Computers are gaining on our botany skills.
Scientists have developed software that combines machine learning and computer vision to guess which plant family a leaf belongs to.
Although it's designed for botanists, it makes a phone app for plant identification — perhaps something similar to Shazam, which can identify music — not seem like such a stretch.
The software was a joint venture by Peter Wilf of Penn State and Thomas Serre, a neuroscientist at Brown University.
Back in 2007, Serre taught computers to differentiate between photos with and without animals. He managed to reach an 82% accuracy rate, besting his human students' 80% accuracy.
Wilf read about Serre's work and realized a similar algorithm might be able to rapidly classify leaves. That got Wilf excited, since identifying plants — especially the ancient variety (leaves are the most common fossils) — remains a big challenge.
Botanists still use a 19th-century method to identify plants known as "leaf architecture," which follows an exhaustive codebook of standards for examining the extraordinary variety of leaf veins and structure. It's thorough but not quick: A person may spend up to two hours to determine a leaf's place in the tree of life.
Software, meanwhile, stands to navigate that complexity in milliseconds.
"I've looked at tens of thousands of living and fossil leaves," Wilf told Wired. "No one can remember what they all look like. It's impossible — there's tens of thousands of vein intersections."
Wilf, Serre, and their colleagues' new program works off of a growing database of 7,597 images of different leaves from 2,001 genera, bleached and stained to highlight vein structures, according to a study published in Proceedings of the National Academy of Sciences. The software is "trained" using the images and
Each image in the database has a "heat map" of factors the software uses to help categorize a leaf:
Wilf and Serre's algorithm isn't perfect, but it can already determine related plant families of a leaf with 72% accuracy. And that includes specimens with holes chewed in them, fungal infections, and other imperfections.
The program is currently unable to identify an exact species — say, one of 600 types of oak— but it can act as a valuable assistant to a botanist, who can pick up the analysis from there.
It's important to note this isn't the first-ever plant-identifying software. Pl@ntNet, for example, can identify plant species from images and was first released in 2013. But unlike although the new software, its database is limited to plants from Western Europe, South America, and the areas around the Indian Ocean.
There are also plant-identifying apps, like Leafsnap for iOS, but they're primarily digital guidebooks with mediocre image recognition.
Whatever the case, it seems like these algorithms are only improving with time — so Tech Insider looks forward to the day when we can whip out our phones like a Tricorder and pretend to be botanists exploring an alien world.
Join the conversation about this story »
NOW WATCH: Someone has designed a pot of flowers that can take and post selfies on its own