A Large-Scale Convolutional Neural Network for Fine-Grained Conus Shell Identification

Introduction

The genus Conus (commonly known as cone snails) represents one of the most diverse groups of marine gastropods, with a current total of 853 recognized extant species according to MolluscaBase and WoRMS (WoRMS, MolluscaBase). The Conus genus is one of the largest in the Mollusca phylum. Until a decade ago, its species were split across 89 genera [1], but are now largely consolidated within Conus genus (WoRMS, MolluscaBase). Despite this tremendous species richness, cone snails exhibit a notable uniformity in general shell shape and pattern, which can differ only by subtle color variations, banding, or small morphological traits. As a result, discriminating among the many Conus species becomes a daunting fine-grained image classification challenge, demanding models that can isolate and interpret minute differences in shell markings.
A Conus CNN model is designed to learn features that capture these nuances, allowing it to separate species based on visual cues in shell images. However, the complexity of this task escalates with the number of species (i.e., classes) included: each new class introduces additional inter-class similarities and increases the potential for misclassification [2]. Moreover, adding classes means collecting and processing more images, thus necessitating greater computational resources and longer training times. For all Conus species together we have a dataset of 130 373 images. Consequently, scaling a Conus CNN model to encompass all 853 species underscores both the difficulty of fine-grained recognition and the importance of robust, efficient training strategies.

The shells of the Conus genus have a characteristic morphology, the most important features for species identification are pattern and colouration. Other important features are the form of the spire and the width versus the length.

Methods

Data Collection

The dataset for the Conus CNN model comprises 130,373 shell images representing 518 Conus species (see table I). From the 529 species for which images were collected, all species with less than 25 images were removed (see Minimum number of images needed for each species). A total of 518 species were used. These images were aggregated from multiple sources, including online databases and museum or field photograph repositories (Identifying Shells using Convolutional Neural Networks: Data Collection and Model Selection). Additionally, broad community-driven efforts (e.g. citizen science platforms) have contributed to the pool of images – modern biodiversity projects have amassed massive image collections of specimens. The dataset comprises images from the following sources: museum collections (4.3%, 5640 images), online citizen science platforms (18.1%, 23661 images) and commercial shell websites (77.6%, 101071 images). The original Conus image dataset is 12% larger, many images were removed because the image quality is bad, or other objects are visible in the picture (hands, other animals, labels, etc.). Also images that contains more than one shell and could not be split in images with only 1 shells were eliminated.

Hardware

For training a HP Omen 30L GT13 is used. It contains a Intel(R) Core(TM) i9-10850K CPU @ 3.60GHz 3.50 GHz processor, with 64GB RAM, Nvidia GeForce RTX 3080 10GB.

Image preparation

All images were pre-processed. If more than 1 shell was visible on the picture, several images were generated, 1 image for each shell. The background was replaced with a uniform black background. A square image was made by padding the black background. All shells were resized (400 x 400 px). A last visual selection was made before producing the final image dataset. Overall, 10-20% of the images were removed for various reasons (when other objects were visible in the picture such hands, habitat, text, etc.).

Annotation and Labeling Challenges

Preparing a labeled dataset of 518 species presents significant annotation hurdles. One major challenge is taxonomic ambiguity. Cone snail taxonomy has been in flux – historically cone snails were split in 89 genera [1], but the last decade most species were merged in the genus Conus (see MolluscaBase/WoRMS). As a result, the same species might be known by multiple names, or what were once separate species might have been merged. Such inconsistencies across image sources can lead to mislabeling (e.g., an image labeled with an outdated name). Careful curation was needed to reconcile synonyms and ensure each image is tagged with a valid, accepted species name.

Another challenge is the morphological similarity among species: many Conus shells differ only in subtle pattern or color variations. Non-experts may confuse one species for another, especially if shell patterns overlap or the specimen is an atypical individual. This means some portion of the images could be erroneously labeled, introducing noise into the training data.

Metrics and confidence intervals

Metrics were calculated using the sklearn.metrics module, functions accuracy_score, precision_score, recall_score, f1_score, confusion_matrix, classification_report were used. To calculate the confidence intervals (95%). A maximum of 200 images were used to calculate the metrics and confidence intervals. Bootstrapping was used [4]. Bootstrapping, being a non-parametric method, does not rely on the normality assumption. A 1000 runs were performed for each species to calculate the intervals.

