robotic systems face aviation regulations, battery limitations, and adverse weather. Many surveillance networks are validated only on test data from the same sites and conditions as their training data, leaving them unverified in new ecological or geographic settings. Citizen science platforms carry biases, validation challenges, and privacy concerns, and public trust and perceived disease risk strongly determine whether UAV-based surveillance is accepted. The authors call for large, multimodal, geographically diverse training repositories, self-supervised and transfer learning to help underrepresented species and life stages, low-power IoT devices and smart traps for resource-constrained settings, and continued investment in explainable AI frameworks such as SHAP and Grad-CAM to build the administrative trust needed to turn these technologies into equitable, scalable tools for global mosquito-borne disease control.

Subject of Research: Applications of artificial intelligence in mosquito surveillance and mosquito-borne disease research

Article Title: Bridging technology and epidemiology: artificial intelligence-driven innovations in mosquito and mosquito-borne disease research

Article References: Pacheco, P. R. D. C., Hila, A. M. J., Francisco, J. C., Francisco, M. E., & Carvajal, T. M. (2026). Bridging technology and epidemiology: artificial intelligence-driven innovations in mosquito and mosquito-borne disease research. BMC Artificial Intelligence, 2(1), Article 19. https://doi.org/10.1186/s44398-026-00039-4

Image Credits: AI Generated

DOI: 10.1186/s44398-026-00039-4

Keywords: artificial intelligence, mosquito-borne disease, deep learning, mosquito surveillance, dengue, malaria, vector control, machine learning, remote sensing, early warning systems, explainable AI, citizen science