Back Home: A Machine Learning Approach to Seashell Classification and Ecosystem Restoration

In Costa Rica, an average of 5 tons of seashells are extracted from ecosys- tems annually. Confiscated seashells, cannot be returned to their ecosystems due to the lack of origin recognition. To address this issue, we developed a convolutional neural network (CNN) specifically for seashell identifica- tion. We built a dataset from scratch, consisting of approximately 19000 images from the Pacific and Caribbean coasts. Using this dataset, the model achieved a classification accuracy exceeding 85%. The model has been integrated into a user-friendly application, which has classified over 36,000 seashells to date, delivering real-time results within 3 seconds per image. To further enhance the system’s accuracy, an anomaly detection mechanism was incorporated to filter out irrelevant or anomalous inputs, ensuring only valid seashell images are processed. 1 Introduction Seashells are essential components of ecosystems, playing a vital role in maintaining ecological balance. As highlighted by Cheng et al. (2023) [3], seashells possess .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 13, 2025. ; https://doi.org/10.1101/2025.01.08.632036doi: bioRxiv preprint unique structural and functional characteristics, such as high hardness, toughness, corrosion resistance, and bioactivity, which make them crucial for various ecological and industrial applications. Many marine organisms rely on the work of seashells for their survival. Therefore, preserving seashells should be a matter of concern for everyone. However, in recent years, a significant decline in seashells has been observed due to tourists collecting them from beaches in Costa Rica. Fortunately, these seashells are often confiscated at the Juan Santamaria Airport (Costa Rica’s na- tional airport), but they cannot be returned because their exact ecosystem of origin (Pacific or Caribbean) is unknown. To address this issue, this project developed a classification model capable of detecting the origin of each species from a single image. A total of 516 species registered in Costa Rica were analyzed, grouped into Pacific and Caribbean cat- egories to perform binary classification. This resulted in a model suitable for returning confiscated seashells at the Juan Santamaria Airport to their respective ecosystems. 2 Related Work Several studies have explored the classification of seashells and related marine ob- jects, focusing on aspects such as feature extraction, classification methods, and dataset development. However, these approaches differ from our work, as they pri- marily emphasize individual feature extraction or taxonomic family classification rather than ecosystem-level insights. Xue et al. (2021) addressed the challenge of identifying deep-sea debris using deep convolutional neural networks. They introduced the DDI dataset, compris- ing real deep-sea images, and developed the Shuffle-Xception network. This model incorporated innovative strategies, such as separable convolutions, group convo- lutions, and channel shuffling, to enhance classification accuracy [22]. Zhang et al. (2019) introduced a pioneering shell dataset containing 7,894 species with over 59,000 images, facilitating research in feature extraction and recognition. They applied traditional machine learning techniques, such as k-NN and random forest, to validate features like color, shape, and texture extracted from the dataset [23]. Yue et al. (2023) proposed ”FLNet”, a convolutional neural network designed to address challenges in shellfish recognition, such as high feature similarity and imbalanced datasets. Their framework included innovative mechanisms for filter pruning and repairing to enhance feature representation, alongside a hybrid loss function tailored for unbalanced datasets. [24] 2 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 13, 2025. ; https://doi.org/10.1101/2025.01.08.632036doi: bioRxiv preprint 3 Our Work The classification model was designed to address the critical challenge of determin- ing the origin of confiscated seashells (Pacific or Caribbean). Due to the specificity of this task, we were required to build a comprehensive dataset from scratch, en- compassing nearly every seashell species found in Costa Rica. This effort resulted in the creation of the first seashell database that integrates species from both the Pacific and Caribbean coasts of the country. By leveraging advancements in convolutional neural network (CNN) architec- tures and integrating anomaly detection mechanisms, this model ensures high ac- curacy and reliability for real-world deployment. A user-friendly web application was also developed to facilitate the classification process (see Appendix A). This application enables users to upload images of confiscated seashells and obtain real- time predictions regarding their ecosystem of origin. The app was designed with simplicity in mind, ensuring accessibility for a wide range of users. This section outlines the data preparation, model selection, and training strate- gies employed to achieve the desired goal. 3.1 Data Collection and Categorization The dataset used for this research comprises 19,058 images representing 516 species of seashells, collected over 10 months. From the total of images, 9,553 are from the Caribbean and 9,505 from the Pacific. All of these species were categorized into four distinct groups based on their taxonomy and ecosystem (see Appendix B) Category