Drone AI spots cattle sale window, saves 15,000 reais

A drone-and-AI system saved 15,000 Brazilian reais in a 110-head Nelore test by identifying the best slaughter window a week earlier than conventional management projected; scaling the result to a 14,600-head feedlot could yield about 2 million reais.

Researchers supported by the São Paulo Research Foundation (FAPESP) described the system in Computers & Electronics in Agriculture. Jayme Garcia Arnal Barbedo, a researcher at Embrapa Digital Agriculture, one of the decentralized units of the Brazilian Agricultural Research Corp. (EMBRAPA), presented it Oct. 28 at FAPESP Week Netherlands. FAPESP reported the development Oct. 30, 2026.

Barbedo said: “We’re very close to obtaining a complete, functional prototype of the system. The next crucial step will be to establish partnerships with companies in the sector or ag techs [startups in the agricultural sector] to transform that research into a commercially available product for ranchers.”

What changes for ranchers?

The system uses aerial images captured by drones flying 15 meters above feedlots. Deep-learning algorithms segment each animal’s body and extract morphological measurements that map the herd’s fattening curve in real time. Field validation put the optimal sale date up to a week ahead of conventional methods.

Two AI models do the work. The first, fine-tuned to locate each Nelore steer individually, achieved 95.56% accuracy. The second delineates body contours at the pixel level.

From those outlines, the model calculates each steer’s width relative to length. As muscle and fat develop, the animal becomes wider relative to its length. The ratio resists small drone-altitude changes and tracks gain without depending on absolute size in the photograph.

Barbedo described the biology this way: “The weight gain curve of a group of animals intended for slaughter has a characteristic shape of an ‘S’: It starts slowly during the adaptation phase, enters a period of very accelerated gain and then slows down in the final stage.”

Once that final stage begins, additional feed no longer produces meat at the same rate, Barbedo said. He estimated that, sometimes, a one-day difference in the decision to slaughter or sell the herd can increase an operation’s profit by 5% to 10%. That changes the sale math.

For ranchers, the practical benefit is less need to pen and weigh every animal. The drone replaces routine physical restraint with aerial observation.

Traditional measurement forces cattle through a narrow restraint chute for individual weighing, requiring labor and risking stress, injuries, bruises and equipment damage. Ranches usually weigh only on feedlot entry and again months later at shipment, leaving managers without reliable weekly weight data for weeks.

To keep the average reliable, the AI analyzes the entire batch and rejects animals that are lying down or held at unfavorable angles. It retains only standing, straight animals within the central window of the image, reducing the chance that temporary poses distort the herd estimate.

A 112-day ranch test

Campanário Ranch, in Laguna Carapã, Mato Grosso do Sul, housed approximately 14,600 head in a feedlot system. Researchers monitored 110 Nelore cattle with an initial average age of 18 months over a 112-day fattening cycle. Twelve aerial missions, spaced every 10 days, fed data into the model.

The mathematical model identified three biological phases: adaptation, rapid growth and the economic saturation plateau, when an animal reaches its body-weight limit. The system detected the herd’s weight-gain deceleration one week before conventional management projected, flagged the optimal slaughter window and prevented unnecessary feed consumption.

The 110-animal batch produced 15,000 reais in savings. The roughly 2 million reais for the ranch’s 14,600-animal capacity represented an extrapolation of operational and financial gains, not a reported ranch-wide saving. The first is a test result. The second is an extrapolation.

The published figures did not disclose feed prices, baseline intake or the calculation method behind either estimate, so producers should test the assumptions against their own costs, feed-conversion records and margins.

During field tests, analysts evaluated flights at 5, 10 and 15 meters. At 15 meters, the cattle did not notice the drone and continued grazing and resting without an impact on animal welfare.

Beyond sale timing

The research team also is adapting the platform to identify stress-related behaviors in feedlots, including disputes over feed and crowding at troughs. Images showing animals unable to line up side by side can indicate inadequate space or pushing during feeding.

EMBRAPA researchers also are studying dairy applications for early mastitis detection and analysis of dairy cows’ feeding patterns, prompted by requests from industry and international partners.

The work ran under the auspices of Semear Digital, the FAPESP-supported Science Center for Development in Digital Agriculture. Four other FAPESP projects funded it: 23/03870-8, 24/15430-5, 25/05985-2 and 25/10602-5.

Barbedo described technology transfer to the market as the next step. The next test is whether partnerships with cattle-sector companies or ag-tech startups can turn the near-functional prototype into a product ranchers can buy.