Abstract
Deep-learning-based classification of pollen grains has been a major driver towards automatic monitoring of airborne pollen. Yet, despite an abundance of available datasets, little effort has been spent to investigate which aspects pose the biggest challenges to the (often black-box- resembling) pollen classification approaches. To shed some light on this issue, we conducted a sample-level difficulty analysis based on the likelihood for one of the largest automatically-generated datasets of pollen grains on microscopy images and investigated the reason for which certain airborne samples and specific pollen taxa pose particular problems to deep learning algorithms. It is here concluded that the main challenges lie in A) the (partly) co-occurring of multiple pollen grains in a single image, B) the occlusion of specific markers through the 2D capturing of microscopy images, and C) for some taxa, a general lack of salient, unique features. Our code is publicly available under https://github.com/millinma/SDPollen.