SYRENE: An Underwater Embedded Artificial Intelligence Camera for Invasive Fauna Monitoring
Invasive species have a major impact on marine ecosystems by disrupting the natural balance and biodiversity. Introduced mainly through human activities such as maritime transport, fishing and aquaculture, these invasive species can supplant endemic species and cause changes in the functioning of ecosystems. These changes can have significant negative consequences, both ecological and socio-economic, including biodiversity loss, leading to a decline in species stocks. In the Mediterranean Sea, one of the most affected regions, the introduction of invasive species has already caused profound changes in marine habitats and food webs. This amplifies the vulnerability of ecosystems to global changes.The SYRENE project and the need for an invasive fauna active monitoringStarted in 2020, the SYRENE project (SYstème in situ de REconnaissance en temps réel de faune invasive sous-mariNE) aims to use embedded artificial intelligence to help local authorities detect and quantify invasive underwater species that threaten the balance of local marine ecosystems. Artificial intelligence makes it possible to detect complex shapes that were impossible to distinguish with any degree of accuracy just a few years ago.Some invasive species, such as lionfish, are active at night or at dusk, making them difficult to count or detect by divers. Active, real-time monitoring would enable rapid detection of their appearance and assessment of their impact. A solution based on embedded AI, capable of continuously analyzing images over a long period of time, offers a promising alternative to post-process detections running on data centers, without energy constraints.The team focused on Pterois volitans/Miles (lionfish), a species widespread across the globe. Accidentally introduced in the early 1990s to the Caribbean Sea and western Atlantic, it has invaded this area where it causes serious damage to marine ecosystems. For example, the biomass of small fish has declined by 65% in the Bahamas reefs in just two years due to its intense predation. It disrupts biodiversity and local food chains. [1]A lionfish in its natural habitat. Credit: Adobe Stock/crisodSYRENE V1, a lionfish detectorThe SYRENE prototype was initially developed in 2020 as a demonstrator to illustrate the capabilities of embedded deep learning image processing. The inner hardware architecture is simple and low cost, with an electronic board (Raspberry Pi 4B) allowing connection to an Intel Neural Stick V2 VPU (Vision Processing Unit), which accelerates the processing of images or videos (640*480 pixels) from a Raspberry Pi OV5647 5 Mpixels camera module.As image processing requires significant computing power, the challenge was to achieve minimum performance on consumer hardware with limited energy, as the system runs on battery power: a compromise on performance was therefore made with this hardware.SYRENE V1 ready to be vertically immersed with its mirror. Credit: IFREMER S. BarbotThe detection model we used in 2020 is YOLOv4 (You Only Look Once), a deep learning algorithm designed for real-time object detection. The model was adapted to take full advantage of the VPU's capabilities. We used transfer learning to adjust the model parameters to our dataset.Lionfishes’ tank with SYRENE in the right corner. Credit: IIFREMER S. BarbotCreating a dataset and training the modelTo train our specialized model for Pterois Volitans/Miles, or lionfish, a specific dataset was compiled. This dataset was assembled from various sources, including open-access resources and recordings made in natural environments and aquariums. This diversity of sources provided a broad representation of different lighting conditions and viewing angles, reinforcing the robustness of the future model.The creation of this dataset was a crucial step in ensuring the relevance and performance of the model. By choosing to target a single species exclusively, we significantly reduced the risk of false positives (detection errors with other species), while optimizing inference time and, indirectly, the energy consumption of the device.Detector testThe prototype was tested in March 2021 under near-real conditions, with budget in mind: submerged in a tank containing lionfish belonging to the Océanopolis marine center, with a return via the mobile telephone network enabling detection alerts and snapshots remote transmission.Maximum performance with this hardware architecture enabled real-time processing at approximately 5 FPS (frames per second). The ultimate goal is to achieve 1 FPS in order to find a better compromise in terms of power consumption, as this species moves slowly.Realtime embedded detection with SYRENE. Credit: IFREMER S. BarbotThe detector makes a