Brazilian researchers win plant identification challenge

Scientists from the São Carlos Institute of Physics-USP win first place in an international ImageCLÉF competition using a new computer system that recognizes plant species from photos of leaves.

By Fábio de Castro

Agência FAPESP – Since 2003, the French organization ImageCLÉF has promoted an annual challenge to the international scientific community to stimulate advances in pattern recognition research.

A group of researchers from the Universidade de São Paulo (USP) won first place in the 2011 ImageCLÉF “Plant Identification” category, which was begun in 2010 with the support of the French Botanical Society.

The Brazilian scientists won the challenge of developing a computer system that can recognize European plant species from a databank of leaf photos.

Coordinated by professor Odemir Martinez Bruno from USP’s São Carlos Institute of Physics (IFSC), the study was part of the project “Computer Vision Methods Applied to the Identification and Analysis of Plants”, funded by FAPESP under its Regular Research Support program. Dalcimar Casanova and João Florindo were also on the research team: both are doctoral students being advised by Bruno, and Casanova has a FAPESP fellowship.

According to Bruno, the competition is important because it foments the solution of complex and important computing problems while also leading to advances that can make contributions to other fields—in this case, botany and the conservation of biodiversity.

“The task was to identify arboreal plants from Europe. We have been studying the application of computer vision methods to identify plants for about a decade. The method we used is based on the recognition of plants through visualization of the leaves and proved to be highly efficient. Now we need to perfect it so it can be used in a simple, standardized way,” Bruno told Agência FAPESP.

“We mainly focused on the detection of leaf shape, innervation and texture, taking pigment patters into consideration. With this, we established a databank that served as a recognition point in the image for species recognition. We achieved a success rate of nearly 50%, much higher than the other competing institutions,” Bruno affirmed.

The efficiency shown by the system that won the ImageCLÉF challenge was encouraging, but a much greater challenge lies ahead for the group: the identification of Brazilian forest species. “Our goal is to identify Brazilian biodiversity; this is much more challenging than European biodiversity, which is much simpler by comparison,” said Bruno.

Aside from seeking methods that allow fast and efficient forest study, the researchers want the technique to be able to help in the study of physiological and evolutionary phenomena that can predict the best climate, the best soil and the best environment for the growth of a species.

“We used many approaches for recognition, which can be performed via photos, scanned images, or reflection and transmission microscope images,” explained Bruno.

Simplifying the work of botanists

In 2011, the research group published many articles focused on the computer identification of Brazilian plants. One prominent article was published in the journal Plant Systematics and Evolution in partnership with Professor Rosana Kolb from Unesp’s Assis campus and botanist Davi Rossato.

“We studied the Melastomataceae family, which grows in the Brazilian cerrado (savanna) and is difficult to recognize. We used a conventional scanner to digitize images of the leaves, and we achieved 100% accuracy in identification of these species,” said Bruno.

According to Bruno, uniting physics, mathematics and computer sciences to analyze biodiversity through images means that researchers can increase knowledge and generate new products and methods. In 2011, the team, together with other USP groups, patented a technology that uses the computer vision method for evaluating the images of leaves, making it possible to mathematically detect a lack of nutrients in corn plants.

“That was a very successful methodology that emerged during the project without being the project’s main focus. But our final goal is to one day arrive at a plant recognition technology that allows biologists to easily survey the species in a forest while armed with just one portable device,” explained the IFSC-USP professor.

The surveying of plant species in forests is traditionally based on flowers and fruits, whereas leaves are not generally used because of their immense variability.

“What makes this approach important to us is the great variability of the leaves because it means we have to create new methods. From the exact-science point of view, plant identification is a type of challenge that leads to huge advances in physics, mathematics and computation. The more complex the problem, the greater the opportunity for scientific advancement,” affirmed Bruno.

Once the new methodologies have been perfected, he says that botanists won’t need to wait for plants to flower or bear fruit before identifying them.

“Each species has a different life cycle, blossoming and bearing fruit in distinct periods. This makes surveying the forest in a region very difficult. Our goal is to overcome the difficulties in the computational field in order to facilitate botanical work,” he said.