Deep learning accurately identifies fjord benthic foraminifera

Over the past several decades, there has been increasing interest in using foraminifera as environmental indicators for coastal marine environments. Foraminifera provide equally good environmental quality status assessment as compared to large invertebrates (macrofauna), which are currently used as biological quality elements. However, foraminifera offer several distinct advantages as bioindicators, including short response and generation times, a high number of individuals per small sample volume, and hard and fossilizing shells with a potential of paleoecological record. One of the major challenges in foraminifera identification is the reliance on manual morphological methods, which are not only time-consuming and error-prone but also highly dependent on the expertise of taxonomic specialists. Deep learning, a subfield of machine learning (ML), has emerged as a promising solution to this challenge, since a neural network can learn to recognize subtle differences in shell morphology that may be difficult for the human eye to distinguish. In addition, the speed and ease afforded by deep learning methods would allow experts and non-experts alike to use foraminifera more extensively in their work, thus helping to integrate the use of foraminifera in biomonitoring programs by agencies and industry. This study focuses on benthic foraminifera from several Skagerrak fjords, including Gullmar Fjord, Hakefjord, Sannäs Fjord, and Idefjord (Fig. 1a). Sediment archives from these fjords provide extensive records of past and ongoing climate and environmental changes. Fjord foraminifera mounted on microslides were imaged using a stereomicroscope (3003 images), and individual foraminifera were labeled using the Roboflow online platform (22 138 individuals). Using the labeled images, we trained a You Only Look Once (YOLO) v7 deep learning model, which demonstrates state-of-the-art speed and performance for object detection as of the time of writing. The models can distinguish among 29 species with 90.3 % and 78.8 % mean average precision in the best- and the worst-performing models, respectively. Even though the imaging and labeling was done in a short amount of time (∼ 300 h over a course of 2 months), the results show that even a relatively small dataset can be used for training a reliable deep learning species identification model.