Pacific (Species) Caribbean (Species) Gastropods 130 149 Bivalves 107 130 Table 1: Comparison of species counts for Gastropods and Bivalves in the Pacific and Caribbean regions. To construct the dataset, we curated a comprehensive list of species by referenc- ing specialized resources such as Conchology Inc. [11], iNaturalist [10], the Florida Institute of Technology [14], and ConchyliNet [4]. While other seashell datasets exist for image classification [13,23] they are not tailored to the unique biodiversity of our ecosystems. Therefore, we were required to source specific images for each species, capturing their distinct characteristics and possible variations to ensure accurate representation. This categorization resulted in a total of 279 species from the Caribbean and 237 species from the Pacific. Each seashell was grouped by family, genus, and 3 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 13, 2025. ; https://doi.org/10.1101/2025.01.08.632036doi: bioRxiv preprint species to ensure accurate taxonomic representation. A detailed list of all catego- rized species is provided in the Appendix for reference. The dataset was metic- ulously verified to ensure the quality and consistency of the images, incorporat- ing diverse positions, backgrounds, and environmental conditions to enhance the model’s robustness. Figure 1: Representative Pacific seashells from the dataset. Figure 2: Representative Caribbean seashells from the dataset. The dataset was divided into three subsets for training, validation, and testing, with 70% allocated to the training set for model training, 15% to the validation set for hyperparameter tuning and performance evaluation during training, and 15% to the test set for evaluating the final model performance on unseen data. This split ensured a balanced representation of families across all subsets, supporting a rigorous and reliable evaluation of the model’s performance. All images were 4 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 13, 2025. ; https://doi.org/10.1101/2025.01.08.632036doi: bioRxiv preprint resized to 224x224 pixels to maintain consistent input dimensions for the model architecture, while preserving key visual features necessary for classification. To enhance the model’s ability to generalize and handle variations in real- world data, several data augmentation techniques were implemented. These in- cluded random rotations (±20 degrees), horizontal flips, and subtle adjustments to brightness and contrast. Additionally, random cropping and zoom operations were applied to simulate different viewing distances and perspectives. These augmen- tation techniques effectively expanded the training dataset and helped the model become more robust to natural variations in species appearances across different lighting conditions and viewing angles. These categorizations were critical in train- ing the final classification model, as they allowed the system to learn distinctions not only between seashells but also between ecosystems. By leveraging this struc- tured grouping, the model could focus on subtle morphological, color, and texture variations that characterize species from different ecosystems. 3.2 Model Training The final classification model was built using the ConvNext architecture [12]. We deliberately excluded Vision Transformers (ViT) from our experimentation due to their substantially larger model size and computational requirements, which would have hindered deployment in resource-constrained environments. We experimented with several established architectures including ResNet50 [7], DenseNet121 [8], and MobileNetV2 [18]. Despite extensive training, these models consistently achieved accuracy scores below 81%, even when using identical training configurations (Ta- ble 5 - Appendix A). ConvNext demonstrated superior performance, which we attribute to its 7×7 kernel design that effectively captures global features without relying on computa- tionally expensive attention mechanisms. This choice not only improved accuracy but also reduced the model’s computational overhead, making it more scalable when deployed for many simultaneous users as we did during the classification process. To train the model we used the ConvNext-Tiny version from PyTorch library with ImageNet-1K weights. The following hyperparameters were used: The learn- ing rate was set to 0.01, and the optimizer used was Stochastic Gradient Descent (SGD) with a momentum of 0.9 and a weight decay of 0.001. The batch size was 16, and the model was trained for 100 epochs. Starting from epoch 65, the learning rate was decayed by a factor of e−0.05. Thirty layers were unfrozen during training, allowing for fine-tuning of the deeper layers, while a final dropout layer with a rate of 0.3 was applied to reduce overfitting. 