From the list of attributes identified from the literature as potentially related to forest-fire analysis ( Table 2), not all of them are available for the entire Amazon Rainforest, let alone for the study period 2001-2020. Specifically, variables such as protected zones, isothermality, and maximum cumulative water deficit were only available for certain regions and for a particular time period. Similarly, elevation-related attributes were only available for certain years between the period 2001-2020. In this study, attributes that were available for the complete Amazon region and for the selected time period of 2001-2020, are identified and further acquired, as detailed in Table 3, with Date of Access: 01 May 2022. This section details the complete data-acquisition process related to each collected attribute.

Burnt Area (BA)

The data product acquired was MODIS/Terra+Aqua Direct Broadcast Burned Area Monthly L3 Global 500 m SIN Grid V006 MCD64A1 Version 6.1, which is a gridded burnt area product at a resolution of 500 m, available in Hierarchical Data Format (HDF) format. The product provides the date of burn (in the form of the day of the year) for individual cells with additional classes, such as unburnt, missing data, and water. The data product is available for the period 2000 to the present (2022), with global spatial coverage in the form of regional subsets. The layers extracted from the data source are for regions 5 and 6, which cover the Amazon area. The data layer values are in units of a day, with a valid range of data values as between 1-366 (representing the day of the year). Further details related to the product, including the quality assessment and known issues, are available at MODIS MCD64A1 ( https://lpdaac.usgs.gov/products/mcd64a1v061/ ).

As the burnt area product is available at the regional level, an additional data processing step for the burnt area product is the merging of two separate regional-level products to cover the entire region of the Amazon basin boundary. Additionally, the data were re-classified to assign a single value of 1 to all burn dates (1-366) to identify the cell with burn data as simply burnt. Hence, working data has four classes (burnt, unburnt, missing, and water) with values (1, 0, -1, and -2), respectively.

Land Cover (LC)

The data product acquired was MODIS/Terra+Aqua Land Cover Type Yearly L3 Global 0.05Deg CMG V006 MCD12C1 Version 6, which consists of three gridded land cover classification schemes at a resolution of 5,600 m, available in the HDF format. The three available classification schemes include Maps of the International Geosphere-Biosphere Programme (IGBP) providing 17 classes, University of Maryland (UMD) providing 16 classes, and Leaf Area Index (LAI) providing 11 classes. LAI classification schemes are extracted from the data product as 11 classes are sufficient for representation of different land covers in terms of Water, Urban, Forest, Grassland, etc., and additional classes available in other schemes are further subdivisions of forests and grassland types. The data product is available for the period 2000 to the present (2022) with global spatial coverage. The details of the land cover classes of the LAI scheme are provided in Table 4. The name of the layer extracted from the data source is Land Cover Type-3, with a range of data values between classes 0 and 10. Further details related to the product, including the quality assessment and known issues, are available at MODIS MCD12C1 ( https://lpdaac.usgs.gov/products/mcd12c1v006/ ).

Precipitation

The data product acquired is Integrated Multi-satellite Retrievals for GPM (Global Precipitation Measurement)-based multi-satellite precipitation product, Version 06 B, available in Hierarchical Data Format version 5 (HDF5) format. The product provides a monthly product of average precipitation rates at a 0.1 °× 0.1 ° (approximately 10,000 m at the equator) spatial resolution, estimated from numerous precipitation-relevant satellite passive microwave (PMW) sensors. The dataset is available for 2000–2021 with global spatial coverage. The values are represented in millimeters per hour ( mm/hr), with a scale factor of 1000 and missing values marked with -9999. Thus, a value of 500 indicates 500/1000 mm/h. Further details related to the product are available at the GES-DISC GPM IMERG Final Precipitation L3 ( https://disc.gsfc.nasa.gov/datasets/GPM_3IMERGM_06/summary ).

Soil moisture

The data product acquired is a model-calculated (not directly observed) averaged soil moisture water height equivalent, namely CPC Soil Moisture Version 2, available in the GEOTIFF format. The data are a monthly product of 0.5 °× 0.5 °(approximately 37,000 m at the equator) spatial resolution, with data available from 1948 to the present (2022). The spatial coverage of the product is 89.75N–89.75S, 0.25E–359.75E. The values are represented in millimeters ( mm), with missing values marked as -9999. Further details related to the product are available at CPC Soil Moisture ( https: //psl.noaa.gov/data/gridded/data.cpcsoil.html ).

