Abstract Detail

Biodiversity Informatics & Herbarium Digitization

Fochesatto, Stefano [1], Shavlik, Matthew [2], Ickert-Bond, Steffi [3], Hodel, Richard [4], Wen, Jun [5].

Using deep learning image clustering to aid species delimitation within the Vitis arizonica complex.

Currently the delineation of plant species and infraspecific taxa is largely achieved through the physical examination of herbarium specimens, from field observations, by anatomical comparisons, observed geographic separation, and by comparing DNA sequencing results. Taxonomic resolution within the Vitis arizonica complex has been a challenge due to its broad geographic range, evolutionary divergence and potential gene flow between taxa. Based on recent field work in the S and SW United States (Texas, New Mexico, Arizona, Utah, and California), additional morphological variation in this complex has been observed that might merit specific status. The digitization and online mobilization of herbarium specimens, in combination with advances in parallel computing, is rapidly advancing our ability to characterize morphological diversity using machine-learning approaches. The goal of this research is to use machine learning to capture and quantify the morphological variation inherent in the ~ 370 herbarium specimen images from several herbaria with large Vitis arizonica complex collections (ASU, RSA, and US). More broadly we hope to identify a workflow which leverages the vast amounts of digitized herbarium specimens, to aid in the identification and delimitation of closely related or interspecific plant taxa. We generated hundreds of high resolution masks for the Vitis arizonica complex using scripted computer vision tools and manual tools for image editing. We then used a convolutional variational autoencoder to compress these images into a latent space where a further cluster analysis might reveal biological insights that could aid in further species delimitation. Ultimately, these insights could help in an explicit statistical framework for morphological variation, aiding in delineating species by integrating morphological and phylogenetic data.

1 - University of Alaska Fairbanks, Mathematics and Statistics, P.O. Box 756660, Fairbanks, AK, 99775-6660, United States
2 - University of Alaska Fairbanks, Department of Biology & Wildlife, 101 Murie Building, Fairbanks, AK, 99775, United States
3 - University Of Alaska Fairbanks, Herbarium (ALA) And Dept. Of Biology And Wildlife, University Of Alaska Fairbanks, 1962 Yukon Dr., Fairbanks, AK, 99775, United States
4 - National Museum of Natural History, Botany, MRC 166, Smithsonian Institution, Washington, DC, USA
5 - Botany, MRC-166 National Museum Of Natural History, 10th St. & Constitution Ave., NW, Mrc 166, Washington/DC, 20013, United States

Keywords:
deep learning
Machine Learning
Vitis arizonica complex
species delimitation
herbarium specimens.

Presentation Type: Oral Paper
Number: BI&HD I002
Abstract ID:734
Candidate for Awards:None