Images were taken automatically by scanning the foraminifera or raw sediment using a modified 3D printer acting as a gantry system for X, Y, and Z hyperfocal image acquisition (“Sashimi system”; https://github.com/microfossil/particle-scanner, last access: 18 March 2026). A 5-megapixel camera (Basler ace acA2440-35uc), a telecentric lens (VS Technology VS-TCH4-65) with 4× magnification, and a ring light (VS Technology VL-LR2550W) for constant illumination were attached to the 3D moving head of the printer. Using the software called Sashimi, the 3D printer and camera could be controlled in the X, Y, and Z directions and enabled automated imaging of foraminifera and particles (Fig. 1). Stacked images were acquired with constant parameters: exposure (1500 µs), stack height (1600 µm), and stack step (20 µm; see example of stacking and fused image in Appendix B). Fusing the individual images into a fully focused image was done using the Helicon Focus Pro software (Helicon Software: https://www.heliconsoft.com/heliconsoft-products/helicon-focus/, last access: 9 May 2023). Automatic white balance was activated using the Basler Pylon Camera Software Suite (https://www.baslerweb.com/de-de/software/pylon, last access: 18 March 2026). Only the stacked images were used in this study. A complete scan of a micropaleontological slide takes approximately 1.5 h. Sashimi and the instructions for building the 3D printer can be found on GitHub (https://github.com/microfossil/particle-scanner, last access: 18 March 2026).

The image acquisition process yielded images with multiple foraminifera per field of view. The full-size images were cropped to obtain individual image objects using a CNN-based object detection workflow (Godbillot et al., 2024). This process requires the constitution of an annotated dataset to train the model. We annotated images from both MUDSURV and MEDIT using the open-source Computer Vision Annotation Tool (CVAT.ai Corporation, 2023). For the annotation step, we drew a bounding box for all occurrences of living foraminifera on the images, which were pooled into a single “generalForam” category. We used this image dataset to train a faster region-based convolutional neural network model (He et al., 2020), using a ResNet50 pre-trained on COCO as the backbone with a threshold of 0.5 (Godbillot et al., 2024). Bounding boxes were drawn for 20 min, which were then used to train a starting model. This model was applied, and we corrected the inference using this as a new batch for the end model. This step was repeated, and, in the end, 10 micropaleontological slides of each region (MUDSURV: 5846 boxes, 1134 pictures; MEDIT: 1406 boxes, 682 pictures; total of 3605 foraminifera; Appendix A) were used. Object detection was done on a shared server hosted by the OSU Institut Pythéas, geared with a Quadro RTX 4000 GPU card. To train this “generalForam” object detection model, 10 micropaleontological slides of each region were used.

A test step was added to evaluate the performance of the object detection model. We selected one sample from both regions with 100 pictures containing foraminifera. None of the images had previously been used during training, ensuring an independent analysis of the model. We applied the object detection model to these 200 pictures using a detection threshold of 0.5 to detect the highest possible number of foraminifera, i.e., increase the recall. This low threshold was chosen to guarantee that a maximum of instances of foraminifera was detected on the images while ensuring that multiple bounding boxes were not generated for single instances of foraminifera (Godbillot et al., 2024). For all images, we manually counted the true positive, false positive, and false negative instances, which we then used to derive values for accuracy, precision, recall, and F1 score.

Three classification models were trained, two on the Atlantic MUDSURV low-diversity time series and one on the highly diversified Mediterranean Sea. The MUDSURV intertidal Atlantic samples were used to build the first model, called “MUDSURV species model”, distinguishing the species of foraminifera, regardless of whether the specimens are dead or alive (Rose Bengal-stained). The second model, named “MUDSURV living dead model”, includes two classes for each species, one for living and one for dead specimens, differentiating the staining as well. The third model is the “MEDIT species model”, used for differentiating species from the Mediterranean coast.

Individual images (13 679 images) of foraminifera, obtained from the object detection workflow, were manually labeled using the software ParticleTrieur (version 3.0.5e; GitHub: https://github.com/microfossil/particle-trieur, last access: 18 March 2026) to train a CNN for the automatic recognition of different species (Marchant et al., 2020). Duplicates and images with multiple objects or with blurred, partial, or unrecognizable foraminifera were manually discarded. When possible, manual foraminifera determination based on pictures was performed to species level according to the literature (Barras et al., 2014; Hayward et al., 2020; Jones and Brady, 1994; Jorissen et al., 2023; Milker and Schmiedl, 2012) and reference slides from the Laboratory of Planetology and Geosciences (LPG) at the University of Angers (France). After completion of the dataset's labeling process, the particle/object classification models were trained with the MISO library, containing the scripts to train a CNN and the TensorFlow framework to be able to run machine learning tasks (Marchant et al., 2020; TensorFlow Developers, 2021; GitHub: https://github.com/microfossil/particle-classification, last access: 18 March 2026). ParticleTrieur offers several CNNs to choose from (Marchant et al., 2020) and was run on an HP ENVY x360 Convertible 15 laptop equipped with an AMD Ryzen 7 4700 CPU, 16 GB RAM, and an integrated AMD Radeon Vega 7 GPU. The training of the models was carried out with the foraminifera classes containing more than 10 specimens. A total of 80 % of the dataset was used as “training” to train the model, while the remaining 20 % was grouped into a “validation” dataset and used to evaluate the model performance at each iteration.

