Computer vision species identification of lichens and bryophytes from biocrusts in Australian drylands.

Premise: Due to their small size and lack of easily visible macroscopic characters, the identification of cryptogam species has always been challenging. Here, the use of a machine learning computer vision method is explored for the identification of species of lichens and bryophytes from Australian biocrusts. Methods: Three models were trained using mostly images from herbarium specimens. The models were then evaluated based on statistics produced by Microsoft Azure Custom Vision and a bench-test with the CSIRO Horama ID mobile app. Results: Despite the small size and reduced habit of lichens and bryophytes, the Cryptogam (lichens and bryophytes) model performance value is just slightly lower than the performance of a vascular plant model of similar scope (64% accuracy for the Cryptogam model versus 70.3% for vascular plants from Costa Rica). Discussion: The performance of our models suggested opportunities for improvement, including for bias issues caused by imbalanced datasets, white background, and mixed specimens, as well as the difficulty in stabilizing live images at high magnification when using a mobile device to deploy the model. Further opportunities to improve model performance for these small and character-poor organisms, including data augmentation and image segmentation, are also discussed.