As a biologist, imagine you’re installing camera traps in Florida waterways or collecting algae samples from algae bloom in the savanna. You might be able to draw boundaries for a new preserve if you pinpoint migration routes for endangered species. Algae blooms could implicate mass public health if we learn why they are blooming. Identifying thousands of species of algae in water samples or tens of thousands of nighttime photos of animals takes hundreds of hours. But now you can get help without needing a degree in biology.
AI advances have in the past five years made it possible for biologists to leave tedious tasks of species identification to machines trained to do it instead due to rapid developments in artificial intelligence (AI). Moreover, this explosion is also spreading into the public domain, allowing everyday citizens, travelers, landscape architects, farmers, and other types of individuals to identify what plant and animal is it.
Artificial intelligence algorithms developed for species identification are known collectively as Species-Identifying Artificial Intelligence (SIAI), which uses statistical methods to train machines to do what tools like the British app Shazam do for music: identify plants and animals. Deep learning, the process of analyzing multiple layers in an image, is particularly helpful for identifying wildlife.
It has been downloaded 5 million times on Google Play, a plant-identification app that launched in 2010 but switched to deep learning in 2017. Since its launch in 2014, the Merlin Bird ID app from Cornell Lab of Ornithology has identified 650 American bird species. A deep-learning app developed by iNaturalist in 2017 claims a 86 percent genus-identification accuracy rate. The German botanists, AI, and geographic data of Flora Incognita were used to identify more than 2,700 types of German plants earlier this year.
A Cornell professor who was part of the Merlin Bird ID team believes that deep learning technology is fundamentally superior to conventional computer vision tools, which analyze images.
Many people in the computer vision field were already doing object recognition before deep learning. According to Belongie, most of their work involved identifying everyday objects that normal people could recognize. Using the programs, you could identify a mushroom from a photo, but they couldn’t tell you if it was poisonous.
Artificial intelligence is truly a brute force endeavor with deep learning. In order to train it, you need a lot of data. There have been a number of open access databases that provide reams of images of thousands of labeled plants and animals in recent years. There are 1.2 million images available on Zooniverse, 675,000 photos on iNaturalist, and 1.8 million at iDigBio.
The power of deep learning is enhanced by access to those databases. Tools that perform deep learning studies the most basic aspects of images, such as the edge of a leaf. Patrick Mäder, a software engineer at the Max Planck Institute for Climate Impact Research, says they then examine more complex characteristics of the image – such as the texture of the leaf – before differentiating between the leaf and stem, before identifying a complete species. Backpropagation allows the program to self-correct errors. First, it determines if the misses were severe, and then it penalizes the worst errors. A change in its analytical process is then passed on to future analyses.
Deep learning capabilities were first demonstrated in 2012 when AlexNet, an AI platform, won the image-identification challenge ImageNet, making half the number of errors as the next best competitor. There was soon a race between Google, Baidu, and Facebook for deep learning experts. Funding and processing power continued to improve. A deep learning project was carried out by Google in 2012. This number increased to 1,000 by 2016. Open source deep learning architectures have been released by several tech companies. Further bolstering the image-identification AI pipeline was AlphaGo’s ability to defeat the human Go champion in 2016.
It was then that machine learning was being used to identify species for the first time. Scientists in England launched the Warblr app in 2015, which identifies birds by analyzing recordings of their songs. Sound-recognition tools can be useful for birds, but images are better for plants and millions of tiny species whose sounds can’t be recorded.
Identifying plants and animals with apps didn’t all start in their current form. As an example, Pl@ntNet, a great alternative to What is the Plant, was launched in 2010 by four French academic institutions. NIRCSC researcher Alexis Joly, a Pla@ntNet collaborator, says the purpose was to encourage citizen science. At that time, 200 contributors helped identify 70 European species. Using deep learning, they also identified plants from South America, North Africa, and the U.S. by 2017. Additionally, iNaturalist initially used community members to identify plants and animals before recognizing the advantages of deep learning.
In part inspired by his ornithologist sister, Belongie developed an identifying app at the Cornell Lab of Ornithology before using TensorFlow, a deep learning tool developed by Google.
They’re not perfect, but they’re worth checking out. For one, there are more species on Earth than there are species in the southern hemisphere, and there are very little data available on species for SIAI. The easiest way to identify species – like flowers – is to observe their seasonal characteristics. It is possible that photos used to create these models do not always have the clarity needed to distinguish between wild carrots and poison hemlocks. Plants that aren’t native to the United States, such as Oriental trees that are often used in American landscaping, may throw off these AI tools.
Although apps like Flora Incognita, iNaturalist, Merlin Bird ID and Pl@ntNet are the vanguard of a revolution, deep learning research in dozens of labs around the world is leading to animal species recognition by AI. Scientists and ordinary people alike are increasingly using smartphones to identify exotic species.