System anticipates soybean productivity using AI and satellites.

The combination of satellite imagery, climate variables, and artificial intelligence is gaining ground as a tool for predicting agricultural productivity in a scenario marked by climate change and extreme events. Brazilian researchers have developed a computational model capable of estimating the productivity of soybean crops in the Midwest even before harvest, using satellite data, meteorological information, and machine learning.

The study, conducted by master's student Ester de Carvalho Pereira from the Luiz de Queiroz School of Agriculture at USP (Esalq/USP), was published in the scientific journal Big Earth Data e The study analyzed municipalities in Goiás, Mato Grosso, and Mato Grosso do Sul between the 2019/2020 and 2021/2022 harvests. The developed model achieved 72% accuracy and an average error of less than 302 kilograms per hectare in productivity estimates.

The research used images from the Sentinel-2 satellite, climatic variables, and historical data from the IBGE (Brazilian Institute of Geography and Statistics) to build predictive models capable of monitoring different phases of soybean crop development. This work is part of the project “Precisia – Harvest Prediction using Satellite Imagery and Artificial Intelligence,” funded by the CNPq's RHAE Program and coordinated by the company Espectro Ltda.

According to Michel Eustáquio Dantas Chaves, a researcher at the São Paulo State University (Unesp) in Tupã and one of the study's authors, advances in agricultural monitoring technologies have significantly expanded the capacity for crop analysis. "Satellite data allows us to monitor harvests and production cycles, something that until recently was unfeasible, especially at the crop level," he states.

AI assists in data identification.

Researchers highlight that the use of artificial intelligence was fundamental in identifying which climatic and spectral factors had the greatest impact on productivity. Among the most relevant variables were accumulated precipitation, solar radiation, and water deficit. In satellite images, bands related to infrared and the so-called red edge, a spectral range associated with the photosynthetic activity of plants, had greater weight.

In total, six distinct models were developed, considering periods between 30 and 180 days after planting. The best performance was obtained in the 150-day model, a phase corresponding to the soybean grain filling stage, considered crucial for determining final productivity.

The research comes at a time of greater climate instability for Brazilian agriculture. In the 2023/2024 harvest, for example, extreme heat events and irregular rainfall significantly reduced the productive potential of soybeans in the country. Initially estimated at 162 million tons by Conab, production ended up closing at 147,7 million tons, according to data cited by the FAO, representing losses of billions of dollars to the sector.

Despite the progress, researchers point to structural limitations in increasing the accuracy of the models in Brazil. One of the main difficulties was the lack of detailed data at the rural property level, which forced the group to use public databases such as IBGE and MapBiomas to identify cultivated areas.

“The biggest challenge is having data from that specific location, but we work with what we have,” says Ana Cláudia dos Santos Luciano, a researcher at Esalq/USP and supervisor of the study. According to her, although the system does not yet offer sufficient detail for individual applications on properties, it has great potential for public policies, regional monitoring, and agricultural planning.

Integration of information

Researchers also warn of the need for greater integration between Brazilian public databases and for strengthening the funding of agricultural monitoring programs. According to Chaves, geoprocessing and artificial intelligence tools already allow for increasingly precise analyses, but many projects face budgetary constraints.

Currently, the group continues to refine the models and develop new applications for other crops, such as sugarcane. In addition, periodic soybean monitoring bulletins are already being produced for states like Goiás, Mato Grosso do Sul, and Paraná.

More information at doi.org/10.1080/20964471.2026.2631900