Hierarchical training improves consistency of pollen classification on the family level

Official URL: https://dx.doi.org/10.1109/ICIPW68931.2025.11386195

Accurate classification of pollen grains is crucial in various fields such as biology, medicine, and environmental monitoring. Even though deep learning methods have improved performance over traditional hand-crafted approaches, existing models often do not take into account the taxonomic hierarchy information of pollen species. In this paper, we propose a hierarchical training framework that jointly optimizes at the family, genus, and species levels using a shared feature extractor and parallel classifier heads. We evaluate our methods using a newly created dataset of 1,787 microscopic pollen images from 274 species, belonging to 145 genera, or 47 families. Experimental results, using two different backbones, demonstrate that the hierarchical training setup improves consistency on the family level while maintaining comparable classification performance at the species level. These findings highlight the value of incorporating biological taxonomy into model training for fine-grained image classification tasks.