Our Achievements
-
Since 2023, PIXSEA works on the WILDDRONE-PIXSEA project to implement marine surveys using electrically powered drones, aiming to replace aircraft, decarbonise and automate wildlife observation, particularly over offshore wind farms.
-
From 2021 to 2024, PIXSEA conducted the ELADOM study on marine megafauna in southern Finistère for the French Biodiversity Agency (OFB), combining visual observations with high-resolution aerial photographs analysed using Deep Learning. The project significantly advanced methods for automatic species detection and identification.
-
From 2019 to 2023, the SEMMACAPE project enhanced the performance of marine megafauna detection algorithms, demonstrating that with expert validation, their results match those of airborne observers, while reducing errors through technical improvements.
WILDDRONE-PIXSEA Project
From 2023 to 2025, the PIXSEA consortium has been the recipient of a call for projects launched by the GIS EMYN, through the WILDDRONE-PIXSEA project. This project, which is still ongoing, aims to implement digital aerial surveys of marine megafauna using drones (Rolland et al., 2024, 2025). Flight tests have been carried out during the construction phase of the Yeu-Noirmoutier offshore wind farm, using Beluga drones developed by DroneVolt. The project seeks to replace aircraft with electrically powered VTOL (Vertical Take-off and Landing) drones in order to decarbonise and reduce the risks associated with maritime observations. On the technical side, the PIXSEA team is responsible for integrating image acquisition systems on drones. In addition, the team performs image analysis using neural networks, expert validation by naturalists, data analysis and report writing.
Rolland, E. G. A., Laporte-Devylder, L., & Christensen, A. L. (2025). Advancing Wildlife Monitoring in Gregarious Species with Drone Swarms. In G. Marreiros, L. Grande, J. P. Llerena, L. Conceição, H. Ko, M. Plaza, & M. Ricca (Éds.), Special Sessions II (Vol. 1151, p. 310‑316). Springer Nature Switzerland. https://doi.org/10.1007/978-3-031-80946-0_33
Access the poster presented at Seanergy 2024 here.
Flight test of the Beluga drone by DroneVolt
Credit: Avion Jaune
ELADOM Project
From 2021 to 2024, the PIXSEA consortium conducted a study on marine megafauna and anthropogenic activities through aerial monitoring over the extension of the Iroise Marine Natural Park in southern Finistère, on behalf of the French Biodiversity Agency (OFB) and the Offshore Wind Observatory, as part of the ELADOM project.
To this end, visual and digital aerial monitoring campaigns were carried out using high-resolution photography in parallel with visual observations. The photographs were analysed by a neural network for the detection and identification of the observations, followed by expert validation of a subset of images. The team was involved throughout the entire project, undertaking sampling design, aerial monitoring, image analysis, data processing and report writing.
The extensive feedback from these campaigns contributed to the continuous improvement of PIXSEA’s methods and techniques:
This project led to the establishment of an average image resolution of 1.7 cm/pixel for both visual and digital campaigns, facilitating species recognition and improving identification accuracy by both observers and the neural network;
The heterogeneity (number of observations per image, species diversity, varying conditions, photographic quality, etc.) and the large volume of photographs collected during the project enriched the neural network’s training dataset;
Optimal threshold values for the classification of observations (species, families, groups) were determined, enhancing the accuracy and robustness of the neural network’s outputs;
The neural network detected 50% more observations than human observers, causing an increase in the proportion of photographs verified by experts;
Automated analysis also enabled improved counts of animal groups by identifying and estimating the number of individuals beneath the water surface and within observation strip transects;
Species-level identification of flying shearwaters reached an accuracy of nearly 90%, thanks to the combination of neural network analysis and expert photo validation.
Access the mid-term report of the ELADOM project here.
Boat followed by a group of birds.
Credit: PIXSEA, 12/06/2023
1:1 photograph of a group of birds detected by the Harmony software as 4 shearwaters in flight, confirmed as Balearic shearwaters in flight by an ornithologist.
Credit: PIXSEA, 12/06/2023
1:1 photograph of a group of birds detected by the Harmony software as 45 shearwaters, both resting and in flight. The software enabled an exact count but could not determine the species of the resting birds (specific characteristics not visible). The individual in flight was identified as a Manx shearwater by an ornithologist (distinctive white markings).
Credit: PIXSEA, 12/06/2023
Zoom at a resolution of 15 mm/pixel on an individual detected as a Manx shearwater in flight by the neural network with a confidence score above 0.95.
Credit: PIXSEA, 04/2022
Density map of Alcids in spring 2022, produced by kriging.
Credit: PIXSEA, 12/2022
Map of the spatial and temporal observations of top predators.
Credit: PIXSEA, 12/2024
SEMMACAPE Project
From 2019 to 2023, the SEMMACAPE project, carried out in collaboration with the French Biodiversity Agency (Agence Française pour la Biodiversité), France Energies Marines, the University of Southern Brittany and financially supported by the ADEME (APRED2019 call for projects), enabled the acquisition of high-resolution images of marine megafauna to improve the detection and species characterisation capabilities of image recognition algorithms (both unsupervised and supervised learning algorithms, i.e. without and with expert image verification). Furthermore, this study allowed comparison between detections and identifications made by airborne observers and those produced by the algorithms from image analyses.
The project demonstrated that supervised learning algorithms with expert validation (i.e. proposals validated and corrected by experts) can produce distribution maps closely matching those based on visual observations recorded by observers onboard the same aircraft. Feedback from this project also permitted the assesment of sun glare impacts on the neural network’s detection and identification capabilities on photographs. Sun glares are now reduced during acquisition by using polarized anti-glare filters and by tilting the imaging system by the operator, thereby decreasing the number of false positives for the neural network in the end.
Access the final report of the SEMMACAPE project for the ADEME here.
Access the final report of the SEMMACAPE project for the French Biodiversity Agency here.
Photograph of a Risso’s dolphin annotated by a biologist.
Credit: PIXSEA
Photograph of a group of striped dolphins accompanied by tunas, annotated by a biologist.
Credit: PIXSEA
95% confidence interval of the densities (individuals/km²) of great gulls, estimated for two survey sessions (05.2020 and 11.2020) according to three observation methods: automated digital method (in yellow), supervised verified automated digital method (in orange) and visual method (in green). The same letter indicates that the results are not significantly different, meaning there is no significant difference between the three methods in this case.
Distribution maps of encounter rates for Alcids (targets/km when detected by the algorithm and individuals/km when confirmed by observers) on 12 November 2020 in the Gironde Estuary and Pertuis Sea Marine Natural Park, produced according to four observation methods: automated digital method (first row, first column), supervised automated digital method (first row, second column), supervised verified automated digital method (first row, third column) and visual method (second row, third column).