Several CNNs proposed by ParticleTrieur were tested, and the best model output for the MUDSURV dataset was obtained using a residual network 50 architecture with transfer learning (ResNet50 TL), while for the MEDIT images, the best results were achieved with a residual network 101 architecture using transfer learning (ResNet101 TL). In both cases, pre-trained weights from the ImageNet database were used, speeding up the training step (He et al., 2015). The ResNet50 TL is a CNN that consists of a total of 50 layers with a special feature that allows it to skip layers if there is a previous strong positive weighting (He et al., 2015). This results in a reduction in gradients, which facilitates classification and makes it possible to form deeper networks with increased accuracy (He et al., 2015). With transfer learning (TL), knowledge already acquired from previous tasks is reused to perform new tasks and facilitate training (Torrey and Shavlik, 2010). Input image sizes are fixed at 224 pixels (224×224×3) and included color images. Class balancing by weighting was enabled. The epoch (training iterations) size was 10, and the batch size (number of pictures that go through the network in one run) was 32.

Mislabeled images were estimated using a K-NN clustering technique implemented in ParticleTrieur during the first training. Using the estimated mislabeled images, the dataset was revised after a manual check. This process was done once for both MUDSURV and MEDIT, and only labels tagged as misclassified were edited.

After the segmentation, the resulting dataset consisted of 6388 images of individual (living and dead) foraminifera showing four dominant calcareous species with similar spiral shell shapes (trochospiral for Ammonia confertitesta and planispiral for Elphidium oceanense, Elphidium selseyense, and Haynesina germanica) and two additional less abundant species: one agglutinated with an initial planispiral turn followed by a uniserial shell (Ammobaculites balkwilli) and one monothalamid species (Psammophaga sp.) (Appendix A). Finally, four very rare species, represented only by a few images (< 10), were not included in the training set: Ammotium salsum (n=5), Lagena sp. (n=3), Leptohalysis scottii (n=1), and Quinqueloculina sp. (n=4), yielding a final dataset of 6375 foraminifera images (Appendices A and C).

This image dataset was used for both models (species with or without a living/dead distinction); only the labels differed according to the purpose of the model: (i) the first model (MUDSURV species model) distinguished the different species (Fig. 2), and (ii) the second model (MUDSURV living dead model) distinguished the different species and included an indication of whether they were living or dead specimens according to the staining.

The MUDSURV species model, intended to distinguish the six different species (i.e., six classes), shows an accuracy of 96.0 %, a precision of 96.1 %, a recall of 95.5 %, and an F1 score of 95.8 % (Fig. 2; Appendices C and D). Five of the six species are recognized with an accuracy of over 96 %, while Ammobaculites balkwilli and Psammophaga reach 100 %. Three of the dominant species reach an accuracy around 96 %–98 %, namely Ammonia confertitesta, Elphidium oceanense, and Haynesina germanica. Elphidium selseyense (104 specimens) has the lowest accuracy (83 %), showing the highest misidentifications with the species Ammonia confertitesta (9 %) and Elphidium oceanense (7 %) (Fig. 2). For the MUDSURV species model, 61 foraminifera were mislabeled (0.96 %).

The MUDSURV living dead model, designed to distinguish the different species and their vital status defined by the RB stain, achieves an accuracy of 94.2 % for 10 different classes (see Fig. 2; Appendices C and E). This results in a precision of 91.5 %, a recall of 93.7 %, and an F1 score of 92.2 (Fig. 2). Herein, 8 of the 10 classes are recognized with an accuracy of over 90 %. The Ammonia confertistesta dead specimen (101 specimens) class shows an accuracy of 80 %, followed by the Elphidium dead specimen class with an accuracy of 86 % (66 specimens), the only two classes being under 90 %. All Psammophaga sp. specimens (n=2) were misclassified as dead. Of the dead Ammonia confertitesta, 14 % were classified as alive with the CNN, and, in contrast, 6 % of living Elphidium were classified as dead. Between the Elphidium species, 7 % of Elphidium selseyense were wrongly classified as Elphidium oceanense and vice versa for an extra 3 %. For the MUDSURV living dead model, 138 foraminifera were mislabeled (2.16 %). The MUDSURV species model was trained in 540 s and the MUDSURV living dead model in 600 s.

