AI Turns Ordinary Smartphones Into Mosquito Egg Counters for Dengue Surveillance

Aedes aegypti is one of the most consequential insects on the planet. The mosquito, which has spread far beyond its African origins to colonize tropical and subtropical cities across the globe, is the primary vector for dengue, Zika, chikungunya, and yellow fever viruses. As urbanization and climate shifts expand its range, public health agencies face a persistent problem: knowing where the mosquito is breeding, and in what numbers, before outbreaks erupt. The standard surveillance instrument for this task is the ovitrap, a simple container holding an oviposition substrate on which female mosquitoes lay their eggs. Ovitraps are cheap, effective, and easy to deploy, but they generate a hidden cost that has long limited their use at scale. Every substrate that comes back from the field must be examined, and every egg on it must be counted, almost always by a trained human eye. A new study published in Parasites & Vectors by Juan Almanza and Diego Montenegro of Fundación Chilloa in Santa Marta, Colombia, presents a system designed to break that bottleneck, and it does so using nothing more exotic than a smartphone and a neural network.

The system, called Col-Ovo, is an artificial intelligence tool that automatically detects and counts Aedes aegypti eggs on photographs of oviposition substrates taken under real field conditions. Its development grew directly out of an operational surveillance program rather than a laboratory exercise. The researchers trained their model on 275 oviposition substrates, each a strip roughly 20.5 centimeters long, collected through routine monitoring activities run in partnership with the Vector-Borne Diseases Surveillance and Control Program of Santa Marta. That provenance matters. Laboratory image datasets are typically clean, well-lit, and carefully framed, which makes them poor proxies for what health workers actually encounter. Field substrates arrive stained, dirty, and inconsistently photographed, and any counting system that cannot cope with that messiness will fail the moment it leaves the bench.

The technical heart of Col-Ovo is a YOLOv11m object-detection model. YOLO, short for You Only Look Once, is a family of convolutional neural network architectures that performs object detection in a single forward pass through the network, rather than the multi-stage proposals used by older detection pipelines. That design choice makes YOLO models fast enough to run near real-time inference, a property the Colombian team exploited to reduce the processing time for a single substrate from approximately fifteen minutes of expert manual counting to less than three seconds of automated analysis. The model was trained on smartphone images captured in situ, with no preprocessing, no standardized lighting rig, and no laboratory-grade camera equipment. The images included substrates treated with bioattractants, specifically blackstrap molasses and the yeast Saccharomyces cerevisiae, which are used in the low-cost OviCol strategy to make ovitraps more attractive to egg-laying females. Those attractants stain the substrate darkly, and dark stains are precisely the kind of visual noise that confounds naive computer vision approaches, since mosquito eggs are themselves small, dark, elongated objects against a darkening background.

Col-Ovo was also deliberately engineered to tolerate a constraint that most academic computer vision papers ignore: image compression from messaging platforms. In community-based surveillance programs across Latin America, the practical workflow often involves local volunteers photographing substrates in their neighborhoods and sending the images to a central coordinator over WhatsApp or a similar application. Messaging services aggressively compress images to save bandwidth, destroying fine detail and introducing blocky artifacts that can obliterate the subtle texture differences between an egg and a speck of sand. By training and evaluating on images subjected to this kind of compression, the researchers built a system that fits the actual communication infrastructure of the communities it is meant to serve, rather than demanding a workflow the communities would have to build around it. The substrates in the dataset also contained sand and particulate debris, further stress-testing the detector’s ability to distinguish genuine eggs from lookalike clutter.

To judge whether the algorithm could be trusted, the researchers benchmarked it against two kinds of human evaluators. The first was expert manual counting, the traditional gold standard performed by trained technicians. The second was a virtual-human count conducted through the OviLab platform, in which evaluators examined images at greater than 200 percent magnification to resolve eggs that would be difficult to see at native resolution. Agreement between Col-Ovo and the human references was assessed using correlation and concordance analyses, the statistical tools of choice for asking whether two measurement methods agree closely enough to be interchangeable. The results were strong. Col-Ovo achieved a Spearman rank correlation of 0.91 with human evaluators, indicating that when the system reports a high count, humans report a high count too, and the ordering of substrates by infestation level is preserved. Lin’s concordance correlation coefficient, which penalizes deviations from the identity line and not merely correlation, came in at 0.89, showing that the automated counts were not just proportional to human counts but closely matched them in magnitude.

