Reclassification rules
This section provides a high-level introduction to reclassification rules used in the LUC-Brasil project. This stage implements 27 specialized reclassification rules to refine and standardize land cover classification. Temporal trajectory rules analyze multi-year patterns to detect transitions, enforce consistency, and correct isolated pixel misclassifications. Each rule addresses specific classification challenges such as pasture-agriculture confusion, urban expansion detection, or temporal consistency in perennial crops.
The heart of the system is the reclassification and remapping pipeline, which transforms diverse input classifications into a standardized output schema. The package implements 27 specialized rules organized into the following categories:
- Forest Rules: Uses PRODES to distinguish between Primary Forests and Secondary Vegetation;
- Agriculture Rules: Uses TerraClass to improve the agricultural classes, especially in the case of Perennial and Semi-Perennial Agriculture;
- Infrastructure Rules: Includes urban and mining areas from TerraClass;
- Water Rules: Includes water mask from TerraClass;
- Multi-year transitions: Consider three consecutive years to ensure consistency of more difficult classes (e.g., conflicts between Shrubby Pasture and Secondary Vegetation).
The rules are described below, in the order on which they are applied.
Base Mask Preparation
Base Mask Preparation prepares the reference masks from two authoritative external datasets: PRODES deforestation monitoring data; TerraClass land use and land cover maps, and water masks. These base masks serve as reference for all downstream processing stages.
The base mask preparation includes geometrically adjusting the PRODES deforestation mask to remove some inconsistencies in the border pixels, which distort the overall statistics.
Reclassification Part 1: Yearly Rules Defined by Masks
Secondary Vegetation Rule
This rule distinguishes forested areas in Amazonia between primary forests and secondary vegetation. To that aim, it uses the PRODES mask, which indicates the primary forest area. The secondary vegetation areas are those that have been deforested and then abandoned, so that a forest regeneration process has started.
Current Deforestation Rule
For each year, the PRODES mask also contains information on forest removal under the current rule. This rule ensures that all areas deforested in a given year are not marked as “Pasture” or “Agriculture”, since TerraClass also follows the same convention of not classifying land use for areas deforested in the year it is produced.
Residuals Rule
This rule is necessary due to problems in the PRODES mask. Since PRODES only observes a few images per year for a given Landsat tile, some areas are not observed due to clouds and are masked as such. When such areas are visible in later years and then assigned as “residuals”. Since LUC-Brasil uses time series which enable removal of most cloudy areas, this rule assigns PRODES residuals to the LUC-Brasil classes.
Silviculture Rule #1
This rule reclassifies the LUC-Brasil pixels matching “Silviculture” in TerraClass. The pixels already classified as “Silviculture” in LUC-Brasil are kept unchanged. The respective years used are defined in Appendix A.
Silviculture Rule #2
This rule reclassifies as “Pasture” the LUC-Brasil pixels which are classified as “Silviculture” in LUC-Brasil and as non “Silviculture” in TerraClass. The respective years used are defined in Appendix A.
Semi-Perennial Agriculture Rule #1
This rule reclassifies the LUC-Brasil pixels matching “Semi-Perennial Agriculture” in TerraClass. The pixels already classified as “Semi-Perennial Agriculture” in LUC-Brasil are kept unchanged. The respective years used are defined in Appendix A.
Semi-Perennial Agriculture Rule #2
This rule reclassifies as “Pasture” the LUC-Brasil pixels which are classified as “Semi-Perennial Agriculture” in LUC-Brasil and do not match any temporary agriculture in TerraClass. The respective years used are defined in Appendix A.
Semi-Perennial Agriculture Rule #3
This rule reclassifies as “Pasture” the LUC-Brasil pixels which are classified as “Semi-Perennial Agriculture” in LUC-Brasil and as non “Semi-Perennial Agriculture” in TerraClass. The respective years used are defined in Appendix A.
Annual Agriculture Rule
This rule reclassifies the LUC-Brasil pixels matching temporary agriculture in TerraClass. The pixels already classified as “Annual Agriculture” in LUC-Brasil are kept unchanged. The respective years used are defined in Appendix A.