To verify the applicability of the MUDSURV species model, six independent test samples from MUDSURV (monthly monitoring, October 2020 to April 2021, except February 2021) with a minimum of 145 (B-01/21) and a maximum of 1650 (B-10/20) specimens were used (Appendix A). The number of foraminifera per species counted manually was compared with the number automatically counted using the CNN model for each station (Fig. 3). Overall, the CNN correctly identifies 91.48 % of the specimens on average. This value ranges from 85.11 % for the January 2021 sample to 93.34 % for the October 2020 sample. The absolute densities of the four major species exhibit similar temporal trends between monthly samples for human and CNN analyses (Fig. 3). For Ammonia confertitesta and Haynesina germanica, the absolute densities are similar between both approaches, whereas they are slightly different for the Elphidium species, especially for E. selseyense, which also exhibited lower accuracy according to the model performance (Figs. 2, 3). When a difference is noticed, this is mostly the CNN that underestimates the absolute density of the species. This is mainly due to the classification of specimens in the “unsure” class when the threshold of 0.8 is not reached.

The dataset for the Mediterranean region comprises a total of 7291 images of individual foraminifera out of which 7148 were used for the CNN training (species with fewer than 10 pictures were discarded) of 63 species of foraminifera, which correspond to the 63 classes defined in the model (Appendices A, C, and F). The training of the MEDIT species model took 873 s.

The MEDIT species model reaches an accuracy of 82.2 % in species recognition. It achieves a precision of 77.3 %, a recall of 70.5 %, and an F1 score of 72.3 (Fig. 4). More than half of the species (40 classes) reach an accuracy of over 70 %. For 10 of them, the accuracy lies between 70 %–80 % and for an extra 15 species between 81 %–90 %, while 16 species are recognized with an accuracy > 90 %. Finally, a total of seven species are recognized with 100 % accuracy. Identification rates of 100 % were achieved for Ammodiscus planus, Cornuspira foliacea, Nonion scaphum, Nonionella turgida, Spiroloculina grateloupi, Trochamminula, and Valvulineria bradyana. In contrast, a total of 12 species present an accuracy below 50 %, with Biloculinella irregularis, Miliolinella subrotunda, and Quinqueloculina spp. completely unpredicted by the model (accuracy = 0 %). For the MEDIT species model, 918 foraminifera were mislabeled (12.84 %).

The MEDIT model above was applied to the six “independent test samples” (Appendices A and G) to compare its identification with an expert ground truth. The foraminiferal abundance in samples ranges from 36 to 327 specimens. The accuracy of correct predictions of the model spans 63.75 % (mid, September 2014) to 79.26 % (proximal, May 2014), with an average value of 69.69 %. Two species were present in the test samples but were not identified as a class in the model since they were absent from the training set (Fursenkoina acuta and Cassidulina oblonga). Figure 5 shows the comparison between human counts and CNN predictions for the four most common taxa in the test samples (Eggerella scabra, Reophax sp., Rosalina sp., and Textularia sp.). Overall, the CNN prediction follows similar trends to human counts across all samples, although discrepancies in magnitude occur.

Out of a total of 28 species declared as tolerant in Parent et al. (2021), 8 were present in the test samples. Of the six samples analyzed by the CNN, only one resulted in an EQS different from the expert assessment (Fig. 6). In sample ref. 05/14, the EQS changes from “very good” to “good” after the CNN classification (Fig. 6).

A general trend where higher training accuracy is associated with lower uncertainty (number of unsure specimens of a species divided by total specimens of this species) is observed in Fig. 7, although this relationship is not strictly linear. Several species with high accuracy (above ∼ 85 %) exhibit low unsure rates, indicating confident model predictions when sufficient training performance is achieved. Larger bubble sizes (representing more images) are mostly concentrated in this high-accuracy, low-uncertainty region, suggesting that greater data availability improves both accuracy and confidence. Seven species have unsure rates of 100 %. Two of them were not present in our dataset (Fursenkoina acuta (11) and Cassidulina oblonga (15)), having an unsure rate of 100 % and accuracy of 0 %. Cassidulina oblonga, being notably a tolerant species, was used to calculate for the TSI-Med. On the other hand, two species reach unsure rates of 100 % but have accuracies of over 80 % (Reussella spinulosa (18) and Asterigerinata mamilla (22)). Additionally, several species with moderate accuracies (40 %–70 %) display relatively high unsure rates (40 %–60 %), suggesting transitional performance.

