Retrieving fine-grained diatom classifications from deep learning models trained on coarse-grained data by over-clustering

Benthic diatoms, unicellular microalgae, are important bioindicators of freshwater ecosystem health. However, identifying diatoms using light microscopy is very challenging due to the existence of thousands of species, limited human resources, and the close morphological similarity between species. Automating this process using machine learning could enhance biodiversity monitoring; however, the scarcity of training data across species limits its effectiveness. For biodiversity monitoring and ecological status assessment, species-level identifications are preferred. Genus-level identifications are a first approach until training data covering a sufficient number of species is available. Moreover, models trained on genus-level labels might still learn to separate species in their feature space. We refer to the application of a clustering method to representations learned by models trained on a coarse (genus-level) classification task, with the aim of resolving clusters at a finer granularity (species), over clustering. This study evaluates the over-clustering capabilities of different deep learning architectures: a convolutional neural network (CNN), a vision transformer, and MAPLE, a CNN-based model incorporating X-means clustering during training. For CNN and vision transformer, features were clustered using HDBSCAN or X-means; for MAPLE, we used its internal clustering. Comparing the performance of the models by visual ordinations and clustering metrics, we found that although none of the models come close to separating all species within a genus, according to most metrics, MAPLE can most effectively separate species in the latent space. We suggest using over-clustering on pre-clustered novel diatom images to speed up expert annotation at the species level. This will contribute to an efficient collection of training data with broad species coverage.