Robust Decision‐Making via Uncertainty Quantification in Deep Learning Models for Marine Microfossil Classification

Abstract

Taxonomic identification of marine microfossils such as foraminifera and coccolithophores is essential for reconstructing past oceanographic and climatic conditions. However, this process remains a major bottleneck in paleoclimate research due to its dependence on expert knowledge and time‐consuming manual analysis. Recent advances in deep learning, especially the convolutional neural networks (CNNs), have spurred efforts to automate such process. Yet, standard CNN models tend to produce overconfident predictions for out‐of‐distribution specimens, as they do not explicitly account for epistemic uncertainty. Such overconfidence can lead to misclassifications, undermining the reliability of downstream paleoclimate reconstructions. In this study, we implement and evaluate three uncertainty‐aware deep learning methods, CNN with Monte Carlo Dropout (MCD), Bayesian CNN, and Deep Kernel Learning (DKL) to enhance robustness of microfossil classification. Using a curated data set of modern planktonic foraminifera, we assess the models' ability to produce uncertainty estimates over predictions. Our results show that these methods can flag out‐of‐distribution samples with large inference uncertainty, thereby significantly reducing both false positive and false negative errors in species identification. We suggest that uncertainty‐aware models, particularly when combined in ensembles, can provide a reliable framework for automated microfossil analysis to enable scalable, high‐throughput analysis in paleoceanographic and paleoclimatic research.