Abstract

The Brazilian Agricultural Census conducted by the Brazilian Institute of Geography and Statistics (IBGE) is one of the largest in the world, covering a territory of 8.5 million km² and encompassing more than 5 million agricultural holdings. As a cornerstone of the national economy, the sector accounts for nearly 30% of the GDP and is characterized by a sharp dynamism, with constant increases in productivity and rapid changes in land use. To ensure the accuracy of this large-scale operation, pilot tests are conducted to validate planning assumptions, the collection system, and the overall workflow.

This initial test was conducted across six municipalities in four Brazilian states, encompassing the Northeast and Southeast regions. The operation involved approximately 200 collaborators and resulted in the collection of 1,200 questionnaires over a 12-day period.

IBGE is transitioning its Agriculture Survey operations toward an earth-observation-driven design. As a result, one of the primary innovations tested was the use of automated parcel delineation (agricultural plot segmentation), aimed at improving planning, refining agricultural holding characterization, and generating a robust database for the supervised training of machine learning models. The ultimate goal is to enable automated crop mapping and production yielding for future and ongoing Brazilian agricultural surveys.

The pilot test sought to evaluate four fundamental pillars: the cognitive capacity and willingness of producers to identify their land on a map; the accuracy of automated mapping compared to territorial reality; the time required for field collection; and the functionality of the data collection application. Consolidated results both qualitative, obtained through observation reports, and quantitative indicate the success of the proposal. Respondents were receptive and demonstrated the ability to identify their plots on mobile devices.

Although automated mapping correctly delineated productive areas, the use of optical imagery from the Sentinel-2 constellation, combined with the automated plot detection model, imposed an analytical limitation for areas smaller than 1 hectare. Additionally, challenges were observed in more arid zones, where the current method and the scarcity of training data produced automated detection results that still require improvement. Conversely, collection time exceeded expectations, frequently taking less than two minutes per questionnaire. The flaws identified in the application’s data flow have already been mapped for correction.

With the 12th Agricultural Census scheduled for 2027, there is sufficient time to implement the necessary improvements. This innovation will provide a qualitative leap in data granularity, utilizing Earth observation as a strategic input across all phases of the statistical operation from real-time supervision to the final data analysis.