Create your own conference schedule! Click here for full instructions

Abstract Detail

Themes of land plant evolution, a celebration of the contributions of Leo J. Hickey

Wilf, Peter [1], Chikkerur, Sharat [2], Little, Stefan [3], Wing, Scott [4], Serre, Thomas [5].

Leaf Architecture: Computer Vision Cracks the Leaf Code.

Leo Hickey pioneered the modern study of leaf architecture, and in 1975, he and Jack Wolfe published what is still the only systematic survey of leaf architectural character states of angiosperms. In large part due to this work, phylogenetic signal in leaf architecture at a coarse scale is widely acknowledged. However, much remains to be done to refine understanding, especially in light of the more recent, large-scale re-alignment of angiosperm phylogeny from molecular data. Thus, the greatest intellectual challenge Leo left us may be to greatly improve evolutionary understanding of leaf architecture and thus to "unlock" the most abundant but underused data source in paleobotany, fossil angiosperm leaves. Leaves show tremendous variation both among and within the hundreds of thousands of extant plant species, and a single leaf can have tens to hundreds of thousands of vein intersections and free endings. Although Leo probably came closer than any other human to understanding this variation, machine vision offers the possibility to simulate the human processing pathway in amplified form, helping us to recognize much more and to record the information quantitatively. Using a large database of >7000 images of cleared angiosperm leaves (randomized halves to train and test), we trained a biologically inspired computer-vision system, based on a hierarchy of visual processing stages that closely mimics the first 0.1 seconds of perception in the primate visual cortex. We found that the computer algorithm repeatedly classified the test set elements into order, family, and other natural categories with accuracies greatly exceeding chance (>40% recognition for 29 APG III families and for 15 orders), despite the fact that nearly all samples had typical imperfections such as rips, insect damage, and low image quality. Classification accuracy also surpassed that of state-of-the-art machine-vision algorithms. After training, the learning algorithm produces visual word dictionaries corresponding to the most diagnostic leaf-image regions, and these can be displayed using simple heat maps for botanical interpretation. Preliminary observations show that, as for primates, there is strong detection of oriented lines. Accordingly, much "hidden" information lies in neglected variation of strong-line features, especially slight asymmetries in bases, apices, marginal curvature, secondary vein angles, and intercostal vein angles. As Leo always knew, leaves have tremendous phylogenetic signal, probably comparable to that of flowers.

Broader Impacts:

Log in to add this item to your schedule

1 - Penn State Univ., 537 Deike Bldg., UNIVERSITY PARK, PA, 16802, USA
2 - Google, 6425 Penn Ave. Suite 700, Pittsburgh, PA, 15206, USA
3 - University of Victoria, Biology, Cunningham Bldg. 202, Victoria, BC, V8P 5C2, Canada
4 - Smithsonian Institution, Dept. of Paleobiology NHB 121, PO Box 37012, WASHINGTON, DC, 20013-7012, USA
5 - Brown University, Department of Cognitive, Linguistic and Psychological Sciences, Brown Institute for Brain Science, Providence, RI, 02912, USA

leaf architecture
machine vision
leaf venation
primate visual cortex

Presentation Type: Symposium or Colloquium Presentation
Session: C6
Location: Belle-Chasse/Riverside Hilton
Date: Tuesday, July 30th, 2013
Time: 4:15 PM
Number: C6011
Abstract ID:578
Candidate for Awards:None

Copyright 2000-2012, Botanical Society of America. All rights reserved