Early-Season Field Reference Dataset of Croplands in a Consolidated Agricultural Frontier in the Brazilian Cerrado

Abstract

This dataset presents field observations collected in the municipality of Goiatuba, Goiás State, Brazil, a consolidated and representative agricultural frontier of the Brazilian Cerrado biome. The region presents diverse land use dynamics, including annual cropping systems, irrigated fields with up to three harvests per year, and pasturelands. We conducted a field campaign from 3 to 7 November 2025, corresponding to the beginning of the 2025/2026 Brazilian crop season, when crops were at distinct early phenological stages. To ensure representativeness, we delineated 117 reference fields prior to the field campaign, and an additional 463 plots were surveyed during work. Geographic coordinates, crop types, and photographic records were obtained using the GPX Viewer application, a handheld GPS receiver, and the QField 3.7.9 mobile GIS application running on a tablet uploaded with Sentinel-2 true-color imagery and the municipal road network. Plot boundaries were subsequently digitized in QGIS Desktop 3.34.1 software, following a conservative mapping strategy to minimize edge effects and internal heterogeneity associated with trees and water catchment basins. In total, more than 26,000 hectares of agricultural fields were mapped, along with additional land use and land cover polygons representing water bodies, urban areas, and natural vegetation fragments. All reference fields were labeled based on in situ observations and linked to Sentinel-2 mosaics downloaded via the Google Earth Engine platform. This dataset is well-suited for training, testing, and validation of remote sensing classifiers, benchmarking studies, and agricultural mapping initiatives focused on the beginning of the agricultural season in the Brazilian Cerrado.

Dataset: Goiatuba Early Season—https://doi.org/10.5281/zenodo.17612590.

Dataset License: Creative Commons Attribution 4.0 International license (CC BY 4.0)

1. Summary

Population growth and rising food demand have intensified global concerns related to food security [1]. Addressing this challenge requires increasing agricultural productivity without proportionally expanding cultivated land, supported by scientific research and technological innovations that promote environmentally sustainable intensification. In this context, remote sensing and machine learning have emerged as key tools for monitoring agricultural dynamics because of the increased availability of satellite data and the development of robust supervised classification algorithms [2].

Accurate reference data are critical for crop type classification, yet such data remain scarce and are often costly and time-consuming to collect [2,3]. In tropical agricultural regions, where rapid crop rotations and highly dynamic land use practices are common, the lack of spatially and temporally consistent in situ observations represents a major constraint for the development and validation of remote sensing-based classification algorithms [4].

Recent initiatives in Brazil have demonstrated that open-access in situ datasets can significantly accelerate the development of crop classification algorithms and support benchmarking efforts [5,6,7]. However, the most recent multiple crop datasets made available in Brazil date back to 2020 [7], which limits their applicability to current agricultural conditions. To the best of our knowledge, despite the recognized importance of early-season mapping for in-season crop classification [8], there are currently no publicly available in situ datasets collected during the early season in recent years in Brazil, largely because of high costs associated with field data acquisition [2,3]. One exception is the dataset published by Parreiras et al. [9], which provides sample points from fieldwork conducted in 2023, although it focused specifically on coffee.

To address this gap, we selected the municipality of Goiatuba, located in southwestern Goiás State, Brazil, to conduct a field campaign with the objective of producing a detailed reference dataset that can be freely used for multiple agricultural and remote sensing applications. This municipality represents a consolidated, large-scale agricultural frontier characterized by the use of large-scale mechanization, double-cropping systems, and high management intensity for food and energy production in the Brazilian Cerrado [10,11]. Nearly all farmers employ a no-till farming system over large rural properties, where deep, well-drained, and naturally acidic soils predominate [12]. The area corresponds to a representative region of the consolidated Cerrado agricultural model that emerged in the 1970s, marked by widespread adoption of annual crops, transition to intensive mechanized systems, and strong integration with agroindustry and global commodity chains. Other consolidated agricultural frontiers in the Brazilian Cerrado, such as the ecotonal region between the Cerrado and the Amazon in Mato Grosso State and western Bahia State, present similar agricultural practices, crop types, and, to some extent, climatic seasonality. As a result, these regions offer strong potential for spatial transferability of methods and findings obtained in this study [13,14].

