Marine wildlife trafficking: use of AI algorithms for the autodetection of shark fins, seahorses and sea cucumbers

1 Introduction

Wildlife trafficking is an illegal activity involving live wildlife and/or wildlife whole animals/parts being illegally traded, often across borders (Pirotta et al., 2022). The physical removal and/or introduction of exotic species into new environments has the potential to impact all species on earth, across fauna and flora, including humans, and all ecosystems (‘t Sas-Rolfes et al., 2019). The illegal wildlife trade (IWT) is a highly lucrative global problem, worth billions annually, making it one of the world's most profitable criminal activities (INTERPOL, 2025). The IWT has various drivers, some of which include; demand for medicinal uses (Wang et al., 2022; Zanvo et al., 2024), the illegal pet trade (Toomes et al., 2023; Padang et al., 2025) and ornamental collections (UNODC, 2024). Illegal wildlife trade increasingly intersects with other serious crimes and illicit commodities including drug trafficking, arms trafficking, human trafficking, illegal mining and counterfeit goods trade (Anagnostou et al., 2025; Van Uhm et al., 2021; INTERPOL, 2025). In some cases, researcher observe ‘product convergence' where wildlife and illegal drugs are combined within the same products (Anagnostou et al., 2025).

Marine wildlife trafficking (MWT) involves the illegal movement of whole marine species (alive or deceased) and/or parts of marine species, and is a lesser documented component of the IWT (Shaw, 2024). MWT may target a range of marine species such as, elasmobranchs (e.g., sharks and rays), teleost fishes (e.g., tropical fish, tuna, sturgeons, eels, aquarium fish, and fish parts (e.g., fish swim bladders)), syngnathiformes (e.g., seahorse), marine reptiles (e.g., turtles), mollusca (e.g., giant clams and conch), corals to whole marine mammals (e.g., dolphins) (Foster et al., 2025; Pavitt et al., 2021; Shaw, 2024; Wright and Mashalla, 2025). MWT occurs at various scales, including at sea (e.g., illegal, unreported, and unregulated (IUU) fishing (Miller et al., 2019)) and on land (Pavitt et al., 2021).

Globally, the international trade (import, export, re-export, and introduction from the sea) of wildlife, including marine species, is regulated by the Convention on International Trade in Endangered Species of Wild Fauna and Flora (CITES), which is aimed at preventing overexploitation of listed species. CITES aims to control the environmental harms associated with international wildlife trade through a permit-based regulatory approach to promote ‘legal, sustainable, and traceable' trade (CITES, 2025). Annual Government reporting of seized CITES-listed species and online seizure records (which tend to have a “taxonomic bias” for terrestrial species Foster et al., 2025), provides some insight into the movement of species involved in the IWT. Despite this, specific reporting on MWT is lacking and more generally, the IWT does not account for all wildlife, as millions of species may be traded and harvested internationally, many of which are not covered by CITES (UNODC, 2024).

Reporting and documenting MWT and more widely, the entire IWT, is reliant on detection.

Despite the IWT being one of the largest organized crimes, publications detailing detection methods are scarce and under researched (Moloney and Chaber, 2024). Ongoing and emerging efforts to physically detect and dismantle the IWT are largely focused on being able to detect trafficked species in the field (e.g., airports, mail/traveler pathways). Currently, this includes a combination of; (1) human detection methods (e.g., manual bag inspections, use of 2D X-ray screening), (2) biosecurity detection dogs and (3) use of new autodetection capabilities using Artificial Intelligence to create wildlife detection algorithms technologies, which is largely focused on terrestrial species (Pirotta et al., 2022).

In this paper, we expand on the innovative use 3D X-ray CT technology using Real Time Tomography (RTT) for the specific detection of MWT. We describe findings on a trial utilizing computer-based algorithms created from 3D X-ray images of three animal subclasses commonly reported in the MWT, both internationally and within Australia, including; (1) sharks (fins), (2) seahorse and (3) sea cucumbers. This research builds off initial research using AI and volumetric 3D computed tomography (Pirotta et al., 2022) as a promising complimentary tool for IWT detection along with existing detection methods.