5 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 13, 2025. ; https://doi.org/10.1101/2025.01.08.632036doi: bioRxiv preprint Figure 3: Classification process utilizing a modified ConvNext Tiny architecture 3.3 Anomaly Detection System Anomaly systems have been broadly studied over recent years, especially for Large Language Models (LLMs). They have shown promising results in filtering out inputs that differ substantially from labeled datasets used for training [19–21]. Image analysis in real-world scenarios often requires systems that can identify when an input cannot be reliably assigned to any of the known classes, which is crucial in classification or detection tasks deployed in practical settings. Deep learning has emerged as a key enabler for such anomaly detection systems, such as Autoencoders and Generative Adversarial Networks to learn latent representations of normal data. By modeling what is typical, these methods can highlight inputs that fall outside the learned distribution. [1,17] Other approaches employ neural networks to estimate elements such as depth, perform feature extraction, or localize specific regions within an image, thereby providing a representation of the anomalous portion [2, 15]. Also, certain ap- proaches emphasize feature extraction and are thus more aligned with our goal: to obtain a robust image representation and subsequently determine whether an input is anomalous relative to a trained reference dataset. [5,6,16] Our approach focuses on using a vector representation of each training image and storing these vectors in a vector database. This enables efficient comparison against images provided by volunteers, resulting in a more accurate classification based on their similarities. This method ensures that only images containing the seashell as the primary element of interest are selected, reducing the impact of extraneous noise that could adversely affect the classification. Such an approach is particularly beneficial for a web application utilized by a wide range of users, many of whom may lack professional or scientific photography skills. This system identifies images that do not correspond to seashells, preventing incorrect classifications during the real-time volunteering. The process involves creating image embeddings using a convolutional neural network (SqueezeNet1.0) [9]. Each image was transformed into a vector of 1,000 scalar values, representing 6 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 13, 2025. ; https://doi.org/10.1101/2025.01.08.632036doi: bioRxiv preprint a compact and discriminative feature set. The embeddings were stored in a vector database to facilitate efficient similarity comparisons. When a new image is provided, its embedding is generated and compared against the stored embeddings using cosine similarity. The anomaly detection system operates as follows: 1. Compute the cosine similarity between the new image embedding and the top k closest embeddings in the database. 2. Average the similarity scores. If the resulting score is below a threshold (λ), the image is classified as an anomaly. Category Images Below Threshold* Cats 10/10 People 7/10 Buildings 8/10 Cars 10/10 Trees 10/10 Rooms 9/10 Cows 9/10 Hospitals 8/10 Horses 9/10 Dogs 10/10 Backgrounds 10/10 Ships 9/10 Birds 9/10 Frogs 7/10 Trucks 10/10 Airplanes 9/10 Reptiles 6/10 Electronic Devices 9/10 Insects 8/10 Seashells 0/40 Table 2: Anomaly Detection Performance Across Object Categories (n=10 images per category except seashells with n=40). * Threshold score = 0.955. Images with mean similarity scores below threshold are classified as anomalies. The cosine similarity between two embeddings, u and v, is defined as: cosine similarity(u, v) = u · v ∥u∥∥v∥ 7 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 13, 2025. ; https://doi.org/10.1101/2025.01.08.632036doi: bioRxiv preprint For anomaly detection, let Einput represent the embedding of the input image and {E1, E2, . . . , Ek} represent the embeddings of the top k closest images in the database. The average similarity score (S ) is calculated as: S = 1 k kX i=1 Einput · Ei ∥Einput∥∥Ei∥ If S < λ, where λ is a predefined threshold, the image is classified as an anomaly: Classification = ( Valid Image if S ≥ λ Anomaly if S < λ This method ensures that only valid seashell images are processed by the clas- sification system, improving overall accuracy and reliability. The threshold λ was determined empirically by analyzing the distribution of similarity scores between known seashell images in our dataset. By computing clusters based on the similar- ities between images within the same class (either Pacific or Caribbean), we found that legitimate seashell images typically exhibited similarity scores above 0.955, which we established as our anomaly threshold. Figure 4: Embedding generation from images using the SqueezeNet architecture 4 Results The classification model was evaluated on a test dataset containing 2,300 images, demonstrating its effectiveness in identifying the origin (Pacific or Caribbean) of seashells. The results show a general accuracy of 86.28%, with slight variations between the two locations. These metrics highlight the robustness of the model in real-world applications. 4.1 Results by Location The performance of the model for each location is as follows: 8 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 13, 2025. ; https://doi.org/10.1101/2025.01.08.632036doi: bioRxiv preprint • Pacific: The model achieved an accuracy of 85.43%, demonstrating strong performance in classifying seashells from the Pacific coast. • Caribbean: The model achieved an accuracy of 87.10%, showing slightly better performance compared to the Pacific classification. Table 3: Performance Metrics on Test Dataset (3,363 Images) Location Accuracy (%) Precision (%) Recall (%) F1 (%) Pacific 85.43 88.70 85.43 87.00 Caribbean 87.10 83.45 87.10 85.29 Overall 86.28 86.23 86.13 86.17 The results highlight the model’s reliability in classifying seashells from both the Pacific and Caribbean coasts of Costa Rica. The slight performance variation between the two locations may be attributed to differences in image diversity, environmental factors, or species