Unique Wageningen expertise set the standard in automated seabird survey
2 April 2026
The unique ecosystem of Wageningen University & Research (WUR) and Vision+Robotics (V+R) has repeatedly proven its value in national and international research, on land and at sea. This is also reflected in the results of the public-private partnership (PPS) project AI-MEG.
At the end of 2025, Wageningen researchers completed the nearly four-year project AI-MEG. Or rather AI-MEG-I, because based on its results, Rijkswaterstaat intends to commission a follow-up project: AI-MEG-II. What makes these results so notable? We asked Martin Poot, researcher at Wageningen Marine Research, and Lydia Meesters of Wageningen Food & Biobased Research.
“AI-MEG-I was a PPS project commissioned by the Top Consortium for Knowledge and Innovation(TKI) Agri & Food together with Gemini wind park. The aim was to develop an AI model for detecting and identifying seabirds and marine mammals using high-resolution images from digital cameras. These images came from two aerial monitoring campaigns at offshore wind farms in the Dutch North Sea. The focus was on high-resolution imagery and deep learning to improve efficiency, accuracy and cost-effectiveness,” says Martin Poot, ecologist and ornithologist at Wageningen Marine Research.
“Mapping and monitoring seabird and marine mammal populations from an aircraft is very time-consuming and demanding work. Typically, as an ecologist, you spend several hours in a plane flying at 200 kilometres per hour at an altitude of 75 metres. While the pilot follows transects, you visually count seabirds or marine mammals on the water surface, which passes beneath you like a rapidly moving conveyor belt. You dictate the species observed and the number of individuals per species into a recorder, also noting different distance bands so corrections can later be made for detection loss, observer effects and visibility conditions.
You need to be a highly skilled ornithologist, and even then the work is physically demanding. Species such as the common guillemot, with their dark plumage, are particularly difficult to detect, so some birds are inevitably missed. Different observers also record different counts, especially in large flocks. There is always an unknown counting error, where rare species can easily be overlooked within large groups.
Flying over the North Sea in this way takes three days, and even then only 1.5 to 2 per cent of the total surface area is covered. The remaining area is interpolated from the collected data, taking care not to count individual birds twice. This approach also involves risks. I once experienced a near collision with a gas platform. For safety reasons, we can no longer map populations within offshore wind farms using visual methods that require low-altitude flights. Yet wind farms are precisely the areas of interest when assessing impacts on seabird and marine mammal populations.”
High-resolution cameras and algorithms
How can populations be mapped more effectively, quickly and safely, while also improving data accuracy and quality? To address this, Martin Poot, Lydia Meesters and colleagues collaborated with the Gemini offshore wind farm in the North Sea, the Ministry of Agriculture, Fisheries, Food Security and Nature (LVVN), and Rijkswaterstaat (RWS).
Meesters explains: “At the Gemini wind farm, north of Schiermonnikoog, and at the Borssele wind farm, we worked with data collected by aircraft equipped with high-resolution cameras for automated digital observations. These aircraft carry four high-resolution RGB cameras mounted in an opening in the underside of the plane, allowing observations to be collected at an altitude of 500 metres. This means data can be collected within wind farms, as the aircraft fly above the tallest turbines.
The camera sensors are designed for images of approximately 6,000 by 2,000 pixels. We use four cameras because a single camera cannot capture an entire transect. We analyse video at around seven frames per second, as the animals need to be tracked to avoid double counting. Ultimately, the AI model must, like a human observer, not only produce a species list but also estimate total numbers as accurately as possible.
The images were annotated by experts from the German organisation BioConsult SH, identifying where birds are located and which species appear in each image. These data were then used to develop deep learning models that can fully automate the analysis of camera images, enabling automatic classification and counting of species. That is the ultimate goal.”
Up to 90 per cent detected automatically
Meesters continues: “Although we started on a small scale, the initial results already generated considerable interest in our model. It can automatically detect and classify seabirds on the water surface. At present, the model covers 12 bird species and can automatically detect up to 90 per cent of observed individuals of these species under calm sea conditions. Some species can already be classified well, others less so. The highest classification accuracy for one of the 12 species currently stands at 93 per cent. That is already a high score given the relatively limited dataset.
We are developing an AI model that automates the entire process, and this inevitably involves a margin of error. There is sometimes a misconception that an AI model cannot or should not make mistakes. You need to build trust in such a model and accept that results are not always 100 per cent accurate.”
As an ecologist, Poot was initially sceptical, like many in the field, about the feasibility of detecting and classifying seabirds on the water surface using vision technology and artificial intelligence. “Especially in a constantly moving sea, where white foam and dark plumage complicate identification. The capabilities of AI have nevertheless surprised me, provided you have sufficient data and approach the analysis carefully. In terms of implementation, I believe we now have a reproducible and reliable AI model for the most common species. It is a clear improvement and opens up opportunities for more extensive and targeted research. From a scientific perspective, this represents a major development.”
Why WUR stands out in this type of project
“For me, it is very clear why WUR and Vision+Robotics perform strongly in a PPS project like this,” says Poot. “The level of expertise required for this type of research is only available here. We have the facilities, know-how and integrated knowledge needed to carry out such projects. Other parties are certainly active in this field as well. We believe the market will ultimately benefit from such a tool, provided it is made publicly available with high-quality hardware and facilities. We therefore explicitly invite external parties to join the project and combine strengths. This is recognised in the market. That is why we receive mandates from clients such as Rijkswaterstaat and from government bodies, where market parties also have a role.”
“We are able to address relatively complex challenges thanks to our network and because we bring together all relevant disciplines and expertise,” adds Meesters. “Developing models is something many organisations can do. What distinguishes us is the close collaboration between our AI experts and marine biologists, who contribute ecological knowledge. This combination ensures that we not only build strong AI models, but also understand exactly which biological information is essential for a useful application.
In the case of AI-MEG-I, we have shown what is possible with relatively small datasets. This builds confidence among clients and has already led to plans for a second project: AI-MEG-II.”
Poot adds: “We aim to expand beyond 12 bird species and to apply the developed methodology to the observation and classification of marine mammals. Because species diversity in this group is relatively limited in the North Sea, we also intend to expand internationally.”
Towards safer and more sustainable monitoring
Following the success of AI-MEG-I, Poot is now convinced of the value of vision technology and artificial intelligence for seabird surveys at sea. “Human observers are likely to miss more birds than a camera ever will. A camera does not get tired, and its observation performance does not decline during a flight. The data are reproducible, and by flying at higher altitudes, observations can still be made within wind farms. This is important, particularly for species such as the common guillemot, which is the most abundant in our part of the North Sea, and because we know wind farms affect its population.”
Lydia Meesters concludes: “In my view, the main impact of our research is that human observers are only needed to a limited extent. We will still need a human in the loop, someone who checks automatic detection and classification at key moments. This human step ensures quality control and prevents errors from going unnoticed. However, the labour-intensive step of manually annotating entire datasets can be removed. I am often struck by how relatively straightforward it can be to automate human effort using AI.”
Martin Poot adds: “A next step could be to use drones. Modern drones can carry the same camera technology and operate for up to eight hours. This would make observations even safer and more sustainable. At the same time, I would argue for maintaining validation flights with human observers, so that as ecologists we remain connected to real-world conditions.”