Results

The Conus image dataset

From these 529 species for which images are available, 518 have more than 25 images which are used to construct a model (see Minimum number of images needed for each species). These 518 species and the number of images used are listed in Table II. The distribution of images among species is shown in the next figure.

Many species have a low number of images. The first bin (25-170 images) has 293 species which is almost half of all species. There are a few species with a large amount of images; Conus textile has 2923 images, Conus furvus 1976 images and Conus mercator 1703 images. There is a considerable imbalance in the dataset. However, preliminary tests with oversampling the minority classes with augmented images (rotate, flip, change brightness and contrast) does not improve the results significantly (data not shown).

Image Preparation

The dataset of 130,373 images were split into 80% training and 20% testing data. This means that the species with the least images (total of 25 images) has 20 images in the training dataset. There will be more than 20 images if the original image shows several shells (or views of the same shell) because separate images were made for each shell in the original images.

Model creation

The Conus model was created as described in Identifying Shells using Convolutional Neural Networks: Data Collection and Model Selection. The hyperparameters used are provided in table III.

Some limited parameter tuning was performed, however the initial hyperparameters gave already good results (data not shown). The learning rate was decreased from initial 0.0005 to 0.0002, and the top layer dropout increased from 0.2 to 0.25 (see also Identifying Shells using Convolutional Neural Networks: Data Collection and Model Selection). This limited hyperparameter tuning was done iteratively. The final training was run for 73 epochs using early stopping. Inference was performed on the validation set and analyzed using sklearn.metrics, classification_report. The summary statistics are provided in table IV.

Because a low number of validation images were used to calculate the metrics for a significant proportion of the species (180 species have less than 20 validation images), the confidence intervals were calculated for the F1 Score using bootstrapping. The metrics and confidence intervals are given in table V. Only 6 species have low metrics (and large confidence intervals): Conus adenensis 0.540 (CI: 0.193-0.645), Conus auricomus 0.652 (CI: 0.476-0.792), Conus compressus 0.696 (CI: 0.428-0.880), Conus conspersus 0.640 (CI: 0.363-0.827), Conus gilvus 0.666 (CI: 0.399-0.864), Conus turritinus 0.588 (CI: 0.200-0.823), and Conus vezzaroi 0.741 (CI: 0.500-0.903). The figure below show the distribution of the F1 Score for the validation set. The figure below shows a scatterplot where the number of images in the validation set is plotted against th F1 Score.

When comparing the F1 score based on the validation set with the F1 Score where training set images were added (to a max. of 200 images), we see for the majority of the species no large difference (see figure below). Only for a few species, those that have a low "Validation" F1 Score, we see a large difference with the F1 Score that includes training images (the species Conus compressus, Conus gloriamaris, Conus turritinus and Conus vezzaroi). This is expected.

Confusion Matrix

The confusion matrix shows which species are confused most often. Following species are confused most often:

• Conus ardisiaceus and Conus aemulus
• Conus asiaticus and Conus alabaster
• Conus angasi and Conus adenensis

Visual inspection shows there are similarities between both species: Significant misclassification was confined to this species. In contrast, misclassified individuals of other species were not concentrated in any particular group.

Additional tests

Additional images were collected after model creation. This anecdotal test for several species confirm the performance of the model (Table VI).

Table VI presents the recall and associated confidence intervals calculated from the validation sets for each tested species, alongside the classification outcomes for newly tested images. Two images of Conus striatus were misclassified, although with relatively low prediction probabilities (0.58 and 0.55). More notably, four images of Conus merletti were consistently misclassified. These images depict a single specimen photographed from different angles and under slightly varying contrast conditions. All were incorrectly identified as Conus moluccensis, a species visually similar to Conus merletti. Potential explanations for this misclassification include the possibility that the specimen is genuinely Conus moluccensis rather than Conus merletti, or that there are inaccuracies or overlaps within the training dataset—for example, Conus moluccensis specimens mistakenly included within the Conus merletti class. Another interpretation is that these two species might actually represent varieties of the same species.

