AI at the service of the field: Brazilian project receives support from giant Nvidia
An important reinforcement will help researchers from Rio Grande do Sul develop new technologies that use artificial intelligence (AI) models for precision agriculture.
Having worked in the area since 2017, a group from the Federal University of Pampa (Unipampa) in partnership with Embrapa Pecuária Sul now has the support of Nvidia, a multinational company based in the United States and a global leader in AI components.
Selected for an academic research funding program, the Terrapampa Smart Systems: Bridging AI and Livestock project now has access to state-of-the-art computing resources, estimated at R$1.2 million. The project involves academics in applied computing, agronomists, veterinarians, administrators, and agricultural technicians.
"This international funding reinforces that we are at the forefront of knowledge in AI applied to agribusiness," says research coordinator Sandro Camargo, from the Unipampa master's program in applied computing.
AI model makes instant tick count on animalsOne of the models was trained to count ticks on cattle using images captured by a camera installed in animal handling areas.
Today, the process is usually done manually by a property employee, which, in addition to being time-consuming and costly, can result in human error.
After about six months of training an AI model, the researchers have already achieved a 94% accuracy rate. "This means that for every 100 ticks present on the animal, the algorithm detects between 94 and 106," explains Camargo.
The system already works on smartphones, but the idea is to port the application to a device that can be placed in areas where livestock roam. "The producer can distribute it in two or three locations, where the animals drink water, for example. The camera detects its presence and immediately counts ticks."
By counting and identifying the animal using an ear tag, a notification can be triggered to inform the presence and quantity of ticks on each individual. "The producer can set a minimum threshold, for example, of 10 ticks, to be notified."
On another front, researchers are developing models that aid in the genetic improvement of Hereford and Braford cattle. One possibility already tested is the selection of individuals less sensitive to sun exposure.
Periods of excessive solar radiation cause animals to stop feeding and seek shelter in shaded areas, which directly affects their weight. "For several years, genetic improvement has been underway to prioritize the mating of animals with so-called 'spectacles,' a red pigmentation around their eyes," explains Camargo.
The amount of pigmentation is directly relevant to an animal's resistance to ultraviolet radiation exposure, as it is linked to squamous cell carcinoma. A less sensitive animal spends more time on pasture, gaining more weight even during sunny periods.
Currently, eye evaluations are performed by specialized technicians who visit farms and assign scores on a scale of 1 to 5 to each individual, with the first being a complete absence of pigmentation and the last being an excellent level. Animals with scores of 4 and 5 are the most suitable for breeding, as they produce more adapted calves.
"The big problem is that experts typically visit a farm to analyze up to 500 animals in a day. Besides taking a long time, after a while the human eye begins to tire. There's a significant issue of inter-evaluator inconsistency," says the research group coordinator.
A model already ported to the Android system allows a technician or the producer to use a smartphone to capture an image of the animal and instantly obtain an assessment of the eye pigmentation score, eliminating the possibility of errors and optimizing the process time.
AI identifies keratoconjunctivitis and bovine anemiaAnother model developed by the researchers allows for early detection of bovine keratoconjunctivitis. Animals with inflamed corneas and conjunctiva generally show no behavioral changes in the early stages of the condition, meaning the condition is only identified at a more advanced stage.
Just like the tick counting system, an algorithm developed by researchers at Unipampa and Embrapa, running on a device installed in the field, can detect the passage of an animal in a given location and capture its image.
An instant analysis performed by a trained AI model is now capable of visually identifying signs of keratoconjunctivitis even when there are no other visible symptoms.
The same technique, with the help of another AI model developed by the group, allows the detection of anemia in animals without the need to directly photograph the ocular mucosa of each individual.
“The idea is that we can have, at any point on the farm, like in a chute [structure used to contain cattle], for example, a minicomputer, the size of a cell phone and powered by solar energy, which will cost the producer between R$700 and R$800,” says Camargo.
"When the animal passes by, the system will instantly detect whether it has ticks, if it's at risk of anemia, keratoconjunctivitis, and several other diseases." The system would eliminate the need for a technician to take photos of each animal. "That's what we'll reach soon. Everything will be automatic," he says.
The researcher explains that training the AI models involves capturing up to two thousand photos of animals at Embrapa stations and processing each image individually. This process, he says, took weeks to months because it required a lot of computational resources.
With access to Nvidia's cloud infrastructure, each hour it takes to train a model now takes about three minutes, allowing us to test new application possibilities, address more complex problems, and fine-tune AI agents.
Models are also trained to detect pests in agricultureThe same technology used in livestock farming can also be applied to detect invasive species in agriculture. Another project developed within the research project involves identifying annoni grass in native pastures, a common problem in Rio Grande do Sul.
The grass species is considered a pest on rural properties and depends on rapid management to avoid an infestation, since a single plant is capable of producing around 14 thousand seeds per year, with 90% viability.
An AI model is already being trained by researchers to detect the presence of the grass. The technology can also be replicated for other invasive species, such as pigweed. According to the research coordinator, an image can be processed in three milliseconds, which allows for the analysis of approximately 350 photos per minute.
Currently, the model's accuracy is around 88%, but the group is working to bring the algorithm to a 98% accuracy rate, which Camargo considers entirely feasible with the available resources. "It's just a matter of time before we reach that result."
The idea is that the products, which are already being field-tested experimentally through Embrapa, will soon be available on the market. "It's a very easy-to-scale technology because it's essentially software that can be installed on an Android phone or a solar-powered minicomputer."
gazetadopovo