representation in the dataset. Even, with that accuracies, the model can be used in productive environments with security, as can be able to classify correctly 8 of 10 seashells, 5 Future Work Future research should be focus on extending the model to other ecosystems by using the already trained weights. This approach could facilitate adaptation to new environments, improving the model’s versatility while reducing the need for extensive retraining. Additionally, enhancing the anomaly detection system for robustness and adaptability to diverse input conditions remains a promising direc- tion. 6 Conclusions This research successfully developed a robust classification model capable of iden- tifying the origin of seashells from both Costa Rica’s Pacific and Caribbean coasts, effectively handling specimens from 515 different families. The results highlight the potential of machine learning to support ecological restoration efforts by facil- itating the proper return of confiscated seashells to their native ecosystems. In addition, we present a new dataset focus on seashell recognition from Pacific and Caribbean that can be used in future scientific research, we set a benchmark of 86.28% on the test dataset, that hopefully can be surpass by other scientists. This dataset not only served as the foundation for model development, but also 9 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 13, 2025. ; https://doi.org/10.1101/2025.01.08.632036doi: bioRxiv preprint represents a valuable resource for various institutions, enabling further research and advancements in marine biology and conservation. Furthermore, the incorporation of an anomaly detection mechanism signifi- cantly enhanced the classification system’s reliability. Through effective filter- ing of anomalous and invalid inputs, the system maintained high accuracy levels and exhibited improved robustness when deployed in real-world environments. This implementation underscores the importance of developing systems capable of identifying and filtering out low-quality data before it affects the performance of production models. 7 Acknowledgements We gratefully acknowledge the funding and help provided by FIFCO and its brand, Imperial. Their contribution was very important during this research and ensured its successful completion. Additionally, we wish to thank Dr. Yolanda Camacho for her invaluable work and dedication in compiling the lists of seashells. We also express our gratitude to the Sistema Nacional de ´Areas de Conservaci´ on (SINAC), Universidad de Costa Rica (UCR), and AERIS for their support throughout this process. A Web Application Details To make the seashell classification model accessible to a broad audience, a user- friendly web application was developed. This application allows users to upload images of confiscated seashells and receive real-time predictions regarding their origin (Pacific or Caribbean). The interface was designed to ensure usability for non-experts, enabling seam- less interaction and accurate results within seconds. The application integrates the classification model and anomaly detection system described earlier in this docu- ment, providing a robust platform for real-world use. It supports images captured in various conditions, ensuring flexibility and reliability. The key features of the application include: • Image Upload: Users can upload single or multiple images of confiscated seashells. • Real-Time Classification: The app delivers predictions in under three seconds per image, indicating the ecosystem of origin (Pacific or Caribbean). • User-Friendly Design: The interface is intuitive and accessible to a wide range of users, requiring no prior technical expertise. 10 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 13, 2025. ; https://doi.org/10.1101/2025.01.08.632036doi: bioRxiv preprint Figure 5: The web application for seashell classification, enabling real-time pre- dictions. Table 4: Summary of Model Results and Parameters Mo del Version unfreeze la yers Optimizer learning rate n um ep ochs lr ep och T est Accuracy (%) Con vNext Tiny 0 SGD 0.01 50 25 83.24 Con vNext Tiny 0 SGD 0.01 50 25 81.11 Con vNext Tiny 10 SGD 0.01 100 50 80.15 Con vNext Tiny 10 SGD 0.01 150 75 84.56 Con vNext Tiny 10 SGD 0.01 250 125 84.16 Con vNext Tiny 13 SGD 0.01 150 75 84.16 Con vNext Tiny 30 SGD 0.01 100 65 86.28 The Test Accuracy (%) column reflects the model’s performance on the test dataset. Notably, the final version of the model achieved the highest accuracy of 86.28%, obtained by unfreezing 30 layers, training for 100 epochs, and applying a learning rate decay starting at epoch 65. Architecture Accuracy (%) ResNet50 78.3 ± 0.4 DenseNet121 80.2 ± 0.3 MobileNetV2 79.3 ± 0.5 Table 5: Accuracy scores (mean ± standard deviation) for models over 50 runs 11 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 13, 2025. ; https://doi.org/10.1101/2025.01.08.632036doi: bioRxiv preprint B List of Species Table 6: Seashell Species - Caribbean Gastropoda Species Acmaeidae Lottia antillarum, Acteonidae rictaxis punctostriatus, Architectonici- dae Philippia krebsi, Areneidae Arene cruentata, Buccinidae Hesperisternia karinae, Buccinidae Pollia auritula, Buccinidae pisania auritula, Bullidae Bulla punctulata, Bullidae bulla mabillei, Bullidae bulla striata, Bursidae Bursa cubaniana, Bursidae bursa thomae, Calliostomatidae Calliostoma jujubinum, Calyptraeidae Bostrycapulus aculeatus, Calyptraeidae Crepidula convexa, Calyptraeidae Ergaea walshi, Calyptraeidae crepidula