Generation and validation of a high-resolution fossil diatom record from Saliña Bartol, Bonaire, based on convolutional neural networks

Diatom microfossil remains preserved in lake and marine sediments are widely used as proxies for reconstructing past environmental conditions. However, generating high-resolution diatom records is time-consuming and relies on the potentially subjective view of expert taxonomists. Recent developments in deep learning have led to an increase in proof-of-concept studies using object detection models for diatom classification and segmentation, but their practical use in creating reliable, continuous paleoenvironmental records has yet to be demonstrated. Here, we present the first application of an object detection model (YOLO11) trained on fossil diatom images to create a high-resolution diatom record for Saliña Bartol, a hypersaline lake in Bonaire. The model was trained on 3242 annotated objects representing 34 diatom taxa, and achieved a mean F1-score of 0.946. To evaluate performance in a real-world paleoenvironmental context, the model was first applied to create a continuous record of 22 virtual microscope slides that were also manually counted. Comparison of automated and manual counts showed highly consistent patterns in downcore relative species abundance, with differences mostly caused by larger or morphologically variable taxa. The model was applied to the full dataset of 399 virtual slides, producing a continuous diatom record spanning ∼ 905 years, with a resolution of ∼ 2 years per sample. The complete record reveals decadal-scale oscillations in aerophilous and brackish-water indicator species, suggesting variability in the hydrology of the catchment that in a traditional, low-resolution analysis would not have been detected. These results demonstrate that deep learning can be applied to automate fossil diatom quantification on a scale that manual analysis could not realistically achieve, marking the beginning of a new era in applying deep learning to create diatom-based paleoenvironmental reconstructions. • Developed a workflow for the automatic segmentation and classification of fossil diatoms using YOLO11. • Achieved strong model performance (mean F1-score of 0.946 across 34 classes). • Constructed the first high-resolution ( ∼ 2 yr/sample) diatom record using a CNN-based object detection model. • Identified challenges in full-slide inference. • Opened new avenues for CNN-based object detection models in paleoenvironmental reconstructions.