Killing mosquitoes- with technology sprays like TinyML, CNN and acoustic sensing

twentieth August occurs to be the ‘World Mosquito Day’; and final week the World Mosquito Program (WMP) noticed the ‘World Mosquito Week’. A very good time to be reminded that mosquito-borne ailments kill multiple million folks all throughout the globe and that mosquitoes infect as much as 700 million folks every year. We don’t want any report back to consider that just about one in ten persons are affected by them- simply the irritating buzz of those mosquitoes round our houses and workplaces is sufficient to vouch for that. And are we to not blame for this nuisance- because of local weather change, globalisation, and reckless urbanisation that assist these ailments transfer throughout areas with never-before velocity and impression? 

A mosquito is greater than a sound to be swatted with a bat. Mosquito-borne viral ailments corresponding to dengue, Zika, chikungunya and yellow fever pose a hazard to billions of individuals worldwide. Currently, six areas in Europe have reported regionally acquired infections with WNV for the primary time ever (As shared by the European Centre for Illness Prevention and Management (ECDC) ). In 2025, chikungunya virus illness, transmitted by the invasive mosquito species Aedes aegypti and Aedes albopictus, has surged; and over 445,000 suspected and confirmed instances have been reported throughout 40 nations (with quickly increasing outbreaks seen within the Indian Ocean area and the Americas (As per numerous experiences by European Fee (EC), World Well being Organisation (WHO) and ECDC). Within the newest menace emergence, there have been 299 West Nile virus illness instances in 2026 (identified by the Facilities for Illness Management and Prevention (CDC)).

Are there any solutions to combat this seemingly-small menace? Sure- many methods have been tried for many years – from controlling vector-borne ailments by eradicating or treating larval habitats, avoiding stagnant water storage, to approaches like fogging or ultra-low-volume spraying with biocidal merchandise or Sterile Insect Method (SIT), or Wolbachia symbionts and gene-drive applied sciences. However making these methods cost-effective has all the time been a challenge- a barrier amplified by restricted investments in surveillance and management. Additionally, insecticide resistance in mosquitoes retains growing; whereas using pyrethroids, primarily in agriculture, has its personal ecotoxicological and human well being impact-concerns. 

Amidst this buzz, the College of Wollongong India has include an innovation. Affiliate Professor Kiran Trivedi, who has developed a low-cost AI-powered system that may strengthen mosquito surveillance and assist public well being efforts, tells us all about this answer on this interview. How can a low-cost transportable system (that makes use of Tiny Machine Studying (TinyML) to establish disease-carrying mosquito species inside seconds by analysing the sound of their wingbeats) assist to combat mosquitoes- that too, with out requiring web connectivity or laboratory infrastructure! This innovation was reportedly showcased on the ‘United Nations AI for Good International Summit’ in Geneva, the place Professor Trivedi additionally demonstrated how low-cost edge AI applied sciences may be leveraged to deal with urgent public well being challenges by way of accessible and scalable options. Can this work? At scale? With infrastructure limitations? For surveillance? In distant elements too? Past sensing and all the best way to therapy? Let’s bat some questions on this new AI-spray with Dr Kiran Trivedi, Affiliate Professor, College of Wollongong India

What triggered you in direction of this space? Any precedent work that you’ve got achieved already much like this?

The concept originated from my earlier work on sound classification utilizing TinyML, with machine studying operating instantly on a small system. Throughout the COVID interval in 2021, I used to be experimenting with TinyML and exploring whether or not a small, embedded system might be used to recognise sounds. I initially labored with Google’s bird-sound datasets and was capable of establish hen calls, together with that of a cuckoo, from recordings.

This led to an extra experiment impressed by a dialogue with my brother, who’s a paediatrician. We mentioned how a child’s cry can convey details about what the newborn wants, whether or not the newborn is hungry, sleepy or uncomfortable. I subsequently collected baby-cry recordings and developed a small TinyML system able to classifying completely different crying patterns.

These experiments established a broader curiosity in utilizing acoustic signatures to establish and interpret organic phenomena. This led me to discover whether or not an analogous method might be utilized to mosquitoes. In contrast to the sooner examples, the potential public-health relevance was significantly important on this case. In accordance with WHO information, the mosquito is the world’s deadliest animal, inflicting greater than 700,000 deaths worldwide yearly.

