The value of machine learning has been demonstrated by scientists at the University of New South Wales and the Sydney Botanic Gardens, who trained and applied artificial intelligence (AI) algorithms to measure the leaves of approximately 3,800 plant specimens in minutes.
It would take a human researcher years to complete the same task.
“Leaf measurement is not a complex task, but it is labor intensive, making traditional methods impractical at scale. Machine learning approaches dramatically improve the process of extracting and editing these data. can be accelerated exponentially,” said the paper. American Botanical Journal.
Researchers used machine learning techniques to analyze digital plant specimens to focus on the relationship between leaf size and climate.
“Within each genus, leaf size was positively correlated with temperature and precipitation, consistent with previous observations. Within species, however, the association between leaf size and environmental variables was weaker. ,” says the paper.
Researchers used high-resolution images to train machine learning models. fig and Czygium Botanical specimen. These were extracted from his collection of over one million digital images created by the Sydney Botanic Gardens by scanning herbariums at the National Herbarium of New South Wales.
Associate Professor Will Cornwell, a researcher at the School of BEES and a member of the UNSW Data Science Hub, said the Royal Botanic Gardens “embarked on a very large-scale effort to scan every image in the collection.” I’m here.
“They got this huge machine from Holland, and they ran all herbarium sheets from the last 250 years on this scanning machine. […] All of Australia’s flora can be found there. ”
Cornwell worked with Dr. Jason Bragg of the Sydney Botanic Gardens and New South Wales researcher Brendan Wilde to create an algorithm that could automate leaf detection and size.
The researchers started by using the samples to train a machine learning model. Syzygium Commonly known as liripili, blush cherry and satina. fig, A genus of about 850 species of trees, shrubs and vines.
These species were chosen in part because the leaf structure is simple, which the researchers thought would be easier for machines to learn.
Researchers first trained a machine learning model using a relatively small set (35). Czygium image.
Machine learning (a convolutional neural network commonly known as computer vision) is trained on human examples to identify the leaves in each image. In this case, Wild drew contours around many digital images of leaves and entered them into the system.
After that, the model “learns what a leaf is and can find it very reliably in new images,” Cornwell says.
To make the measurements, the machine learning algorithm basically draws a contour around each leaf and counts the pixels within the contour.
Once the model was trained, the team applied it to measure leaves (leaf area, length and width) on digital images from 1227 Syzgium specimens and 2595 Ficus specimens.
Fifty images of each set were manually verified by Wilde to ensure accuracy.
According to Cornell University, even though machine learning has greatly accelerated the measurement of thousands of leaves, the accuracy and usefulness of the model’s output depends on the people who collected, labeled, and scanned the samples. , to Wilde’s careful work in training and validation, it relies heavily on human data. machine learning model.
“Human processes, humans using machine learning [are] It’s really, really important,” Cornwell says.
Now that the researchers have developed a leaf size model, they hope to work with more complex leaf shapes and analyze more specimens at the NSW National Herbarium.
They were also interested in whether the model would work using less controlled imagery, such as photos of plants taken by citizen scientists in a non-standardized way on different models of mobile phones. increase.