The dataset documents field conditions at the beginning of the 2025/2026 Brazilian crop season, providing essential ground information to support local research initiatives and broader methodological advancements in agricultural mapping. Early-season crop mapping in tropical regions is especially challenging because of the persistent cloud cover in optical satellite images, the low vegetation canopy typical of early crop growth stages, and variations in solar illumination [15]. The dataset adheres to the Findable, Accessible, Interoperable, and Reusable (FAIR) principles [16]. The application of FAIR principles has been increasingly emphasized in recent agri-food datasets to ensure long-term interoperability and usability over diverse analytical workflows [17,18]. By providing interoperable and openly accessible reference information, this dataset helps address existing limitations in the availability of ground truth data, fosters the development of more robust classification models, and serves as a benchmark for evaluating and comparing different classification algorithms and workflows in agricultural mappings.

2. Data Description

This paper presents the Goiatuba Land Use and Land Cover Dataset, made available as a single GeoPackage (.gpkg) file and deposited on the open-access Zenodo platform (https://doi.org/10.5281/zenodo.17612590). We selected the GeoPackage format because it is open, non-proprietary, and standards-based, allowing efficient data management while preventing file fragmentation, a common limitation of other formats. The GeoPackage is also widely supported by major geographical information system (GIS) software packages [19]. The dataset includes plot boundaries and reference information collected during the 2025/2026 Brazilian crop season in the municipality of Goiatuba, Goiás State, Brazil.

The main layer contains 580 plots, each labeled according to the observed land use and land cover (LULC) category. Figure 1 illustrates the spatial distribution of these plots, while Figure 2 presents panoramic photographs depicting the main LULC classes identified in the study area. Each feature represents an individual field polygon, with its corresponding attributes organized in columns within the attribute table (

Table 1. Name of field attributes and their corresponding description.

Feature	Description
fid	Unique identifier of each plot
in_situ	Crop type or land use and land cover observed in situ
date	Date of the plot observation
area_ha	Polygon area in hectares
reference	Identifies if the polygon had GPS coordinates

).

Ancillary layers representing water bodies, urban areas, fragments of natural vegetation, and other landscape features were also included. These layers were refined using high-resolution satellite images available on the Google Maps platform to support the identification of human infrastructure (e.g., sugarcane processing facilities). In addition, Global Positioning System (GPS) coordinates of reference points collected during the field campaign were incorporated. Overall, this dataset provides standardized and spatially explicit reference information suitable for crop type mapping and LULC supervised classifications, including applications that integrate multiple sensors (e.g., combinations of optical and radar data).

3. Methods

We collected field data in the municipality of Goiatuba, Goiás State, located in Brazil’s Central-West Region. The municipality comprises 2475 km² in a consolidated region of the Brazilian agricultural frontier [20], which motivated its selection as a representative agricultural area in the Brazilian Cerrado. This region plays a significant role in national crop production and in global climate mitigation discussions, given the increasing pressure on tropical croplands to support land-based strategies aligned with the 1.5 °C climate target [21].

The main land use practices in this region include annual cropping systems, mainly soybean as the first crop and maize as the second, as well as sugarcane plantation, cultivated pastures dominated by Urochloa species, and irrigated croplands operated with a center pivot irrigation system during the dry season. The remaining native vegetation is composed mainly of forest formations, Cerradão, according to the Brazilian vegetation classification system for the Cerrado biome, and riparian forest fragments [11]. In this biome, the six-month rainy season typically ranges from October to March, while the dry season ranges from April to September [22]. The mean annual rainfall is approximately 1500 mm [23].

The field campaign was conducted from 3 to 7 November 2025, a period corresponding to the beginning of the rainy season in the region, when most crops had already been planted and were in their early phenological stages. For each reference plot, geographic coordinates were recorded using a Garmin 78S GPS receiver, expressed in decimal degrees and referenced to the World Geodetic System 1984 (WGS-84; EPSG: 4326). The Garmin 78S provided horizontal accuracies between 3 m and 5 m, consistent with manufacturer specifications [24]. No additional control-point validation was performed, as the dataset was designed as an observational, non-probabilistic survey. For each site, we also identified the crop type and acquired panoramic photographs to document local field conditions.

The open-source QField 3.7.9 mobile GIS application, installed on a 14.5-inch tablet, was used to record photographs and annotate relevant plot information, including geographical coordinates and the corresponding LULC type. Prior to the field campaign, we uploaded the following background data to the application: 117 point-based reference coordinates planned for visitation; official federal and state road maps for the municipality, complemented by the OpenStreetMap road network; and an RGB color composite mosaic derived from Sentinel-2 Harmonized Surface Reflectance imagery acquired on 23 October 2025 and 2 November 2025. These mosaics were accessed through the Google Earth Engine platform and include atmospheric correction, radiometric harmonization between Sentinel-2A and Sentinel-2B, and 10 m spatial resolution for the RGB bands.