How complicated is the technical half right here?

The applying was additionally technically attention-grabbing as a result of completely different mosquito species have distinct wingbeat traits. I discovered the Oxford Humbug mosquito acoustic dataset, which comprises mosquito wingbeat recordings, and commenced experimenting with species corresponding to Anopheles, Culex and Aedes aegypti. To introduce larger selection into the info, we additionally synthetically generated further information. The analysis paper was submitted co-authored with my former pupil Harsh Shroff to an ITU-Kaleidoscope and international hackathon in 2021, the place it grew to become the worldwide successful challenge. Subsequently, I used to be invited as a keynote speaker and to reveal the idea on the AI for Good occasion in Geneva final month.

The mosquito challenge subsequently emerged from a broader line of labor centered on the concept sound comprises significant info and that AI may be educated to interpret these acoustic patterns. The development from birds and child cries to mosquito wingbeats was a pure extension of that analysis.

How does the answer work, particularly with out connectivity and lab gear? How sturdy and novel are the sensor parts?

The elemental precept is to establish the mosquito by way of its wingbeat relatively than bodily capturing it and transporting it to a laboratory. A microphone captures the mosquito’s acoustic sign. The audio is then processed and transformed right into a spectrogram, which represents the sound as a picture exhibiting how its frequency content material adjustments over time. We then used a Convolutional Neural Community (CNN) to be taught the acoustic patterns related to completely different mosquito species.

The educated mannequin can run on a small embedded system. In our prototype, we used an Arduino Nano 33 BLE, an open-source and comparatively cheap growth platform able to operating TinyML fashions regionally. That is important as a result of the system doesn’t basically require an web connection to carry out the classification. The intelligence is delivered to the sting, relatively than requiring the audio to be repeatedly uploaded to a cloud server.

I’d describe the novelty much less by way of creating a brand new microphone and extra by way of integrating acoustic sensing, sign processing, machine studying and edge computing into a conveyable, low-cost organic sensing system. The last word objective is to allow an preliminary identification on the location the place the mosquito is discovered, relatively than requiring each pattern to be transported to a laboratory.

What sort of information and coaching have been used for the machine studying on the edge?

The preliminary work used mosquito wingbeat recordings from the Humbug mosquito acoustic dataset related to the College of Oxford. We experimented with three essential mosquito teams: Anopheles, Culex and Aedes aegypti. I subsequently added synthetically generated audio information, together with ambient noise information, to assist establish silence when no mosquitoes are current.

The uncooked acoustic recordings weren’t instantly appropriate for classification due to noise and the traits of the recordings. We carried out audio processing and transformed the alerts into spectrogram pictures, which grew to become the enter to a CNN-based classifier. The mannequin was subsequently designed for TinyML and edge inference, permitting the educated intelligence to run on a small, embedded system relatively than requiring a standard pc or cloud infrastructure. Within the prototype experiments, the general classification accuracy was roughly 89 per cent, whereas Anopheles reached roughly 92 per cent within the observations reported in my demonstration.

Nonetheless, I’d emphasise that these are prototype outcomes and shouldn’t be interpreted as a declare of field-ready medical or epidemiological accuracy. Actual-world efficiency is significantly affected by environmental noise, which stays one of many predominant areas I’m working to enhance.

Does it leverage IoT capabilities to allow real-time information switch and motion?

Sure, IoT is a crucial extension of the system, though connectivity shouldn’t be required for the essential identification course of. The system can carry out the preliminary AI inference regionally and, when connectivity is on the market, ship the classification end result, location and time to a central dashboard. That is at present on the starting stage.

The idea is to deploy these units at places the place mosquito breeding is more likely to happen. As an alternative of manually gathering water samples and transporting them to a laboratory, the system present steady info on mosquito presence and exercise. The dashboard might finally perform in a fashion much like a Google Maps site visitors map, however as a substitute show mosquito-vector exercise. For instance, one space might present elevated Aedes exercise, whereas one other might present Culex or Anopheles. Public-health authorities might then establish potential hotspots and examine them earlier than an outbreak develops. This represents the longer-term imaginative and prescient described in my prototype work.

What’s the scalability and applicability to different use-cases?

