Artificial intelligence makes mapping agricultural intensification in the Cerrado more precise

• Pioneering methodology was developed with algorithms for the digital classification of satellite images based on Artificial Intelligence (AI).

• Application in image analysis of Cerrado areas in the limits of Sorriso, MT, Brazil allowed accuracy of up to 97%.

• Such precision favors studies and monitoring of land use and agricultural intensification.

• The distinct approach of the methodology, marked by the harmonization of Landsat and Sentinel-2 data, combined with the use of algorithms enabled the identification of areas with up to three harvests in the same crop year, which is not common in most existing mappings.

• The methodology was developed by researchers from Embrapa, the State University of Campinas (Unicamp), the National Institute of Spatial Research (INPE) and the Federal University of Uberlândia (UFU)..

• The work is published in the International Journal of Geo-Information (IJGI). Details of the methodology, findings and resulting maps can be found on Embrapa's Repository of Research Data (Redape).

HLS image cutout, true color composition (Bands 4-3-2), January 26, 2022 in Sorriso, MT, Brazil    Photo: Embrapa

A pioneering methodology, developed with the support of Artificial Intelligence (AI), allowed the achievement of an accuracy level of up to 97% when applied the anaysis of satellite images of Cerrado areas in the town limits of Sorriso, MT, one of the main agricultural focal points in Brazil. Accuracy is a relevant aspect in surveys through remote sensing.

The tool attributes higher precision to studies, monitoring and planning related to land use and the practice of agricultural intensification, and contributes to public and private decision-making based on qualified geospatial information.

The methodology was developed with AI-based digital satellite image classification algorithms. It stems from the work of researchers from Embrapa, the State University of Campinas  (Unicamp), the National Institute of Spatial Research (Inpe) and the Federal University of Uberlândia (UFU), published in the International Journal of Geo-Information  (IJGI), in July 2023, with free access for the general public.

"The results demonstrate the robustness of the methodology that was developed with a focus on the identification of land use dynamics processes like agricultural intensification", evaluates Édson Bolfe, a researcher at Embrapa Digital Agriculture, and coordinator of the project Agricultural Mapping in the Cerrado through a combination of multi-sensor images - MultiCER, funded by the São Paulo Research Support Foundation (Fapesp).

Bolfe explains that one of the main differentials of the methodology is the generation of a expanded geospatial database based on harmonized images from the United States National Aeronautics and Space Administration's (NASA) Landsat and European Space Agency's (ESA) Sentinel-2 satellite images (an initiative also called HLS, acronym for Harmonized Landsat Sentinel-2), combined with the use of AI-based digital classification algorithms. The approach enabled mapping crops at three different hierarchical levels, and indicated areas with one, two and up to three harvests in the same crop year.

The succession of harvests of different crops in the same area and in the same agricultural calendar, aiming to increase production without suppressing native areas, is a growing practice in Brazil, and its mapping and monitoring can guide decision makers in analyzes aimed at agri-environmental planning, in particular.

Synthesis of the methodological approach to map agricultural intensification.

Speed and precision, the role of AgroTag

Remote sensing products and AI models for pixel-by-pixel image classification have demonstrated high reliability in agricultural mapping, Bolfe explains. With HLS it is possible to obtain up to two images per week in the same agricultural regions of interest.

One of the challenges of the research team lies in obtaining qualitative and quantitative field information, which is fundamental for remote sensing in agriculture. For this purpose, the researchers used the app AgroTag, which was developed by Embrapa Environment to map the main crops at regional and national scales with speed and precision.

″AI-based algorithms heavily rely on a massive amount of input data to perform so-called 'machine learning'. The latter comprises processes in which reference sample data ('ground truths') are used to train the algorithms to identify targets under investigation in large areas, in this case using satellite images, that is, large-scale mapping″, comments Luiz Eduardo Vicente, a researcher at Embrapa Environment, a remote sensing expert, and one of the coordinators of the AgroTag project.

In this sense, according to Vicente, the use of AgroTag was fundamental, as it allowed the quick and accurate collection of field information, such as the type of land use and land cover in each sampling point, and their automated transfer to the online data cloud, enabling their use in said algorithms.

In contrast with traditional collection methods, AgroTag showed a 25% increase in sampled areas during the project. "The project reaffirms one of the reasons why Agrotag was created," Vicente states.

Examples of second and third crops produced in the same crop year, in Sorriso, MT, in June 2022. Photo Credit: Taya Parreiras, 2022.

Dynamic mappings

"The study mapped agricultural production in the 2021-2022 crop year in Sorriso, MT, a town chosen due to its economic and agri-environmental relevance in the context of the Cerrado and the country," observes Edson Sano, a researcher at Embrapa Cerrados and member of the MultiCER project.

Most existing mappings do not follow the evolution of ″land-saving″ agricultural intensification practices - such as the production of up to three crops in the same area - staying at the level of the first crop. "Some surveys evolved to identify the number of crops planted; however, they do not detect specific crops," Sano concludes.

″In order to produce dynamic, detailed and accurate mappings, a large volume of 'ground truth' information is needed, which are samples labeled according to the types of land use or land cover that are obtained during the field activities,″ notes Taya Parreiras, a PhD student at Unicamp's Institute of Geosciences and a member of the MultiCER Project.

According to the researcher, it also requires regular time series of high temporal resolution satellite images; and, to that end, the harmonization of Landsat and Sentinel-2 data is a special approach. Parreiras points out that in order to deal with the volume of such databases and information, machine learning algorithms like 'Random Forest' or 'Extreme Gradient Boost' are fundamental.

″As part of the AI, such algorithms can analyze and learn complex spectral and textural patterns from extensive and varied agricultural datasets, enabling the accurate identification of different crop types, soil conditions, and environmental variables,″ she argues.

As 'Random Forest' creates multiple independent decision trees and combines them, it can produce more reliable estimates. 'Extreme Gradient Boost' also creates several decision trees, but with the advantage of allowing those with low predictive power to be adjusted. "Both algorithms are highly scalable, which allows them to quickly process large volumes of data, contributing to the generation of detailed up-to-date agricultural maps," he concludes.

Digital classification of land use and land cover, with emphasis on second crops in Sorriso, MT, Brazil, in January 2022, generated from HLS images and AI-based algorithms.

Read the '2023 Latin America Focus' magazine.

If you would like to share your company's story. Please contact Christina Xie at christina@agropages.com