In the preprocessing of the Soil Moisture data product, data transformation is implemented as an additional step. As the source data have a spatial offset, not aligning with the reference base map, the data are transformed to correct alignment using the Geospatial Data Abstraction Library (GDAL). ¹³

Elevation

The acquired data product is a global multivariate package related to terrain features, which can serve many large-scale research publications. The data product is based on a 250 m Digital Elevation Model (DEM), available in Tagged Image File Format (TIF) format, from Global Multi-Resolution Terrain Elevation Data 2010 (GMTED2010). ¹⁴ This data product provides many topographic variables, such as elevation, slope, aspect, northness, elasticity, roughness index, and topographic position index at different resolutions of 1, 10, 50, or 100 km, with global spatial coverage; however, our focus is only on elevation. The Elevation values are represented in meters ( m). Further details related to this product are available at ( https://www.earthenv.org/topography ).

Land Surface Temperature (LST)

The data product acquired was MODIS/Terra Land-Surface Temperature/Emissivity Monthly Global 0.05Deg CMG MOD11C3 Version 6, which is a monthly Land Surface Temperature & Emissivity (LST&E) value product at a spatial resolution of 0.05 ° (approximately 5,600 m), available in the HDF format. The data product provides values for both daytime and nighttime observations, along with other details related to the quality assessment. The data product is available for the period 2000 to the present (2022) with global spatial coverage. The temperature values are represented in kelvin ( K), with a scale factor of 0.02 and a range of values between 7,500 and 65,535. Thus, the LST value equal to X represents X*0.02 kelvin. Further details related to this product are available at MODIS MOD11C3 ( https://lpdaac.usgs.gov/products/mod11c3v006/ ).

Specific humidity, Evapotranspiration (ET), wind and air temperature

The acquired data provides a set of parameters related to land surface observations. The data is a simulation-based product of the Noah 3.6.1, model from Famine Early Warning Systems, Network (FEWS NET) Land Data Assimilation System (FLDAS). All the provided variables are available as a monthly product in a 0.10 degree spatial resolution (approximately 1,000 m at the equator) and available (as a layer) in NETCDF file format. The dataset is available for the period from 1982 to the present (2022) with global spatial coverage. The values of Specific Humidity are represented as ( kg/kg), using a ratio between kilogram of water (moisture) per kilogram of air; whereas Evapotranspiration, Wind and Air Temperature are measured in ( kg/m ² s), ( m/s) and kelvin ( K), respectively. Further details related to the product are available at the GES DISC-FLDAS Noah Land Surface Model L4 ( https://disc.gsfc.nasa.gov/datasets/FLDAS_NOAH01_C_GL_M_001/summary ).

While these model-derived variables (including Soil Moisture) introduce inherent numerical uncertainties compared to direct field observations, they are incorporated to expand the suite of available environmental covariates, providing the multivariate depth necessary for robust spatiotemporal analysis across the Amazon. As the primary objective of this work is the curation and standardization of a basin-wide Amazon dataset, a formal independent uncertainty analysis remains beyond the scope of this data curation effort. By providing these products in a common, analysis-ready format, this work establishes the necessary foundation for future studies to conduct such analytical sensitivity assessments and empirical validations.

All of the various attributes collected in the database have different spatial resolutions, as described in Table 3. Similarly, not all variables are available at monthly resolution, as Land Cover and Elevation are annual and one-time, respectively. Moreover, all of these variables cover different spatial extents and have dissimilar spatial orientations. To obtain a dataset with all the variables at a fixed spatial extent and resolution, we constructed a spatial grid of 500m resolution covering the Amazon region and obtained the cell values for this raster following the steps described below. Similarly, we executed the process to achieve a monthly temporal resolution for all variables, with the data period from 2001 to 2020.

To achieve temporal harmonization across these differing frequencies, we adopted a temporal expansion framework where annual and static values are mapped consistently across the corresponding twelve monthly increments of each year. This ensures that every monthly snapshot in the 240-month time series contains a complete suite of environmental covariates, enabling dynamic analyses such as fire risk modeling or temporal trend assessment. This approach maintains temporal sensitivity by allowing dynamic climate variables to fluctuate monthly while the slower-evolving landscape attributes such as Land Cover provide a stable structural context for each year.

Although the collected data packages for different attributes have heterogeneous specifications, their processing generally follows a common workflow. A methodological baseline of the processing steps is shown in Figure 3. Specifically, Accessed Data refers to the downloaded data package from data sources in various formats, such as HDF, HDF5, NETCDF, GEOTIFF, and TIF. Accessed data in source data formats, such as HDF, HDF5, and NETCDF, contained several layers with different attributes, and the layer related to the subject attribute was extracted from this set of layers. Accessed data with the source data formats of GEOTIFF or TIF contained only the required layer that was extracted. These extracted layers are referred to as the raw data.

