Towards the automatized identification of moss species from their spore morphology.

Background and aims: Automatized species identification tools have massively facilitated plant identification. In mosses, spore ultrastructure appears to be a promising taxonomic character, but has been largely under-exploited. Here, we test artificial intelligence-based approaches to identify species from their spore morphology. In particular, we determine whether the number of spores, their polarity, and variation among populations and capsules affect model accuracy.

Methods: Scanning electron microscopy spore images were generated for five capsules of five populations in ten species. Convolutional neural networks with a highly modularized architecture (ResNeXt) were trained to identify the species, population and capsule of origin of a spore. The training set was progressively sub-sampled to test the impact of sample size on model accuracy. To assess whether variation in spore morphology among populations affected model accuracy, one population was successively removed to test a model trained on the four remaining populations.

Key results: Species were correctly identified at average rates of 92 %, regardless of polarity. Model accuracy decreased progressively with decreasing sample size, dropping to about 80 % with 15 % of the initial dataset. The population and capsule of origin of a spore was retrieved at rates >75 %, indicating the presence of diagnostic population and capsule markers on the sporoderm. Strong population structure in some species caused a substantial drop of model accuracy when model training and testing was performed on different populations.

Conclusions: Spore morphology appears to be an extremely promising tool for moss species identification and may usefully complement the suite of morphological characters used so far in moss taxonomy. The presence of spore diagnostic features at the population and capsule level raises substantial questions on the origin of this structure, which are discussed. Substantial infraspecific variation makes it necessary, however, to train an automatized identification tool from a range of populations and capsules.