aculeata, Cassidae Cypraecassis testiculus, Cassidae Semicassis granulata, Cassidae cassis tuberosa, Cerithiidae Cerithium eburneum, Cerithiidae cerithium guinaicum, Cerithiidae cerithium litteratum, Cerithiidae cerithium lutosum, Charoniidae Charonia variegata, Columbellidae columbella mercatoria, Columbellidae mazatlania fulgurata, Columbellidae mitrella ocellata, Columbellidae nitidella laevigata, Columbellidae nitidella nitida, Columbellidae parametaria ovulata, Conidae Conasprella mindana, Conidae Conus cardinalis, Conidae Conus mus, Conidae Conus regius, Conidae Conus spurius lorenzianus, Conidae conus daucus, Conidae conus jaspideus, Conidae conus spurius phlogopus, Cymatiidae Monoplex nicobaricus, Cypraeidae Luria cinerea, Cypraeidae Macrocypraea zebra, Cypraeidae Naria aci- cularis, Cypraeidae cypraea acicularis, Cypraeidae cypraea cinerea, Cypraeidae cypraea zebra, Cystiscidae persicula interruptolineata, Ellobiidae Melampus coffea, Elysiidae elysia ornata, Fasciolariidae Fasciolaria tulipa, Fasciolariidae Hemipolygona carinifera, Fasciolariidae Poligona angulata, Fasciolariidae Polygona infundibulum, Fasciolariidae latirus angulatus, Fasciolariidae leucozonia nassa, Fasciolariidae leucozonia ocellata, Fissurellidae Diodora listeri, Fissurellidae Fis- surella barbadensis, Fissurellidae Hemitoma octoradiata, Fissurellidae diodora jaumei, 12 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 13, 2025. ; https://doi.org/10.1101/2025.01.08.632036doi: bioRxiv preprint Fissurellidae fissurella angusta, Fissurellidae fissurella fascicularis, Fissurellidae fissurella nodosa, Fissurellidae fissurella rosea, Fissurellidae hemitona octoradiata, Fissurellidae lucapina aegis, Fissurellidae lucapina sowerbii, Fissurellidae lucapina suffusa, Harpidae Morum oniscus, Hipponicidae hipponix antiquatus, Littorinidae Cenchritis muricatus, Littorinidae Echinolittorina ziczac, Littorinidae Littoraria angulifera, Littorinidae echinolittorina meleagris, Littorinidae echinolittorina tuberculata, Littorinidae littoraria tessellata, Lit- torinidae littorina ziczac, Littorinidae nodilittorina angustior, Marginellidae prunum holandae, Melon- genidae melongena melongena, Modulidae modulus modulus, Muricidae Stramonita floridiana, Muricidae Stra- monita rustica, Muricidae Vasula deltoidea, Muricidae chicoreus florifer, Muricidae chicoreus po- mum, Muricidae muricopsis deformis, Muricidae muricopsis oxytatus, Muricidae plicop- urpura patula, Muricidae thais haemastoma floridana, Muricidae thais rustica, Nassariidae Nas- sarius consensus, Nassariidae nassarius albus, Nassariidae phos antillarum, Naticidae Natica marochiensis, Naticidae Polinices hepaticus, Naticidae naticarius canrena, Naticidae polinices lacteus, Naticidae sinum perspectivum, Neritidae Nerita fulgurans, Neritidae Nerita peloronta, Neritidae nerita pelonronta, Neritidae nerita tessellata, Neritidae nerita versicolor, Neritidae vitta virginea, Olividae Oliva fulgurator, Olividae oliva reticularis, Olividae olivella minuta, Olividae olivella nivea, Ovulidae Cyphoma gibbosum, Ovulidae Cyphoma signatum, Phasianellidae eulithidium bellum, Phasianellidae tricolia tessellata, Pisaniidae Engina turbinella, Pisaniidae pisania pusio, Planaxidae Supplanaxis nucleus, Planaxidae planaxis nucleus, Pseudomelatomidae Crassispira harfordiana, Ranel- lidae Cabestana labiosa, Ranellidae Charonia tritonis, Ranellidae Monoplex nicobaricus, Ranellidae Mono- plex pileare, Ranellidae Monoplex vespaceus, Ranellidae Septa occidentalis, Rissoidae rissoina elegantissima, Rissoidae rissoina sagraiana, Strombidae Aliger gigas, Strombidae Lobatus rani- nus, 13 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 13, 2025. ; https://doi.org/10.1101/2025.01.08.632036doi: bioRxiv preprint Strombidae strombus pugilis, Tegulidae Agathistoma viridulum, Tegulidae Cittar- ium pica, Tegulidae tegula excavata, Terebridae hastula cinerea, Tonnidae tonna galea, Tonnidae tonna maculosa, Triviidae trivia nix, Triviidae trivia pediculus, Trochi- dae Calliostoma javanicum, Turbinellidae Turbinella angulata, Turbinellidae Volutella muricata, Turbinidae Astraea caelata, Turbinidae Lithopoma tectum, Turbinidae astraea phoebia, Turbinidae astraea tecta, Turbinidae turbo cailletii, Turbinidae turbo castanea, Vitrinellidae solariorbis corylus, Vitrinellidae vitrinella elegans, Volutidae voluta virescens Table 7: Seashell Species - Pacific Gastropoda Species Architectonicida Heliacus areola bicanaliculatus, Architectonicidae Architectonica karsteni, Architectonicidae Heliacus areola bicanaliculatus, Architectonicidae Heliacus cae- latus, Batillariidae Rhinocoryne humboldti, Bullidae Bulla punctulata, Bursidae Alanbeuella corrugata, Bursidae Bursa rugosa, Bursidae Dulcerana gran- ularis, Bursidae bufonaria rana, Bursidae bursa granularis, Calyptraeidae Bostrycapulus aculeatus, Calyptraeidae Calyptraea chinensis, Calyptraeidae Calyptraea conica, Calyptraeidae Crepidula lessonii, Calyptraeidae Crepidula marginalis, Calyptraeidae Crepidula striolata, Cassidae Cypraecassis coarctata, Cerithiidae Cerithium adustum, Cerithiidae Cerithium atromarginatum, Cerithiidae Cerithium browni, Cerithiidae Cerithium muscarum, Columbellidae Anachis boivini, Columbellidae Anachis lyrata, Columbellidae Anachis rugosa, Columbellidae Columbella haemastoma, Columbellidae Columbella labiosa, Columbellidae Columbella major, Columbellidae Columbella paytensis, Columbellidae Columbella socorroensis, Columbellidae Columbella strombiformis, Columbellidae euplica