To resolve the inherent inconsistencies in spatial orientation and resolution, we developed a standardized processing framework. First, all raw data layers are projected onto EPSG:102033-South America Albers Equal Area Conic, to acquire the Projected Layer. This equal-area coordinate reference system was specifically chosen to maintain geometric fidelity across the vast longitudinal expanse of the Amazon basin, ¹⁵ minimizing the distortion of area and shape that typically occurs in global projections. To correct for grid misalignments and boundary distortions caused by the differing source orientations of the datasets collected, we introduced a standardized spatial grid as a master template. By mapping all attributes onto this fixed spatial grid, we ensured that every pixel across all variables represents the exact same geographical footprint, thereby eliminating spatial offsets and ensuring seamless interoperability between datasets.

The projected layers are at either a global or regional-level (based on the specifications of the data source), and to confine them all to the Amazon Basin boundary, these layers were further clipped using a shapefile-based (vector) Amazon Basin boundary. This clipped layer is labelled as pre-processed data.

Although all layers are cropped to the Amazon basin boundary, their respective cells may not exactly align with each other owing to differences in their source data extent, cell-grid orientation, and spatial resolution. To obtain layers of the same spatial extent, resolution, and orientation, we executed a rigorous two-step disaggregation and resampling procedure to transfer cell values from the pre-processed data to the fixed spatial grid (master template). The spatial grid covered the entire Amazon Basin boundary and had a cell resolution of 500 m. The value of each cell was transferred to this grid for each attribute, and the process was repeated for all attributes, thereby creating a separate spatial grid for each attribute.

First, layers with resolution coarser than 500 m (ranging up to 37 km) were disaggregated to approximately 500 m. The disaggregation factor varied for each attribute, based on the spatial resolution of the source data. Following this, the terra:resample function in R was used to transfer information from the attribute layers to the fixed spatial grid. To minimize spatial inaccuracy in this environmentally heterogeneous region, the resampling method was tailored to the variable type: the ‘near’ (nearest neighbor) method was employed for Land Cover, Burnt Area, Soil Moisture, Specific Humidity, Evapotranspiration, Near Surface Wind Speed, and Near Surface Air Temperature to preserve original discrete values and categorical integrity. Conversely, ‘bilinear’ interpolation was used for Land Surface Temperature and Precipitation to accurately represent the continuous spatial gradients of these atmospheric phenomena.

The resulting spatial grids corresponding to each attribute constitute the final layers available for analysis, hence called working data. This workflow was followed for each monthly file (i.e., 240 files over 20 years) to achieve a consistent monthly temporal resolution from 2001 to 2020. By maintaining this monthly sensitivity for dynamic climate variables while holding slower-evolving landscape variables such as Land Cover constant within each annual cycle, the dataset remains sensitive to the immediate environmental drivers for fire while providing a stable structural context. While an analytical sensitivity assessment regarding the impact of integrated variable frequencies is a valuable research direction, such an analysis is beyond the scope of this data curation work.

Figure 4 illustrates an example of Land Surface Temperature in January 2020 for all three categories of raw data, pre-processed data and working data. Similarly, Figure 5 presents an example of a single monthly instance from January 2020 for all the variables collected.

In terms of implementation, pre-processing work was completed using GIS software, and the processing work was executed in the statistical computing software R. ^{16, 17} All data layers were managed using the SpatRaster data structure in the terra package, ¹⁸ ensuring a transparent, scripted workflow. This algorithmic approach serves as a systematic process log, minimizing accumulated errors across the 240 temporal layers and ensuring the reproducibility of the dataset.

The raster-based dataset of covariates presented in this study is a collection of established datasets that do not include any newly created data records. This work mainly focuses on exhaustive data search and its acquisition process, followed by computer-intensive pre-processing to develop a dataset for the Amazon region. As noted in the Data Collection section, these industry-standard source datasets have undergone rigorous independent validation, as documented in their respective technical documentation; therefore, a secondary validation against field observations is beyond the scope of this curation effort. To ensure process transparency, the dataset follows standardized naming conventions.

While the primary contribution of this work lies in workflow standardization and the resolution of dataset fragmentation, this technical framework serves as the essential infrastructure required for future algorithmic advancements. By delivering a harmonized, analysis-ready foundation, this study enables high-level environmental research and predictive modeling, ¹¹ that were previously hindered by spatial and temporal data incompatibility. This standardized curation ensures that the resulting working data is fit-for-purpose for complex spatiotemporal analyses and provides a reproducible baseline for the wider research community.