I consider scalability is probably probably the most important points of the technology. The elemental structure consists of a low-cost sensor, edge AI, connectivity and distributed sensing. As a result of AI inference can happen regionally, the sensing nodes don’t essentially require highly effective computer systems or steady cloud connectivity. My earlier experiments with birds and child cries demonstrated the broader precept that completely different organic occasions can have acoustic signatures, and that machine studying can be taught these patterns. I see this as greater than a mosquito detector; relatively, it’s an instance of AI-enabled organic sensing on the edge.

Are points like distant reachability, assaults on breeding grounds, illness vectors, and so forth. in your radar? 

Sure, that is an space the place the know-how might have important public-health purposes. Present surveillance can contain gathering samples from suspected breeding places and sending them to laboratories. The method I’m exploring is to position clever sensing units nearer to potential breeding websites, significantly in areas the place mosquito breeding is more likely to happen.

After the monsoon, for instance, stagnant water and different breeding environments can emerge throughout massive city and rural areas. Quite than relying completely on periodic handbook inspections, distributed sensing units might probably present extra steady info on mosquito exercise and the presence of disease-vector species.

This might assist shift surveillance from merely figuring out areas the place mosquitoes could also be current to acquiring extra particular info on which disease-vector species are energetic in a selected location. Such info might assist public-health authorities prioritise areas for additional inspection and intervention. I see the know-how primarily as a surveillance and early-warning layer that may assist, relatively than substitute the work of public-health professionals.

Have you ever tried deploying it wherever? What has been the expertise and classes?

I’ve constructed and examined a working prototype, however I’d describe the present stage precisely as a proof of idea relatively than a large-scale area deployment. Some of the attention-grabbing experiments concerned shifting from recorded mosquito sounds to actual mosquitoes. I used to be capable of seize particular person mosquitoes and produce them near the system whereas they have been buzzing, and the prototype efficiently categorized them. Nonetheless, this experiment additionally highlighted the largest problem: ambient noise.

A mosquito’s wingbeat is a particularly low-amplitude sound. In a quiet surroundings, the system can detect it comparatively effectively, however in an city surroundings with site visitors, folks, equipment, speech and different environmental sounds, the mosquito sign can grow to be troublesome to isolate. One of many main analysis instructions is bettering the microphone and signal-processing pipeline to isolate the precise frequency traits of mosquito wingbeats from environmental noise.

This limitation additionally highlights an essential lesson: a mannequin that works in a managed dataset is barely the start. The actual problem is making AI work reliably in nature.

Can surveillance finally be translated into therapy or mosquito removing?

Sure, there may be potential for surveillance information to tell subsequent therapy or mosquito-control interventions. Nonetheless, I see the AI system primarily because the detection and intelligence layer, relatively than because the therapy or removing mechanism itself. Its function is to establish mosquito exercise and, the place potential, the disease-vector species concerned, offering info that may assist selections on the place and the way interventions could also be required.

How do you see lasers, LiDAR, AI mosquito killers, IoT traps, drones and genetically modified bugs specifically in India?

I see all of those as probably complementary applied sciences. On this broader ecosystem, my know-how would primarily function an info layer that may assist establish the place the issue is and what vector is concerned, whereas leaving intervention selections to public-health and environmental specialists.

How can know-how be used for insect management in public healthcare? What’s the largest alternative and largest barrier?

I consider the largest alternative lies in shifting mosquito management from reactive intervention in direction of extra predictive and data-driven public-health surveillance. Mosquitoes are a serious illness vector, and important sources are dedicated to spraying, fogging, eliminating breeding websites, laboratory testing and different interventions. The know-how I’m creating explores whether or not mosquito wingbeat patterns can be utilized to establish particular disease-vector species, probably offering an extra supply of knowledge for surveillance.

If 1000’s of cheap, clever sensing units might finally be deployed throughout a metropolis or area, they might probably create a real-time map of vector exercise. This would supply public-health authorities with extra particular info on the place explicit disease-vector species are energetic and the place their exercise could also be growing, serving to them prioritise areas for additional investigation and intervention.

The most important alternative, for my part, is early detection. The most important barrier is real-world reliability, significantly the problem of distinguishing the very weak mosquito acoustic sign from environmental noise, validating the fashions throughout completely different geographical areas and mosquito populations, and establishing that the ensuing surveillance can contribute to improved public-health outcomes.