Discussion

Developing a CNN-based classification model for the Conus genus involves a variety of unique challenges stemming from both the taxonomic complexity of this group and the subtle characteristics that distinguish its species. One of the most significant hurdles is that Conus species are primarily differentiated by their color patterns, which can be very similar across species. Even minor variations in lighting, shell wear, and image quality can obscure these differences, making it difficult for a CNN to accurately distinguish one species from another.

Compounding this challenge is the sheer size of the Conus genus—currently recognized to include 853 species—making it one of the largest genera within the Mollusca phylum (WoRMS, MolluscaBase). Handling such a large number of classes naturally demands more computational resources, including significant memory capacity. As a result, training a Conus CNN model can be both computationally expensive and time-consuming, often taking over four hours on our infrastructure.

Another important factor is that Conus shells are highly sought after by collectors, leading to a relatively large pool of publicly available images. On the one hand, this abundance of data provides a rich resource for model training. On the other hand, it requires careful data management to account for variations in image quality, resolution, and lighting conditions, as well as potential imbalances in how frequently each species is photographed.

Altogether, these considerations—high species diversity, subtle morphological and color distinctions, and a sizable but variably curated dataset—highlight the complexity of creating a robust Conus CNN model. Building such a model requires thorough data curation, meticulous preprocessing, and a well-designed computational infrastructure capable of supporting prolonged training periods. Our CNN model achieved an accuracy of 96%, utilizing 130 373 cone snail shell images.

Before model training, extensive pre-processing is performed. All images were analyzed to detect the number of shells in the image and a separate image was made for each shell. If possible, the background was changed to black. A fixed input size of 400x400 pixels was used. Images were made square if needed. A final , manual step was included to select images that clearly show shell features that help in species identification.

Metrics calculated for each species (recall, precision and F1 score) shows that a large proportion of all species can be identified reliable. Calculation of the confidence intervals support this conclusion.

Two other models of the Conus genus were created before, both with good performance [5, 6]. The team at Naturalis, Leiden created several models for several topics, including also a Conus model. The model was trained on 797 Conus species, 15 877 images. Performance of the model is not communicated, but a limited tests (data not shown) show a good performance. N. Qasmi et al. have also made a Conus AI model, based on 47 600 images on 119 Conus species. Their model has 95% accuracy using a combination of Random Forest (RF), XGBoost (XGB) methods and feature extraction using a CNN.

Two other Conus genus recognition models have been previously published, each demonstrating promising but somewhat different approaches and datasets [5, 6]. First, a team at Naturalis in Leiden created several AI-based image-recognition pilots, among which was a Conus model trained on 15,877 images spanning 797 species [5]. Although overall performance results were not formally published, a small pilot evaluations (data not shown) suggested the model delivered robust predictions for many Conus species. This suggests that even with a moderately sized dataset (15,877 images), accurate species-level classification can still be achieved if the training images are curated carefully and taxonomic labels are standardized (e.g., via WoRMS).

Second, a recently reported model by Qasmi et al. [6] employed a combined approach of deep learning (VGG16 for feature extraction) and ensemble supervised learning (Random Forest and XGBoost). Their dataset encompassed 47,600 images of 119 Conus species, achieving around 95% accuracy. Notably, their workflow involved explicit feature-engineering steps—such as color moments, local binary patterns, and Haralick textures—before applying ensemble classifiers. This pipeline effectively demonstrated how hybrid methods (deep feature extraction plus machine-learning classifiers) can yield strong performance in a challenging fine-grained domain.

Compared to these two models, the CNN described in this study substantially broadens the species coverage to 518 species, incorporating 130,373 images—a dataset volume almost three times as large as Qasmi’s and well above the Naturalis pilot. Nonetheless, it attains a similarly strong performance: an accuracy of 96% with a macro-average F1 score around 0.96. This outcome underscores two important points:

1. Broader Taxonomic Scope vs. High Accuracy: Expanding classification from 119 to 518 Conus species introduces further inter-class similarity, increasing the risk of misclassification. Despite this, the final accuracy remains comparable to previous efforts, implying that large-scale coverage does not necessarily diminish model precision—provided the dataset is well curated and sufficient computational resources are available.
2. End-to-End CNN Training vs. Hybrid Feature Extraction: Unlike Qasmi et al.’s approach, which used a CNN (VGG16) mainly for feature extraction before applying classical ensemble methods, this model employs a fine-tuned convolutional neural network pipeline. Both methods illustrate valid strategies for biodiversity image classification. Ensemble approaches may be easier to interpret or to integrate with domain-specific features, whereas fine-tuning a CNN can leverage the model’s internal feature hierarchy, especially when the training set is extensive.

