Embrapa is developing a model that estimates sugarcane productivity with high accuracy using satellite imagery collected throughout the crop growth cycle. By integrating these images with statistical techniques and machine learning, researchers achieved precise predictions. The same methodology was also tested on soybeans, serving as a validation for the newly released biostimulant Hydratus.
The research relies on daily PlanetScope images provided through the Brasil Mais Program of the Ministry of Justice and Public Security. These time-series images allow researchers to identify optimal stages of plant development to calculate vegetation indices used in productivity forecasts. The satellite data is combined with variables such as cultivar, production cycle, and accumulated precipitation to feed the prediction model.
For sugarcane, a collaboration with the Cooperative of Sugarcane Planters of São Paulo (Coplacana) and funding from Finep monitored two crops over three years, achieving a coefficient of determination of 0.89. This means the model’s predictions matched 89% of the productivity measured through traditional agronomic methods — an impressive level of accuracy for forecasting, according to a press release.
Geraldo Magela Cançado, researcher at Embrapa Digital Agriculture, explains that the project started with a simpler model, but new variables such as temperature, soil texture, and water availability will be added to further improve its efficiency.
The research team envisions a tool that can be used by producers and agribusinesses to generate plot-level forecasts, enabling better strategic planning, logistics management, and guidance for interventions. Additionally, the model could support government crop forecasting and national agricultural planning.
“This methodology allows a more objective harvest survey. We want to reduce the subjectivity of this forecast and be more comprehensive. Considering the immensity of this country, only with the use of satellite images does this become possible,” says researcher João Antunes.
Applying the Methodology to Soybeans
Following the sugarcane study, Embrapa applied the same methodology to soybean cultivation to validate the effects of the biostimulant Hydratus, which enhances growth and protects plants against drought. Funded by Finep, the research was conducted in partnership with Embrapa Corn and Sorghum and the company Bioma, monitoring three different fields. In two areas, PlanetScope satellite images were used, while in the third, data was collected via drone imagery.
For sugarcane, productivity forecasts relied on the green normalized difference vegetation index (GNDVI), which uses near-infrared (NIR) and green spectral bands to detect chlorophyll differences. In soybeans, the team used the enhanced vegetation index (EVI2), combining red and NIR bands to capture plant structure and biomass.
The results highlighted differences in productivity across treatments with varying doses of Hydratus compared to control plots. The model achieved a 71% correlation between predicted and observed productivity. While slightly lower than the sugarcane model, this level of accuracy is still considered high and promising for forecasting soybean yields.
“Each culture has a different behavior and this variation between them is normal. Overall, we assume as acceptable correlation levels above 0.6 (i.e., the model is able to explain above 60% of the observed variation). In the case of sugarcane, as production is very linked to the plant’s own canopy (part of the plant on the soil surface, formed by leaves and stems), better results are obtained, as it is almost a direct relationship between biomass and stale productivity (typical stem of grasses, such as sugarcane). In the case of soybeans, as the product is the grain, the relationship between soybeans and productivity is not so direct,” explains Geraldo Cançado.
The good results of the prediction model bring optimism for use in field research, allowing precise and non-destructive monitoring.
“This double evaluation framework, combining agronomic metrics with remote sensing, provides an innovative and economical strategy for evaluating the performance of crops in real time,” says the researcher.
The research has employed two complementary approaches: one using machine learning and the other based on traditional statistical methods. According to Embrapa analyst Eduardo Speranza, because the dataset used to train the algorithm is still relatively small, the statistical model has so far produced more accurate predictions.
“Despite having many experiments, we worked on a publication with 500-600 samples to train an algorithm. This amount for machine learning is small. The machine learning method has the potential to be better, but it needs thousands of samples,” Speranza explains, recalling that the increase in samples depends on on-site validation by the agronomic monitoring method.