varians, Columbellidae mitrella elegans baiyeli, Columbellidae mitrella guttata, Columbellidae pyrene ocellata, Conidae Conus brunneus, Conidae Conus dalli, Conidae Conus fergusoni, 14 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 13, 2025. ; https://doi.org/10.1101/2025.01.08.632036doi: bioRxiv preprint Conidae Conus princeps, Conidae Conus regularis, Conidae conus chaldaeus, Conidae conus ebraeus, Conidae conus gladiator, Conidae conus nux, Conidae conus purpurascens, Conidae conus scalaris, Conidae conus tessulatus, Cymatiidae Monoplex gemmatus, Cymatiidae Monoplex pilearis, Cymatiidae Monoplex vestitus, Cymatiidae Monoplex wiegmanni, Cypraeidae Macrocypraea cervinetta, Cypraeidae Pseudozonaria arabicula, Cypraeidae Pseudozonaria robertsi, Cypraeidae cypraea cervinetta, Cypraeidae cypraea robertsi, Fasciolariidae Granolaria salmo, Fasciolariidae Leucozonia cerata, Fasciolariidae Triplofusus princeps, Fasciolariidae leucozonia rudis, Fasciolariidae opeatostoma pseudodon, Ficidae Ficus ventricosa, Fissurellidae fissurella virescens, Hipponicidae Antisabia panamensis, Hipponicidae Cheilea corrugata, Janthinidae janthina janthina, Littorinidae Echinolittorina aspera, Lottiidae Lottia fascicularis, Lottiidae Lottia filosa, Lottiidae Lottia mesoleuca, Melongenidae Melongena patula, Mitridae Neotiara lens, Mitridae Strigatella tristis, Modulidae Trochomodulus catenulatus, Muricidae Hexaplex brassica, Muricidae Hexaplex princeps, Muricidae Muricanthus radix, Muricidae Neorapana muricata, Muricidae Phyllonotus regius, Muricidae Plicopurpura collumelaris, Muricidae Plicopurpura columelaris, Muricidae Stramonita haemastoma, Muricidae Thaisella kiosquiformis, Muricidae Zetecopsis zeteki, Muricidae acanthais brevidentata, Muricidae cymia tectum, Muricidae stramonita biserialis, Muricidae thais tuberosa, Muricidae vasula melones, Nassariidae nassarius erythraeus, Nassariidae nassarius gemmuliferus, Naticidae Polinices uber, Naticidae mammilla simiae, Naticidae natica fasciata, Naticidae notocochlis chemnitzii, Neritidae Nerita funiculata, Neritidae nerita scabricosta, Olividae Agaronia nica, Olividae Agaronia propatula, Olividae Agaronia testacea, Olividae Oliva incrassata, Olividae Oliva polpasta, Olividae Oliva porphyria, Olividae Olivella volutella, Olividae Pachyoliva semistriata, Ovulidae Jenneria pustulata, Personidae Distorsio decussata, Pisaniidae Engina tabogaensis, Pisaniidae Gemophos ringens, Pisaniidae Hesperisternia vibex, Pisaniidae Solenosteira gatesi, Planaxidae Supplanaxis planicostatus, Planaxidae planaxis obsoletus, Siphonariidae Siphonaria gigas, Siphonariidae Siphonaria maura, Strombidae Lobatus peruvianus, Strombidae Persististrombus granulatus, Strombidae Strombus alatus, Strombidae Strombus gracilior, 15 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 13, 2025. ; https://doi.org/10.1101/2025.01.08.632036doi: bioRxiv preprint Strombidae Titanostrombus galeatus, Strombidae strombus granulatus, Tegulidae Tegula pellisserpentis, Tegulidae tegula panamensis, Tonnidae Malea ringens, Triviidae trivia sanguinea, Turbinellidae vasum caestus, Turbinidae Arene olivacea, Turbinidae Turbo saxosus, Turbinidae Uvanilla buschii, Turritellidae Caviturritella leucostoma Table 8: Seashell Species - Caribbean Bivalves Species Arcidae Anadara brasiliana, Arcidae Anadara chemnitzii, Arcidae Anadara nota- bilis, Arcidae Anadara transversa, Arcidae Arca imbricata, Arcidae Arca zebra, Arcidae Barbatia cancellaria, Arcidae Barbatia candida, Arcidae Barbatia dominguensis, Arcidae Fugleria tenera, Arcidae Lamarcka imbricata, Cardiidae Acrosterigma magnum, Cardiidae Dallocardia muricata, Cardiidae Laevicardium pictum, Cardiidae Pa- pyridea semisulcata, Cardiidae Papyridea soleniformis, Cardiidae Trachycardium isocardia, Cardiidae Trachycardium magnum, Cardiidae Trachycardium muricatum, Carditidae Carditamera gracilis, Chamidae Arcinella arcinella, Chamidae Chama congregata, Chamidae Chama florida, Chamidae Chama macerophylla, Chamidae Chama sin- uosa, Chamidae Pseudochama cristella, Chamidae Pseudochama radians, Corbulidae Corbula caribaea, Corbulidae Corbula contracta, Corbulidae Juliacorbula aquivalvis, Cyrenidae Polymesoda arctata, Donacidae Donax denticulatus, Donacidae Donax striatus, Dreissenidae Mytilopsis sallei, Glycymerididae Glycymeris undata, Glycymerididae Tucetona pectinata, Isognomonidae Isognomon alatus, Isognomonidae Isognomon bicolor, Isog- nomonidae Isognomon radiatus, Limidae Ctenoides scaber, Limidae Lima caribaea, Limidae Lima caribea, Limidae Lima lima, Limidae Limaria pellucida, Lucinidae Anodontia alba, 16 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 13, 2025. ; https://doi.org/10.1101/2025.01.08.632036doi: bioRxiv preprint Lucinidae Callucina keenae, Lucinidae Clathrolucina costata, Lucinidae Codakia orbicularis, Lucinidae Ctena orbiculata, Lucinidae Divalinga quadrisulcata, Lucinidae Divaricella quadrisulcata, Lucinidae Lucinisca centrifuga, Lucinidae Phacoides pectinatus, Mactridae Mactrellona alata, Mactridae Mac- troma fragilis, Mactridae Mactrotoma fragilis, Mactridae Mulinia cleryana, Margaritidae Pinc- tada imbricata, Mytilidae Botula fusca, Mytilidae Brachidontes exustus, Mytilidae Lioberus cas- tanea, Mytilidae Modiolus americanus, Nuculanidae Adrana lancea, Nuculanidae Adrana tellinoides, Ostreidae Crassostrea rhizophorae, Ostreidae Crassostrea virginica, Ostreidae Dendostrea frons, Pectinidae Spathochlamys benedicti, Pectinidae Antillipecten antillarum, Pectinidae Argopecten gibbus, Pectinidae