2 Materials and methods

2.1 Data collection

Data collection took place between September 2024 and February 2025 at Rapiscan Ltd. Head Office, Homebush, New South Wales, Australia. Only samples of deceased marine species were used in this study (Table 1). Preserved (dried) specimens were obtained from the Australian Museum Collections, this included; (1) dried shark fins (all dorsal fins, with the exception of 1 pectoral and 1 caudal fin), species included; Isurus oxyrinchus, Shortfin Mako, Carcharhinus falciformis, Silky shark, Alopias superciliosus, Bigeye Thresher, Carcharhinidae, Requiem sharks, (2) seahorses (West African Seahorse, Hippocampus algiricus) and (3) sea cucumbers (Holothuria nobilis, Black teatfish, Holothuria atra, Black sea cucumber/lolly fish, Bohadschia argus, leopard sea cucumber, Thelenota ananas pineapple sea cucumber/prickly redfish). Sea cucumbers were treated with a fixative and dried as part of the existing Australian Museum collection (Supplementary Table 1). The shark fins were seized in a marine wildlife trafficking incident in Australian waters by NSW Department of Primary Industries and Regional Development (NSWDPIRD) and Australian Fisheries Management Authority and identified to species by the Australian Center for Wildlife Genomics at the Australian Museum using genomic techniques. The fins were then provided to the Ichthyology Collection, fixed in 5% buffered formaldehyde, dried and accessioned. Whole dried seahorses provided by the Australian Museum were also obtained through a marine wildlife trafficking seizure. Individuals were detected and seized in Hong Kong, then donated to the Australian Museum Ichthyology Collection for research purposes under permit from Agriculture, Fisheries and Conservation Department, Hong Kong, reference (13) in L/M 338/2023 in AF GR COB 07/11. The Australian Museum (AU010) is a registered institution for the exchange of specimens for scientific research including CITES listed species in accordance with CITES Article VII.6.

Table 1. Summary of marine wildlife samples used for the production for marine wildlife trafficking autodetection algorithms.

A mixture of dried samples obtained from the Australian Museum and an Asian grocery store were scanned including shark fin, seahorses and sea cucumbers. Only one example of a fresh shark fin was sampled. Samples were chosen based on their documented involvement in real world illegal trafficking scenarios. Seahorses used in this study were obtained from a seizure as part of the global illegal trade of marine wildlife.

Additional shark fins (unknown species) were purchased from an Asian grocer in Sydney, Australia, reflective of available shark specimens for purchase in a retail setting. Fresh shark fins were also collected from a single deceased bull shark (Carcharhinus leucas) to be used as comparison with dried specimens. Fresh bull shark samples were collected under a NSW DPIRD Research Permit Section 37 (Ref. PO1/0059A-2.0) and Animal Care and Ethics (ACEC 07/08) permit. Ethical review and approval were not required for this study as all wildlife specimens were limited to dead and/or preserved Museum specimens.

2.2 3D X-ray scanning

Marine wildlife samples were scanned using the Rapiscan Ltd. Real Time Tomography 110 (RTT^®110) system (Figure 1). This 3D system is currently deployed in-situ globally at various airports and mail/traveler pathways for the purpose of explosives and/or biosecurity detection. The RTT^®110 uses X-ray computed tomography (CT), which is a technique involving the use of X-rays to produce a series of image cross sections through an item. Unlike existing 2D technology, the production of 3D volumetric images provides personnel with information about an item content, with good spatial detail. 3D images are produced in the RTT^®110 via a stationary gantry with multiple X-ray sources to produce high resolution 3D X-ray CT images (voxel size 1.4636mm³) in real time (Thompson et al., 2012). This method allows images to be manipulated to provide a 360-degree view of an item to help identify concealed wildlife in real world detection scenarios e.g., luggage or via post parcel.

A series of treatments were used to scan each sample (Supplementary Table 2). This was based off an initial scientific design by Pirotta et al. (2022). Samples were scanned alone (baseline information) and with a variety of non-biological items, with varying densities and shapes. Scanning of items were also used to recreate known/public previous trafficking cases of smuggled wildlife in Australia e.g. within clothing, with children's toys and under tin material. A total of five treatments were collected for each sample, with additional scanning of samples with multiple samples in the same image.

2.3 Image segmentation

Scanned CT images were first segmented by thresholding (an image processing method) to extract the volumes within the range of organic materials. A morphological opening operation, followed by a connected component analysis, was used to generate individually labeled components for each CT image.

2.4 Detection algorithm

A detection model based on the 3D convolutional neural network (CNN) (Khagi and Kwon, 2020) was used to classify the labeled components into marine life subcategories and non-marine life components, identifying both smaller objects such as seahorses and sea cucumbers, and larger objects including shark fins. The CNN used in this model consisted of seven convolutional layers, each followed by a rectified linear unit as activation function and a max-pooling layer down-sampling the output by a factor of 2. The model parameters were determined through a training process on a collected dataset usually referred to as the ‘training set' by optimizing the parameters to minimize the detection error on this set. To improve the generalization ability of the model, a 3D Threat Image Projection (TIP) (Wang et al., 2020) was used to extend the training data set by inserting 3D marine life objects into CT images of negative scans. This data augmentation method served not only to increase the number of training images, but also to simulate the appearance of the data in real world scenarios, where marine species could be found concealed in baggage. During the training process, the model was optimized over 30,000 training steps.

