AI for difficult herbarium specimens: identification of peat mosses (subgenus Sphagnum) without dissection

Artificial intelligence (AI) for image-based herbarium specimen identification has thus far focused on plants that can be identified by eye. Here, we develop the first AI focused on identifying herbarium specimens of a bryophyte group, peat mosses in Sphagnum subgenus Sphagnum. These plants have substantial morphological plasticity, and confident identifications require time-consuming dissections and microscopy. We hypothesized that AI, using unmagnified low-resolution images, can (H₁) identify species and (H₂) discover novel morphological characters. We collected 4386 publicly available herbarium specimen images of all 10 North American species and imaged an additional 105 specimens with determinations verified by DNA and morphology. AI identification was generally successful with our newly formulated FireNetSEz model (68% AUCPR (area under the curve: precision recall)). We produced a reduced dataset (the five most imaged species) that we, the authors, could attempt. Our identifications took hours and were all lower-scoring than the AI. These H₁ results show that AI can learn hard-to-identify botanical species without microscopy and outperform both generalist botanists and Sphagnum experts. Regarding H₂, we found the AI focuses on edges of organs that humans often ignore. AI holds promise for hard botanical identifications and the potential to rapidly identify Sphagnum, which is important for studying peatlands that strongly impact climate.