Argopecten irradians, Pectinidae Argopecten irradians amplicostatus, Pectinidae Caribachlamys ornata, Pectinidae Caribachlamys sentis, Pectinidae Chlamys ornata, Pectinidae Euvola laurentii, Petricolidae Petricola bi- color, Pinnidae Atrina seminuda, Pinnidae Pinna carnea, Plicatulidae Plicatula gibbosa, Psammobiidae Asaphis deflorata, Psammobiidae Psammotella cruenta, Pteriidae Pteria colymbus, Semelidae Semele proficua, Semelidae Semele purpuras- cens, Solecurtidae Solecurtus cumingianus, Solecurtidae Tagelus divisus, Spondylidae Spondylus butleri, Tellinidae Arcopagia fausta, Tellinidae Eurytellina angulosa, Tellinidae Eurytellina nitens, Tellinidae Eurytellina punicea, Tellinidae John- sonella fausta, Tellinidae Laciolina laevigata, Tellinidae Merisca cristallina, Tellinidae Psam- motreta brevifrons, Tellinidae Scissula similis, Tellinidae Serratina aequistriata, Tellinidae Strigilla carnaria, Tellinidae Strigilla dichotoma, Tellinidae Strigilla pisiformis, Tellinidae Strigilla pseudocarnaria, Tellinidae Tellina punicea, Tellinidae Tellina radiata, Tellinidae Tellinella listeri, Tellinidae Tellinella listeri, Ungulinidae Diplodonta punctata, Ungulinidae Phlyctiderma semiasperum, Veneridae Anomalocardia brasiliana, Veneridae Anomalocardia flexuosa, Veneridae Chione cancellata, Veneridae Chione intapurpurea, Veneridae Chione paphia, Veneridae Chionopsis intapurpurea, Veneridae Dosinia concentrica, Veneridae Globivenus rigida, 17 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 13, 2025. ; https://doi.org/10.1101/2025.01.08.632036doi: bioRxiv preprint Veneridae Gouldia cerina, Veneridae Hysteroconcha circinata, Veneridae Hysteroconcha dione, Veneridae Lamelliconcha circinatus, Veneridae Lirophora paphia, Veneridae Macrocallista maculata, Veneridae Megapitaria maculata, Veneridae Pitar albidus, Veneridae Pitar fulmi- natus, Veneridae Tivela mactroides, Veneridae Transennella cubaniana, Veneridae Transennella stimpsoni, Veneridae Ventricola rigida Table 9: Seashell Species - Pacific Bivalves Species Arcidae Acar gradata, Arcidae Acar rostae, Arcidae Anadara similis, Arcidae Anadara tuberculosa, Arcidae Arca pacifica, Arcidae Barbatia lurida, Arcidae Barbatia reeveana, Arcidae Lamarcka mutabilis, Arcidae Larkinia grandis, Arcidae Larkinia multi- costata, Cardiidae Acrosterigma pristipleura, Cardiidae Americardia biangulata, Cardiidae Americardia planicostata, Cardiidae Dallocardia senticosum, Cardiidae Laevicardium substriatum, Cardiidae Papyridea aspersa, Cardiidae Trachycardium procerum, Carditidae Cardita crassicosta, Carditidae Carditamera affinis, Carditidae Carditamera radiata, Carditidae Cardites crassicostatus, Carditidae Cardites laticostatus, Carditidae Strophocardia megastropha, Chamidae Chama buddiana, Chamidae Chama coralloides, Chamidae Chama echinata, Corbulidae Caryocorbula amethystina, Corbulidae Caryocorbula biradiata, Corbulidae Caryocorbula nasuta, Corbulidae Caryocorbula ovulata, Crassatellidae Eucrassatella gibbosa, Cyrenoididae Polymesoda inflata, Donacidae Donax carinatus, Donacidae Donax dentifer, Donacidae Iphigenia altior, Glycymerididae Axinactis delessertii, Glycymerididae Axinactis inaequalis, Glycymerididae Tucetona multicostata, Gryphaeidae Hyotissa hyotis, Isog- nomonidae Isognomon recognitus, Limidae Limaria tetrica, Lucinidae Codakia distinguenda, Lucinidae Ctena gala- pagana, 18 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 13, 2025. ; https://doi.org/10.1101/2025.01.08.632036doi: bioRxiv preprint Lucinidae Ctena mexicana, Lucinidae Divalinga eburnea, Mactridae Harvella ele- gans, Mactridae Mactrellona clisia, Mactridae Mactrellona exoleta, Mactridae Mactrel- lona subalata, Mactridae Mulinia pallida, Mytilidae Brachidontes puntarenensis, Mytilidae Leiosolenus aristatus, Mytilidae Leiosolenus plumula, Mytilidae Modiolus capax, Mytilidae Mytella guyanensis, Noetiidae Noetia reversa, Ostreidae Crassostrea columbiensis, Ostreidae Cras- sostrea corteziensis, Ostreidae Crassostrea gigas, Ostreidae Saccostrea palmula, Ostreidae Striostrea prismatica, Pectinidae Argopecten ventricosus, Pectinidae Nodipecten subnodosus, Pinnidae Atrina maura, Pinnidae Pinna rugosa, Psammobiidae Gari helenae, Psammobiidae Heterodonax pacificus, Psammobiidae Sanguinolaria tellinoides, Pteriidae Pinctada mazatlanica, Pteri- idae Pteria sterna, Semelidae Semele bicolor, Semelidae Semele elliptica, Semelidae Semele formosa, Semelidae Semele purpurascens, Semelidae Semele verrucosa, Solecurtidae Tagelus affinis, Solecurtidae Tagelus peruanus, Solecurtidae Tagelus peruvianus, Spondylidae Spondylus limbatus, Tellinidae Eurytellina regia, Tellinidae Iridona subtrigona, Tellinidae Psammotreta pura, Tellinidae Strigilla chroma, Tellinidae Strigilla dichotoma, Tellinidae Strigilla dis- juncta, Tellinidae Strigilla serrata, Ungulinidae Zemysina subquadrata, Veneridae Chione subimbricata, Veneridae Cyclinella producta, Veneridae Cyclinella subquadrata, Veneridae Dosinia dunkeri, Veneridae Dosinia ponderosa, Veneridae Hysteroconcha lupanaria, Veneridae Hys- teroconcha multispinosus, Veneridae Hysteroconcha roseus, Veneridae Iliochione subrugosa, Veneridae Lamelliconcha tortuosus, Veneridae Lamelliconcha unicolor, Veneridae Leukoma asperrima, Veneridae Leukoma ecuadoriana, Veneridae Leukoma grata, Veneridae Leukoma histrionica, Veneridae Megapitaria aurantiaca, Veneridae Megapitaria squalida, Veneridae Periglypta multicostata, Veneridae Tivela byronensis, 19 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 13, 2025. ; https://doi.org/10.1101/2025.01.08.632036doi: bioRxiv preprint Veneridae Tivela planulata