The European Space Agency’s Sentinel-2 Multispectral Instrument (MSI) mission provides imagery with 12-bit radiometric resolution and 13 spectral bands located in the visible, near-infrared, and shortwave infrared regions [25]. Depending on the spectral band, images are delivered at spatial resolutions of 10, 20, or 60 m. The mission consists of two satellites with individual revisit cycles of 10 days, allowing a combined revisit frequency of five days at the equator.

After the fieldwork, we used the same Sentinel-2 images to manually delineate plot boundaries in QGIS Desktop 3.34.1. Polygons were manually digitized at scales ranging from 1:2000 to 1:3000, ensuring precise vertex placement over spectrally homogeneous areas. Following a conservative mapping strategy, plot boundaries were delineated slightly inside the visually identifiable edges, typically by 10 to 20 m (1 to 2 Sentinel-2 pixels), to minimize spectral mixing with adjacent land use classes. For cropland plots, internal non-crop elements, such as isolated trees, small water catchment basins (Figure 3), or other anthropogenic or natural features, were removed whenever their pixels displayed visual characteristics distinct from the dominant spectral pattern of the field.

This procedure aimed to avoid the inclusion of mixed pixels that could bias subsequent analysis. In contrast, for pasture plots, internal trees are inherent elements and were therefore retained during the delineation process. This distinction ensured that the mapped polygons accurately reflected the characteristic spatial heterogeneity of each land use type.

We then linked the collected point data to the final polygons, resulting in more than 26,000 hectares of delineated plots in the municipality. To support broader applications of the dataset, we also provided GPS coordinates for all reference fields, along with polygons representing water bodies, urban areas, and fragments of natural vegetation. These ancillary features were delineated using the same procedures to facilitate their integration into LULC classification workflows. Because the survey was not designed to statistically represent the full land use distribution of the municipality, no spatial representativeness metrics, such as χ² tests, are reported. Instead, the dataset documents the conditions of the fields visited during the early 2025/2026 season.

4. User Notes

This dataset was prepared in accordance with FAIR principles, ensuring that the information is retrievable, accessible, interoperable, and reusable across a wide range of remote sensing and geospatial workflows. Users can work directly with the field polygons provided or generate point-based samples within each polygon for applications requiring finer spatial granularity, such as data augmentation, training machine learning and deep learning models, or assessing within-field variability.

When integrating these data with satellite products, users should ensure temporal correspondence between the in situ observations and the image acquisition dates. Transfer learning approaches may also be applied spatially, by extending analyses to areas with similar consolidated agricultural characteristics, and temporally, by examining spectral behavior in previous years. This is facilitated by the use of standardized attribute naming conventions, which enhance interoperability with analytical environments such as QGIS Desktop 3.34.1, Python 3.14, and R-4.5.2. We emphasize that the field survey followed an observational, non-probabilistic design; therefore, no formal accuracy metrics (e.g., GPS root mean squared error or boundary uncertainty estimates) are provided.

To support a wide range of in-season time-series analyses, we recommend using satellite imagery acquired from September to March, which corresponds to the first harvesting period in the region [26]. Although the field data were collected during the first week of November 2025, shortly after major soybean planting, the crop is expected to remain in place until the harvest in February. Therefore, imagery acquired throughout this window still represents the same fields and phenological development, enabling analyses of crop growth dynamics, phenology, biomass accumulation, and other time-series applications.

In the Central-West Region of Brazil, the crop calendar is mainly shaped by the onset of the rainy season, which typically begins between September and October. In 2025, this onset was slightly delayed due to irregular rainfall. Nevertheless, farmers in the study area commonly use millet or sorghum as cover crops and apply desiccation practices prior to soybean planting, ensuring that the sowing remains within the standard seasonal window. This context helps users interpret the temporal behavior of the fields and adapt similar analyses to regions with different climatic conditions or management practices.

A multisensor approach is also encouraged. As illustrated in Figure 4, a center pivot irrigated field displays distinct temporal and spectral behavior when observed using Sentinel-2 Normalized Difference Vegetation Index (NDVI) data and Sentinel-1 VV-polarized backscatter imagery. For analyses of subsequent harvests, users may extend the temporal window to encompass the entire 2025/2026 Brazilian crop calendar, thereby allowing investigations of long-cycle crops as well as double- and triple-cropping systems.