COMPUTER VISION CRACKS THE LEAF CODE

Leaves are the most conspicuous and abundant major plant organs, and compressed leaves are the most abundant type of plant macrofossil. However, assessing the natural variation and systematic value of leaf size and shape (architecture) remains one of the most difficult challenges in botany and paleobotany. Living leaves show tremendous variation both among and within the hundreds of thousands of plant species, and a single leaf can have tens to hundreds of thousands of vein intersections and free endings. Here, we compiled a very large database of annotated, cleared, angiosperm leaves with over 7500 images from multiple sources. We used a biologically inspired computer-vision system, based on experimental knowledge of object perception in the mammalian visual cortex, to automatically classify images into Order, Family, and other natural categories. After training, the system generates a large dictionary of venation patterns, which corresponds to the most informative leaf-image regions and has great potential for novel systematic studies. Despite the fact that nearly all samples had typical imperfections such as rips, insect damage, and low image quality, the classifications were correct many times more often than chance level, and much more frequently than alternative approaches based on leaf shape alone (without using venation). Overall, the success of the approach suggests that leaves have tremendous phylogenetic signal, probably comparable to that of flowers, that is now much more accessible for study.