Brazil-Crop Dataset

A benchmark for agricultural remote sensing applications

The Brazil-Crop dataset is a comprehensive, publicly available benchmark designed for tropical and subtropical agricultural remote sensing applications. Developed to address the critical gap in open-access ground truth data for the Geospatial Artificial Intelligence (GeoAI) era, it provides seasonal land use information across two distinct and highly dynamic agricultural regions in Brazil: Paraná (Atlantic Forest biome) and Mato Grosso (Cerrado biome).

Project Overview

This dataset captures the complex temporal dynamics of Brazilian agriculture, including double-cropping systems, across two major growing seasons. It encompasses a wide range of agricultural profiles, representing the diversity of management practices and environmental contexts found in tropical Brazil.

  • Field Metrics: 9,448 field observations across 5,477 unique field geometries.
  • Visual Evidence: 2,242 geotagged photographs providing in situ ground truth.
  • Land Use Diversity: 30 distinct land use classes, including soybean, corn, sugarcane, cotton, and various perennial crops.
  • Resolution: Field boundaries precisely delineated using 10m Sentinel-2 imagery.

Technical Specifications

| Feature | Details | | :— | :— | | Subject Area | Remote sensing, GIS, Land Use and Land Cover | | Data Format | Vector (Shapefile, GeoPackage, Parquet) and Field photography | | Spatial Reference | WGS 84 (EPSG: 4326) | | Accessibility | Openly available under Creative Commons Attribution 4.0 | | DOI | (Under Construction) |

Methodology

The data collection followed a rigorous three-step workflow:

  1. Field Campaigns: Four extensive field trips (Jan 2020, Aug 2020, Feb 2021, and Jun 2021) to collect coordinates, photographs, and crop classes.
  2. Boundary Delineation: Field polygons were created using Sentinel-2 (10m) imagery as a background to ensure precise geometry.
  3. Label Verification: Each polygon was labeled and cross-verified using Sentinel-2 and MODIS NDVI time series to avoid annotation errors.

Related Research & Data

  • Zenodo Repository: (Under Construction)
  • Oldoni et al. (2025) — Brazil-Crop dataset: a benchmark for agricultural remote sensing applications (Under Construction).
  • Previous Databases: This project complements the LEM+ and Campo Verde initiatives.

This work was supported by CAPES, CNPq, and the Brazil Data Cube project.

References