Introduction
Barbed wire jellyfish (Apolemia sp.) is a marine organism that self-organizes into string-like colonies of length ranging from a few centimetres to over 30 metres. The individuals that make up a colony contain hair-like structures with stinging cells – making a colony resemble barbed wire both in look and touch. The genus is distributed in the Mediterranean, Pacific, and Atlantic seas. In recent years, it has been observed with great abundance along the Norwegian coast, both intact colonies and swarms of fragments. Such swarms have caused significant disruption of aquaculture operations and mass mortality events in salmon fish farms, resulting in economic losses.
The new appearance of barbed wire jellyfish along the Norwegian coast, and the ensuing challenges for Norwegian aquaculture, necessitate a better understanding of their distribution and abundance. However, these gelatinous zooplankton (such as Apolemia sp.) are often overlooked in Norwegian marine monitoring programs. Moreover, traditional sampling techniques are poorly suited for capturing gelatinous organisms, leading to significant knowledge gaps regarding their distribution, abundance, and seasonal dynamics across Norwegian waters. To address these challenges, the Norwegian citizen science portal “Dugnad for Havet” (DfH) [1] was launched in 2019.
Using expert-validated images from DfH, we have developed and benchmarked machine learning models which detect barbed wire jellyfish in citizen science images. We describe these models and their performance below. The best-performing model is intended for integration into DfH to automate identification of Apolemia sp. and is part of the development of an early-warning system through the JellySafe project.
Acknowledgements: We are grateful to Sajidah Ahmed, Marius Andersen, and Øystein Knutsen for discussions and assistance with various activities related to this text, including data cleaning, labelling, and model development. This activity was funded by the FHF through the project JellySafe (FHF/ 901941).
Methodology
The «Dugnad for Havet» (DfH) dataset. The DfH project [1] by the Norwegian Institute for Marine Research invites participants to submit observations of gelatinous zooplankton. Each observation includes fields such as photograph(s) and/or video (optional), proposed species identification, geographical coordinates, date and time. Observations have been submitted by a broad range of users, including recreational divers, fishermen, as well as aquaculture personnel. Most images were captured using smartphones from the sea surface.
Between January 2019 and April 2025, there were 4080 observations of gel. zooplankton recorded. Notably, the siphonophore Apolemia sp. accounted for over 50% of the total observations, with 2068 records and approximately 1400 accompanying images collected to date. To ensure data reliability, each submitted record undergoes a manual validation process conducted by gelatinous zooplankton specialists.
Despite its many advantages, the DfH citizen science approach presents several inherent biases and challenges, including sampling bias, seasonal bias, observer bias, and a high validation workload. By developing models which automate identification, we aim to improve the efficiency of the validation process and leverage the growing volume of submitted images, focusing first on the most reported and harmful species Apolemia sp.
Machine learning: architectures and training. Citizen science images present challenges for object detection such as variable lighting, varying resolutions, camera settings, etc. After strict data cleaning, our Apolemia sp. dataset consisted of 350 images. To deal with small datasets, leveraging pre-trained models is a common strategy. We compare the recent state-of-the-art models YOLOv12 [2] and RF-DETR [3] on our detection task.
The YOLO family, originally based on Convolutional Neural Networks (CNNs), has long dominated one-stage detection due to its efficacy and speed. While early versions primarily relied on local receptive fields, recent models (e.g., YOLOv12) increasingly incorporate attention mechanisms to capture broader context. In parallel, transformer-based detectors such as DETR and its successors (e.g., RF-DETR) fully embrace global attention, offering advantages in cluttered or complex scenes. These architectures may therefore excel in detecting complex or cluttered scenes, such as the elongated and diffuse morphologies of barbed wire jellyfish.
Both models were trained with 10-fold cross-validation and standard augmentations.
Results and discussion
The models were evaluated on their mean Average Precision (mAP) [4] scores, which is the percentage of objects correctly detected. Here, ‘detect’ means that the model predicts a bounding box covering a threshold of the ground truth bounding box. We use a threshold of 50% to indicate detection skill, and average across a threshold range of 50%-95% to reward more precise outlines.
Both the models YOLOv12 and RF-DETR performed well on the detection task, achieving mAP 50% scores of 94.0% and 91.4%, respectively. Their mAP 50-95% scores were essentially identical, at respectively 76.4% and 76.3%. This indicates that the YOLOv12 model was able to correctly detect more of the Apolemia sp. targets, but that the RF-DETR bounding boxes were slightly more precise.
Future directions. Our models are intended to reduce the validation workload of DfH and to be part of an early-warning system for harmful jellyfish blooms along the Norwegian coast. This development is ongoing through the JellySafe project [ref]. As more images become available through DfH, we also aim to include other gelatinous zooplankton.
A drawback of the bounding box models presented above is that they don’t give precise quantity estimates. To address this, we are investigating pixel-wise segmentation models.
References
[1] Norwegian Institute for Marine Research, “Dugnad for Havet.” [Online]. Available: https://dugnadforhavet.no
[2] Y. Tian, Q. Ye, and D. Doermann, “YOLOv12: Attention-Centric Real-Time Object Detectors,” Feb. 18, 2025, arXiv: arXiv:2502.12524. doi: 10.48550/arXiv.2502.12524.
[3] I. Robinson, P. Robicheaux, and M. Popov, RF-DETR. (2025). [Online]. Available: https://github.com/roboflow/rf-detr
[4] T.-Y. Lin et al., “Microsoft COCO: Common Objects in Context,” Feb. 21, 2015, arXiv: arXiv:1405.0312. doi: 10.48550/arXiv.1405.0312.
1 A standard mosaic dataloader was only available for YOLOv12, not RF-DETR.