The detection rate (probability of detection, P_D) for shark fins, seahorses, and sea cucumbers was defined as the proportion of the detected scans among all scans containing these items in each category respectively. The overall false alarm rate (probability of false alarm, P_FA) was determined by the proportion of any false detection among scans containing no marine life items. The false alarm rate on negative samples raised by the detection of each marine life category, was also calculated.

3 Results

A total of 68 marine wildlife samples of shark fin, seahorses and sea cucumbers were scanned for marine wildlife algorithm development (Table 1). This included; 17 shark fins (dried dorsal fins and pectoral fins), 1 fresh shark (dorsal, pectoral and caudal fins), 30 seahorses, and 20 sea cucumbers to produce 3D volumetric data (Figure 2). To address concerns over density and detection capabilities with shark fin, a comparison between fresh and dried shark fin samples proved helpful, with the similar appearance of shark cartilage across all three samples via the CT images (Figure 3).

3.1 Detection algorithm

The training data set consisted of 3,569 bags, including 437 containing shark fins (131 original and 414 TIP), 531 containing seahorses (128 original and 534 TIP), 186 containing sea cucumbers (39 original and 222 TIP), and 2,415 negative bags containing no marine species (Table 2). To assess the model's ability to generalize on the data that was unseen in the training set, we estimated the detection performance of the model on a separate ‘testing set', sampled from the same population as the data used during training. We evaluated model performance on a test data set consisting of 512 bags, among which 108 contain shark fins (26 original and 82 TIP), 131 contain seahorses (25 original and 106 TIP), 75 contain sea cucumbers (15 original and 60 TIP), and 198 negative bags contain no marine species.

Table 2. Dataset of marine life CT scans used for algorithm development for three marine species subclasses; shark fin, seahorse and sea cucumbers.

The performance of the detection algorithm across all marine species subclasses was high (Table 3, P_D shark fin: 95%, seahorse: 96% and sea cucumber: 86%), with low false alarm rates (P_FA shark fin: 2%, seahorse: 9% and sea cucumber: 1%). Examples of detection results (Figure 4) demonstrate successful automated algorithm detection in simple and more complicated, real-world type scenarios.

Table 3. Marine life algorithm detection performance for three animal classes.

The detection rate (probability of detection, P_D) for each marine life subclass is the proportion of the detected scans among all scans containing these items in each category respectively. The overall false alarm rate (probability of false alarm, P_FA) is the proportion of any false detection among scans containing no marine life items. False alarm rate on the negative sample raised by detection of each marine life category was also calculated.

4 Discussion

This paper presents findings from purpose-built AI algorithms for the detection of illegal marine wildlife trafficking in airport and mail/traveler pathways (luggage only). 3D X-ray CT scanning was used to develop marine autodetection algorithms across three marine wildlife subclasses (shark, seahorse and sea cucumber). To the best of our knowledge, these are the first dedicated marine wildlife trafficking algorithms developed for the purpose of detecting trafficked marine species. Findings demonstrate high detection capabilities (P_D overall 92%), with low false alarm rates (P_FA overall 13%). This research demonstrates the use of 3D X-ray technology and AI autodetection as a complementary detection tool, in addition to existing methods (e.g., visual/manual inspections, biosecurity dogs), for identifying cases of marine wildlife trafficking in global airport and mail/ traveler pathways. This study was produced using an existing commercially available product (RTT^®110) and co-authored by employees of Rapiscan. All collaborators, including Rapiscan, were involved in the methodology and algorithm development.

4.1 Auto-detection capabilities to detect IWT

This study builds on relatively recent efforts of using innovative 3D technology, with artificial intelligence, and complex analysis techniques for the detection of the illegal wildlife trade (Pirotta et al., 2022; Rathi, 2025). The adaptability of existing 3D imaging tools has already proven successful for land-based species known to be involved in trafficking, such as birds, fish and lizards, including Australia's most trafficked species, the endemic Australian shingleback lizard (Tiliqua rugosa), which has a distinctive body profile comprising a similarly shaped head and tail to aid in confusing predators (Pirotta et al., 2022). In contrast, species with more general shape profiles e.g., birds and fish, will result in general, non-species-specific detections (Pirotta et al., 2022).

