Speaker
Description
With an estimated 5 million species, fungi represent a remarkably diverse kingdom of life. They play important roles in ecosystems, through nutrient acquisition for symbiotic plant partners, providing habitat for insects by breaking down deadwood and debris in forests. While fungal species are increasingly threatened by human activities such as land use change, nitrogen leaching and pollution, policies protecting biodiversity focus mainly on animals and plant species. This gap in policy is likely due to a lack of awareness about fungal species in the general public and limited data on abundance and distribution of most fungal species. Automated identification applications can help to address both problems by raising awareness through enabling rapid identification from images and large-scale opportunistic data acquisition by citizen scientists. However, these applications need to be reliable in their predictions for them to be useful for identification and conservation purposes. Computer vision models such as Convolutional Neural Networks (CNNs) drive the predictions in identification apps, but it is not yet well understood what kinds of images are most useful for training these models to improve their accuracy. In the field, trained mycologists look at different macroscopic traits of fungal fruiting bodies to identify species, and these could be similar or different from the features that CNNs extract to make predictions. To test what kinds of image perspectives are useful for training CNNs, we designed a study for which we are collecting a systematic dataset looking at up to 9 different pre-defined perspectives for 4 morphological groups of mushroom-forming fungal species (bracket/shelf, gilled bell-shaped, gilled convex/plane/funnel, and boletes). Preliminary results from the early 2026 field season will be presented, with implications for best practice image collection and training protocols for the models powering fungal identification applications.
| Status Group | Doctoral Researcher |
|---|