[1] Jinwon An and Sungzoon Cho. Variational autoencoder based anomaly de- tection using reconstruction probability. Technical report, SNU Data Mining Center, 2015. Technical Report. [2] Liron Bergman and Yedid Hoshen. Classification-based anomaly detection for general data. In International Conference on Learning Representations (ICLR), 2020. [3] Meiqi Cheng, Man Liu, Lirong Chang, Qing Liu, Chunxiao Wang, Le Hu, Ziyue Zhang, Wanying Ding, Li Chen, Sihan Guo, Zhi Qi, Panpan Pan, and Jingdi Chen. Overview of structure, function and integrated utilization of marine shell. Science of The Total Environment, 870:161950, 2023. [4] ConchyliNet. Seashell database, n.d. Accessed: 2024-12-26. [5] Thomas Defard, Aleksandr Setkov, Angelique Loesch, and Romaric Audigier. Padim: a patch distribution modeling framework for anomaly detection and localization, 2020. [6] Thomas Defard, Alex Setkov, Angeline Loesch, and Romaric Audigier. PaDiM: a patch distribution modeling framework for anomaly detection and localization. In International Conference on Pattern Recognition (ICPR), 2021. [7] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770–778, 2016. [8] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In Proceedings of the IEEE Con- ference on Computer Vision and Pattern Recognition, pages 4700–4708, 2017. [9] Forrest N. Iandola, Song Han, Matthew W. Moskewicz, Khalid Ashraf, William J. Dally, and Kurt Keutzer. Squeezenet: Alexnet-level accuracy with 50x fewer parameters and ¡0.5mb model size, 2016. [10] iNaturalist. inaturalist: Connect with nature, n.d. Accessed: 2024-12-26. [11] Conchology Inc. Welcome to conchology, inc., n.d. Accessed: 2024-12-26. 20 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 13, 2025. ; https://doi.org/10.1101/2025.01.08.632036doi: bioRxiv preprint [12] Z. Liu, H. Mao, C.-Y. Wu, C. Feichtenhofer, T. Darrell, and S. Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11976–11986, 2022. [13] Matthewacs. South florida atlantic sea shell types, n.d. Accessed: 2024-12-26. [14] Florida Institute of Technology. Marine & environmental systems: Mollusk shell collection, n.d. Accessed: 2024-12-26. [15] Kevin Roth, Raghavendra Setuka, Julian Ibarz Zepeda, Yannic Kilcher, Thomas Hofmann, and Bj¨ orn Ommer. Towards total recall in industrial anomaly detection. In IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR), 2022. [16] Lukas Rudolph, Bastian Wandt, and Bodo Rosenhahn. Fully convolu- tional cross-scale-flow for image-based anomaly detection and localization. In IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2022. [17] Mayu Sakurada and Takashi Yairi. Anomaly detection using autoencoders with nonlinear dimensionality reduction. In Proceedings of the 2nd Workshop on Machine Learning for Sensory Data Analysis (MLSDA), in conjunction with the 20th ACM SIGKDD International Conference on Knowledge Discov- ery and Data Mining (KDD), pages 4:4–4:11. ACM, 2014. [18] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhmoginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recog- nition, pages 4510–4520, 2018. [19] Weihang Su, Yichen Tang, Qingyao Ai, Changyue Wang, Zhijing Wu, and Yiqun Liu. Mitigating entity-level hallucination in large language models. arXiv preprint arXiv:2407.09417, 2024. [20] Weihang Su, Changyue Wang, Qingyao Ai, Yiran Hu, Zhijing Wu, Yujia Zhou, and Yiqun Liu. Unsupervised real-time hallucination detection based on the internal states of large language models. arXiv preprint arXiv:2403.06448, 2024. [21] S. M. Towhidul Islam Tonmoy, S. M. Mehedi Zaman, Vinija Jain, Anku Rani, Vipula Rawte, Aman Chadha, and Amitava Das. A comprehensive survey of hallucination mitigation techniques in large language models. arXiv preprint arXiv:2401.01313, 2024. 21 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 13, 2025. ; https://doi.org/10.1101/2025.01.08.632036doi: bioRxiv preprint [22] Bing Xue, Baoxiang Huang, Ge Chen, Haitao Li, and Weibo Wei. Deep-sea debris identification using deep convolutional neural networks. IEEE Jour- nal of Selected Topics in Applied Earth Observations and Remote Sensing, 14:8909–8915, 2021. [23] Qi Zhang, Jianhang Zhou, Jing He, Xiaodong Cun, Shaoning Zeng, and Bob Zhang. A shell dataset for shell features extraction and recognition. Scientific Data, 6(226), 2019. [24] Yang Zhang, Jun Yue, Aihuan Song, Shixiang Jia, and Zhenbo Li. A high- similarity shellfish recognition method based on convolutional neural network. Information Processing in Agriculture, 10(2):149–163, 2023. 22 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 13, 2025. ; https://doi.org/10.1101/2025.01.08.632036doi: bioRxiv preprint .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 13, 2025. ; https://doi.org/10.1101/2025.01.08.632036doi: bioRxiv preprint .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 13, 2025. ; https://doi.org/10.1101/2025.01.08.632036doi: bioRxiv preprint .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 13, 2025. ; https://doi.org/10.1101/2025.01.08.632036doi: bioRxiv preprint .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 13, 2025. ; https://doi.org/10.1101/2025.01.08.632036doi: bioRxiv preprint .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 13, 2025. ; https://doi.org/10.1101/2025.01.08.632036doi: bioRxiv preprint .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 13, 2025. ; https://doi.org/10.1101/2025.01.08.632036doi: bioRxiv preprint