Use of autodetection capabilities for wildlife is being documented more widely, particularly via online platforms (Kulkarni and Di Minin, 2023; Cardoso et al., 2023; Kumar et al., 2023), including social media (e.g., using machine learning, Di Minin et al., 2018). In contrast, within airports and mail facilities, existing threat detection tools are already in place, like the RTT^®110 3D x-ray scanner (used in this study) and more general contraband detection tools e.g., manual inspections, detector dogs (Moloney and Chaber, 2024). These existing threat detection tools have the potential to be coupled with AI to expand detection capabilities for the illegal wildlife trade. In parallel, the autodetection of other contrabands at international borders aids IWT enforcement efforts as MWT is becoming an organized crime activity alongside other transnational crimes including drugs, arms and people trafficking (Anagnostou et al., 2025). This presents a symbiotic opportunity to national and international enforcement agencies.

The three animal subclasses of marine species chosen in this paper were based off known examples of marine wildlife trafficking. Shark fin trafficking represents the largest documented subclass of illegally traded marine species, often occurring alongside trade relating to seafood (Hasan et al., 2023). Efforts to detect provenance and understand trade routes of shark fin via mail/traveler pathways and via online platforms have previously been documented and remains ongoing (Partin et al., 2022; Chakraborty et al., 2025; Prasetyo et al., 2024). The trade of dried seahorses, often used for traditional medicine, is also global, with a recent study identifying smuggling occurring across in 62 countries and most commonly intercepted at airports in passenger bags or in sea cargo, with evidence of diversifying trade routes (Foster et al., 2025). Sea cucumbers are the lesser known of the two subclasses, with evidence of their involvement in the illegal trade (e.g., trafficking in Mexico, Felbab-Brown, 2022) however, much of the literature is focused on an overfishing of fisheries perspective, rather than illegal movement (Anderson et al., 2011).

4.2 A combination approach to detecting the marine wildlife trade

This paper demonstrates the potential to expand illegal wildlife trade detection capabilities focused on marine species. While use of AI and autodetection capabilities have proven successful (Pirotta et al., 2022), there are a several limitations to this work. The first being detectability of marine specimens using machines such as the RTT^®110. Depending on how species are trafficked, some items may not be visible through CT scanning due to low density and/or size of individuals. It is likely that trafficked marine species will vary in size and arrangement, and therefore algorithms may at times miss detecting trafficked items. This limitation also reinforces this technology as a complimentary means of detection, and no replacement for manual human-based inspection methods already in place at international borders. The overall false alarm rate (PFA) of 13% currently achieved on this data set still requires operators to clear the false alarms. Physical inspections are crucial to support ongoing detection efforts for MWT as highlighted in shark and seahorse seizures (Partin et al., 2022; Foster et al., 2025). In addition, this study was only able to source a small number of physical samples for algorithm training, and used a higher number of TIP-generated bags. We acknowledge that this may limit algorithm's generalization ability and may not be reflective or as effective in real world deployments if information used to detect marine wildlife is limited to a small sample size. This further highlights a need for more widespread sampling of marine life to support the production of more powerful marine life detection algorithms.

Secondly, the use of 3D scanning capabilities is costly and not implemented in all detection routes around the world. Whilst 3D is currently considered the gold standard of frontline detection for security screening at international borders, the existence of 2D scanning remains in place as the basis for some wildlife trade detection capabilities; e.g., detecting the trade of rhinoceros horn, Project Vikela using 2D to scan rhino horn in Africa (United for Wildlife, 2025; Moloney and Chaber, 2024). This means, the MWT algorithms presented in this paper are inaccessible to countries still using 2D machines.

5 Conclusions

This study successfully demonstrates the application of world-first marine wildlife detection algorithms, developed with the purpose of identifying trafficking of shark fin, seahorse and sea cucumbers. This research highlights the significance of autodetection of illegal wildlife trade as a complementary method needed for effective implementation of CITES. Scaling-up IWT detection capabilities into existing systems around the world presents an opportunity for CITES Parties to better implement and enforce the text of the Convention toward ‘sustainable, legal, traceable trade' (CITES, 2025).

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher's note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/focsu.2026.1776978/full#supplementary-material

Supplementary material Supplementary material at the source

Footnotes

Edited by: Victoria Ndinelago Erasmus, International University of Management, Namibia

Reviewed by: Yagnesh Challagundla, University of Florida, Gainesville, FL, United States

Halil Ibrahim Senol, Harran University, Türkiye