The absolute error figures tell a similarly encouraging story. The mean absolute error between Col-Ovo and human counts was 22.1 eggs per substrate. On substrates carrying hundreds of eggs, an error of that size is operationally negligible for surveillance purposes, where the goal is to rank neighborhoods and time periods by mosquito abundance rather than to certify exact totals. Perhaps most tellingly, when the researchers compared paired Col-Ovo counts against OviLab virtual-human counts, the difference between the two methods was not statistically significant, with a P value of 0.252. In practical terms, the algorithm behaved like another trained counter in the room, one that never fatigues, never needs retraining on visual recognition, and works three hundred times faster than its human colleagues.

The performance held up under the harshest conditions the field could supply. Substrates that were heavily stained by the molasses and yeast bioattractants, and those contaminated with sand and debris, were exactly the samples where manual counting is slowest and most error-prone, and they were precisely the samples Col-Ovo was built to handle. The system successfully detected and quantified eggs across this spectrum of difficulty, and its accuracy remained robust when the input images had been compressed by messaging applications. That combination of robustness and speed transforms the economics of ovitrap surveillance. A program processing a thousand substrates would previously have faced roughly 250 hours of expert counting labor; with Col-Ovo, the same workload requires under an hour of machine time, freeing technicians for the control interventions that actually reduce mosquito populations.

The broader significance of the work lies in how it fits into an emerging digital infrastructure for community-led vector control. Col-Ovo is designed to integrate with OviLab, the platform through which virtual-human counts are performed, creating a flexible pipeline in which routine smartphone photographs flow into automated analysis, with human review reserved for ambiguous or high-stakes cases. The OviCol strategy that underpins the surveillance effort combines monitoring with control in a single low-cost ovitrap design, making it feasible for neighborhoods to run their own mosquito programs. What has been missing is a way to process the resulting flood of substrates without a small army of counters. By eliminating the need for specialized equipment, standardized imaging setups, or any image preprocessing, Col-Ovo addresses what the authors identify as a critical barrier to scaling ovitrap surveillance to the thousands of deployment sites a city-wide program requires.

The study was conducted within the legal mandates and public health competencies of the Santa Marta vector control authorities, and the authors disclose that they were involved in developing both Col-Ovo and OviLab, a relationship declared for transparency. The work was supported by Fundación Chilloa and by a postdoctoral grant from MinCiencias-Colombia. Those details matter because the tool’s credibility rests on having been forged in operational conditions rather than demonstrated in a curated dataset. If systems like Col-Ovo can be validated and deployed across other endemic regions, the implications extend beyond dengue to Zika, chikungunya, and yellow fever, all of which ride on the same vector. Rapid, accurate, and scalable egg counting turns ovitraps from a labor-intensive research instrument into a genuine early-warning network, one in which a volunteer with a phone in a Santa Marta backyard can contribute a data point to a city-wide picture of mosquito activity within seconds of pressing the shutter. In the ongoing contest between urban humans and Aedes aegypti, that kind of speed may prove decisive.

Subject of Research: Smartphone-based artificial intelligence for automated counting of Aedes aegypti eggs in ovitrap-based vector surveillance

Article Title: Col-Ovo: Smartphone-based artificial intelligence for rapid counting of Aedes mosquito eggs under field conditions

Article References: Col-Ovo: Smartphone-based artificial intelligence for rapid counting of Aedes mosquito eggs under field conditions. (n.d.). https://doi.org/10.1186/s13071-026-07734-1

Image Credits: AI Generated

DOI: 10.1186/s13071-026-07734-1

Keywords: Aedes aegypti, Col-Ovo, artificial intelligence, object detection, YOLOv11m, ovitraps, dengue, Zika, chikungunya, vector surveillance, computer vision, smartphone imaging