local copy of the full-resolution photos of the detections made, which will ultimately enrich our learning dataset. It could also send the number of detections via an IoT network (e.g. LoRa), where bandwidth is very limited.Power consumption averages less than five watts at 5 FPS, and its battery life exceeds 100 hours without optimization for the available battery volume in the enclosure.Toward SYRENE V2 and Small Vision Language Models (sVLMs)Updated in 2025, SYRENE has benefited from the hardware and software advances of recent years. We turned to an embedded GPU from Nvidia (Jetson nano Orin NX) to test the potential of a system where power consumption would be less of a constraint. This time, the entire system consists of two modules: a low-power AI module and an AI module with higher computing power.Depending on the trigger conditions and remaining battery life, the low-power module switches the embedded GPU, providing embedded processing power for denser data and therefore more accurate detection than before.Ultra Low Power AIThe module, designed for continuous operation, is based on a microcontroller incorporating an embedded convolutional neural network (CNN) accelerator, optimized for very low energy consumption. The embedded AI unit enables it to finely filter relevant events and trigger the high-computing power section only when necessary.In the future, new types of more advanced neural networks, such as spiking neural networks (SNNs) and adapted processors, could be tested for the purpose, further reducing energy consumption while improving detection quality.SYRENE V2 hybrid architecture. Credit: IFREMER S. BarbotAI with high computing powerThis new module can now integrate more powerful AI models such as YOLOv11, which significantly improves detection performance. This system allows YOLOv11 to take full advantage of the GPU's parallel computing capabilities, benefiting in particular from the optimizations offered by TensorRT, thereby reducing real-time latency by a factor of 4 in some cases we have tested.Re-training (fine-tuning) the neural network significantly improved detection and processing performance. According to our tests, the computing power/energy consumption ratio with the old hardware is greater than six, opening up new possibilities in terms of image resolution: this translates into a greater detection distance, allowing smaller or more distant objects to be detected with greater accuracy.This extra power also allows for the integration of Transformer-type networks [2] as well as small vision-language models (sVLM) [3]. These models favor a textual description of the image rather than a direct transfer of the image, thereby achieving significant information compression. This is particularly useful for acoustic links with seabed equipment, where a detector such as SYRENE will be deployed.Very soon, it will be possible to integrate autonomous embedded AI agents designed to perform specific tasks, such as sending SMS alerts in the event of specific events (e.g. species counting). These agents will be able to analyze data locally, make decisions in real time and act independently, while limiting the energy consumption of embedded systems.ConclusionEvolving to keep pace with recent technological advances in AI, SYRENE is an example of an oceanographic application that improves environmental monitoring and anticipates the impact of invasive marine species, particularly lionfish. Coupled with an antenna and an IoT (Internet of Things) modem for shallow-water applications such as lagoons, or via a surface relay buoy for benthic applications, this type of architecture opens up a wide range of possibilities. It enables long-term monitoring of parameters such as acoustics, video, seismicity and other physicochemical parameters, as well as the compression of useful information.In conclusion, the SYRENE detector thus offers new perspectives for real-time monitoring of underwater fauna via imaging, acoustics or other parameters.AcknowledgmentsWe would like to thank Pierre Ternat et Dominique Bartelemy (Oceanopolis) for having kindly hosted the camera in their lionfish tank, Helene Leau (Ifremer), SincObs project coordinator as well as Amaury Tisseau, Thomas Morales and Alvaro Scarramberg for their involvement in this project.Credits[1] Green, S.J., et al. (2012)[2] Vaswani et al., 2017 : https://arxiv.org/abs/1706.03762)[3] https://hal.science/hal-04889751About the authorsLaurent Gautier is a mechatronics engineer in IFREMER’s Research and development lab. He is involved in the development of several benthic observatories, as well as in technological devs for scientists.Stephane Barbot is a software engineer in IFREMER’s Research and development lab.Damien Le Vourc’h is a mechanical designer in IFREMER’s Research and development lab.