Another distinguishing factor is the volume and diversity of images. In the Naturalis pilot, images came predominantly from a handful of museum collections plus some private collections (15,877 total) ​[5]. Here, over 130,000 images were aggregated from a wide array of sources, including community-driven repositories, potentially bringing greater variance in lighting conditions, viewpoints, and shell morphologies. While this diversity strengthens generalizability, it also escalates demands on data preprocessing, standardization, and computational power. In particular, the workflow included automated image segmentation (one shell per image) and uniform background replacement—steps that appear to significantly streamline model training. Both the Naturalis pilot and Qasmi et al. ​[6] used similarly rigorous approaches for data cleaning, but with smaller datasets and fewer species, the effect of image variation may have been comparatively lower.

At last, a notable distinction in the current study is the use of transfer learning and fine-tuning, particularly leveraging an EfficientNet architecture pretrained on ImageNet. Although ImageNet does not contain seashell images, the extensive and diverse features learned from over a million labeled images still confer a significant advantage when training on the Conus dataset—or any other seashell dataset. By pretraining on ImageNet and then fine-tuning on domain-specific images, the model inherits rich, general-purpose visual representations that aid in discerning even subtle morphological details of shells. This approach is particularly effective for fine-grained biodiversity classification, where expert-labeled data are often scarce and species distinctions can be minute. ​[7].

Although direct performance comparisons can be confounded by differences in taxonomy, image sources, or evaluation protocols, these concurrent findings strongly support the viability of AI-based classification for large, visually diverse mollusk genera. Future work may involve combining the strengths of these approaches: unifying data from multiple sources, benchmarking different architectures or ensemble methods, and assessing the impact of refined taxonomic standards on model reliability.

One of the major challenges in building a robust Conus classification model is the imbalance in species representation, where some species have thousands of images while others have only a few. This imbalance is a common issue in biological datasets, where rare or newly discovered species often have limited available data ​[1, 2]. Few-shot learning techniques based on meta-learning and contrastive learning have been successfully applied in biodiversity classification to address data scarcity [8]. In recent studies, prototypical networks and metric-based learning have enabled models to recognize species with only a few labeled images by learning generalized feature spaces that capture taxonomic similarities [9, 10]. Similarly, contrastive learning, which pretrains models using large unlabeled datasets, has demonstrated superior transferability for species recognition tasks [11, 12]. Future work could explore such approaches to enhance classification performance for Conus species with very few training images, reducing the impact of dataset imbalance. By incorporating these advanced transfer-learning methods, AI models could better support biodiversity research, particularly for rare and underrepresented species.

This Conus AI model is a node of the hierarchical CNN model available at Identifyshell.org.

Conclusion

In summary, this work demonstrates the feasibility and accuracy of a large-scale CNN-based classification model for Conus shells — one of the most diverse and taxonomically challenging groups within the Mollusca. By assembling a dataset of over 130,000 images representing 518 species, we highlight the key hurdles inherent to fine-grained shell identification, including taxonomic ambiguity, limited or imbalanced species-specific data, and subtle morphological differences. Careful data curation, background standardization, and strategic model fine-tuning were crucial in achieving consistent performance across hundreds of species, as evidenced by high macro-averaged metrics and reliable confidence intervals.
The findings underscore that high coverage of Conus species need not compromise classification accuracy, provided the dataset is sufficiently robust and preprocessing steps are meticulously executed. Comparing our results to earlier Conus AI models further illustrates how diverse computational strategies—ranging from end-to-end CNN training to hybrid feature extraction—can yield strong results in challenging biodiversity contexts. These approaches collectively validate the viability of automated shell recognition on a scale that can significantly accelerate research and improve collection management for museums, citizen science platforms, and other stakeholders interested in marine biodiversity.