To evaluate the potential of automated foraminiferal identification in ecological studies, we first focus on low-diversity intertidal environments, where species dynamics are well-documented and can be linked to environmental variability. In the case of the Bourgneuf intertidal mudflat assemblages, Choquel et al. (2026) used species absolute abundances in order to study foraminiferal dynamics through time according to environmental parameters. In particular, reproduction events (high-density events) were counted every time a species' abundance exceeded a certain limit. Benthic foraminifera reproduction follows a change of generation every few months in sexual and asexual phases (Murray, 2006). These reproduction events can be influenced by environmental factors, and because only the > 150 µm fraction is analyzed, a time lag occurs between the actual reproduction and observation in samples (Schönfeld et al., 2012).

The threshold value for a reproduction event was determined to be above 600 living foraminifera per 50 cm³ for Ammonia confertitesta, Elphidium oceanense, and Haynesina germanica and above 100 foraminifera per 50 cm³ for Elphidium selseyense (Choquel, 2026). In the period from October 2020 to April 2021, three high-density events could be observed (Fig. 3). Elphidium selseyense high-density events occurred in October 2020 and April 2021 (Fig. 3). In October 2020, a high-density event took place for Haynesina germanica (Fig. 3). Both of the high-density events which occurred in October 2020 could be identified using the CNN model with an accuracy of 67 % (Fig. 3).

In this case study based on low-diversity assemblages, we can therefore attest that the CNN model is able to reach sufficiently good results to study the seasonal dynamics of the different foraminiferal species. Fully automating the analysis of monitoring studies would enable increasing the sampling frequency to monthly or bi-monthly intervals, allowing a better understanding of foraminiferal responses to environmental changes in this kind of highly dynamic environment. The FOBIMO-group (Schönfeld et al., 2012) developed an international protocol for biomonitoring studies, recommending that sampling should be done once a year, preferably in autumn. To capture the variability of the environment, three replicates are necessary, yielding three samples for one station (Schönfeld et al., 2012). Most biomonitoring studies comprise several stations that are sampled seasonally or monthly, leading to an overwhelming number of samples (Alve et al., 2016; Bouchet et al., 2007; Burone et al., 2007; Dubois et al., 2021; Kenigsberg et al., 2020; Morvan et al., 2004). An automated approach saves time, enables the processing of larger sample sets, and increases the amount of data available on environmental dynamics while also improving taxonomic consistency by reducing operator-dependent bias.

We then assess the applicability of the automated identification model in more complex, high-diversity coastal environments, where ecological indices such as TSI-Med are used to evaluate ecosystem quality (Barras et al., 2014; Cavaliere et al., 2021; Dubois et al., 2021; Parent et al., 2021). Despite an average accuracy of ∼ 60 % for the correct identification of the Mediterranean test foraminifera, the TSI-Med index, on the test dataset, shows almost no differences in the evaluation of ecosystem quality between the manual and automatic methods (Fig. 6). Unlike other indices, such as diversity indices (Bouchet et al., 2012) or Foram-AMBI (Jorissen et al., 2018), the TSI-Med requires only the identification of a limited number of species listed as tolerant to environmental stress and the accounting of the total number of living foraminifera. Our findings indicate that taxa assigned as tolerant according to Parent et al. (2021) are among those yielding good accuracies as enough training pictures were available. Examples include Cancris auricula and Leptohalysis scottii. The human–machine comparison of the TSI-Med values shows reliable automated results, indicating that the combination of automated identification and TSI-Med calculation could be applied in studies of the ecosystem quality. To confirm this result, it would be interesting to test the model on samples including more species listed as tolerant since only 8 out of 28 were present in our samples. For Foram-AMBI (Alve et al., 2016; Bouchet et al., 2021; Jorissen et al., 2018), all species would need to be correctly identified with the CNN to obtain a correct value since all species are classified into five different ecological groups from sensitive to opportunist. Rare species would also need to be identified, which would be more difficult due to the limited number of images available to properly train the CNN.

The use of machine learning for the automated identification of benthic foraminifera has the potential to substantially transform current analytical workflows. Taxonomic identification remains one of the main issues in foraminiferal studies, as it requires time and specialized expertise. It is important to note that the reliability of CNN models is built upon hundreds to thousands of expertly labeled images, underscoring that such approaches depend on the essential role of taxonomists.

By enabling rapid and consistent species-level classification from large image datasets, the approach presented here could significantly reduce the processing time and facilitate the analysis of larger sample sets and higher-temporal-resolution datasets. When combined with automated imaging, such as the modified 3D printer setup used in this study, CNN models could contribute to the development of integrated, high-throughput systems for the analysis of benthic foraminiferal assemblages. Such systems would enhance the practicality of using foraminifera as bioindicators in environmental monitoring programs, where rapid and standardized assessments are increasingly needed.