Perennial Agriculture Rule
This rule reclassifies as “Perene” the LUC-Brasil pixels matching perennial agriculture in TerraClass. In years with no TerraClass, we use the closest TerraClass year (Appendix A). Then, we reclassify as “Perene” the LUC-Brasil pixels matching perennial agriculture in TerraClass and “Secondary vegetation” or “Shrubby Pasture” in LUC-Brasil. For years prior to 2008, before TerraClass started mapping the Perennial Crop class, the 2008 TerraClass map was used as the reference for a retrospective adjustment. Pixels classified as Perennial crop in the 2008 TerraClass map and as Secondary Vegetation or Shrubby Pasture in the corresponding LUC-Brasil classification of the preceding year are reclassified as Perennial crop. This procedure is then applied interactively to earlier years so that, as long as a pixel remains classified as Secondary Vegetation or Shrubby Pasture, it is reclassified as Perennial Crop, ensuring temporal consistency throughout the time series.
Mining Rule
This rule reclassifies the LUC-Brasil pixels matching “Mining” in TerraClass. The respective years used are defined in Appendix A.
Urban Areas Rule
This rule reclassifies the LUC-Brasil pixels matching “Urban Areas” in TerraClass. To ensure temporal and spatial consistency throughout the time series, LUC-Brasil data were adjusted so that pixels once classified as urbanized areas remain in this class in subsequent years. In addition, the extent of urbanized areas in a given year cannot exceed the extent mapped in the following year, with the maximum extent corresponding to the 2024 map. For intermediate years without a corresponding TerraClass product, the adjusted data from the previous year, after applying the rules and corrections, were used as a mask. The respective years used as input are defined in Appendix A.
Water Bodies Rule #1
This rule reclassifies as “Water” the LUC-Brasil pixels which are not classified as “Wetlands” or “Seasonally Flooded” in LUC-Brasil and classified as “Water Bodies” in TerraClass. Since the “Water Bodies” class is considered static in TerraClass, for years without a corresponding TerraClass product the same water extent from the most recent available TerraClass reference year is assumed. The respective years used are defined in Appendix A.
Water Bodies Rule #2
The GLAD data does not have permanent water pixels (e.g., lakes, rivers). To include the “Water” class in years using this dataset, in addition to the Water Bodies Rule #1, we built a static water mask. In this mask, we consider “Water” the NA pixels in GLAD data that never change in the entire series.
Natural Non-Forest Areas Rule
This rule reclassifies the LUC-Brasil pixels matching “Natural Non-Forest” in PRODES. The respective years used are defined in Appendix A.
Reclassification Part 2: Year-Specific Mask Generation
Water Consistency Rule
This rule reclassifies the LUC-Brasil pixels which are classified as “Water” in 2019, but as non “Water” in 2018 or 2020. In the reclassification of 2019, we adopt the class value of 2018. For example:
- 2018 - Pasture
- 2019 - Water
- 2020 - Pasture
In this case, 2019 will be reclassified as “Pasture”.
Perennial Agriculture Trajectory
This rule validates perennial crop classifications by consulting TerraClass reference data from surrounding years (year-1 and year+1). Perennial crop pixels are only retained if TerraClass confirms their presence in adjacent years, reducing false positives.
The rule targets specific years where TerraClass surveys are available:
| Target Year | TerraClass Before (year-1) | TerraClass After (year+1) | Rationale |
|---|---|---|---|
| 2009 | 2008 | 2010 | Between consecutive TerraClass surveys |
| 2011 | 2010 | 2012 | Between consecutive surveys |
| 2013 | 2012 | 2014 | Between consecutive surveys |
| 2019 | 2018 | 2020 | Between consecutive surveys |
| 2021 | 2020 | 2022 | Between consecutive surveys |
| 2023 | 2022 | 2024 | Between consecutive surveys |
Secondary Vegetation - Pasture Consistency
This rule analyzes secondary vegetation to pasture transitions by examining 3-year trajectories. This rule identifies pixels classified as secondary vegetation that are surrounded by pasture classifications in adjacent years, helping distinguish true secondary vegetation from misclassified pasture areas.
The rule processes 21 consecutive years from 2001 to 2025, creating overlapping 3-year windows. For each target year, the algorithm examines year-1, year, and year+1 classifications to detect SV-pasture patterns.
Temporal Water Consistency
This rules ensures temporal consistency for the water areas. It is applied to the period 2015-2022.
Cropland/Pasture Consistency
This rule uses a three-year period to adjust misclassification scenarios and unlikely changes between cropland and pasture. For example, pixels labelled cropland - pasture - cropland are reclassified as cropland - cropland - cropland. The three-year periods are those between TerraClass surveys:
| Target Year | TerraClass Before (year-1) | TerraClass After (year+1) |
|---|---|---|
| 2014 | 2015 | 2016 |
| 2016 | 2017 | 2018 |
| 2018 | 2019 | 2020 |
| 2020 | 2021 | 2022 |