đ Reference · For users who want to look up how a deforestation check result was produced, by methodology version.
A deforestation check assesses whether forest on a production plot was lost after the cut-off date of 31 December 2020 set by the EU Deforestation Regulation (EUDR). This article describes each version of the methodology Prewave uses for that assessment.
Why the methodology is versioned
Every deforestation check result carries the version of the methodology that produced it. Versioning keeps results traceable: for any result, you can look up which data and which decision rules were used, and reproduce how the result came about.
Prewave creates a new version when the data analysis or the decision rules change substantially, or when the underlying data sources change substantially. Prewave decides when a change is substantial enough for a new version. Each new version is added to this article, and earlier versions stay documented here.
Version | Period of application | What defines it |
Version 2 | October 2026 to present | One Satelligence check per origin, with a result for every plot, on an extended set of data sources |
Version 1 | May 2024 to October 2026 | A basic check on open data, followed by an advanced Satelligence check where the basic check returns Non-negligible, with one result per origin |
Version 1 results are shown per origin. Version 2 results show every plot with its own result.
Version 2
Period of application
October 2026 to present.
Changes from Version 1
One check instead of two. Every check runs on the Satelligence service. There is no basic check.
The result is decided per plot instead of per origin.
Only forest loss within the in-scope timeframe of the production counts (see below).
The data sources behind the Satelligence forest baseline were extended. See Version 2 data sources.
Checks and decision rules
Each origin receives one check, powered by Satelligence. The check produces a result for every plot of the origin.
A plot is Non-negligible when Satelligence detects forest loss on it within the in-scope timeframe, and Negligible when it does not.
The in-scope timeframe runs from January 2021 up to and including the month the production ended: the month of a single production date, or the last month of a production period. Forest loss detected after that month is out of scope. For example, a detection in February 2026 does not count for an origin produced in January 2026.
Primary forest on a plot does not make the plot Non-negligible on its own. Only detected forest loss does.
An origin takes the result of its weakest plot. A product and a Product Origin Request each take the result of their weakest origin.
Data analysis
The data analysis is the same as in the Version 1 advanced check: the same hybrid forest baseline approach, the same use of commodity maps to remove false positives, and the same change detection algorithms (SpatioTemporalAdaptiveBareness for optical imagery and Bayesian Iterative Updating for radar imagery). The full description is under Version 1, Data analysis, Advanced check below.
Changes in the data analysis compared with Version 1:
The baseline, now called the Harmonized Nature and Commodity Baseline, also maps non-forest natural land: shrubland, grassland and wetland.
In areas with a lower risk of deforestation, change is detected with open alert datasets: Global Forest Watch Tree Cover Loss and the Global Forest Watch Integrated Deforestation Alerts, which combine the Global Land Analysis and Discovery alerts from Landsat and Sentinel-2 (GLAD-L, GLAD-S2) and the Radar for Detecting Deforestation alerts (RADD). A change counts only where it overlaps forest in the baseline.
Data sources
Version 1
Period of application
May 2024 to October 2026.
Changes from the previous version
None. Version 1 is the first version of the methodology.
Checks and decision rules
Version 1 uses two checks in sequence.
Every origin first receives a basic check.
If the basic check returns Negligible, that is the result for the origin.
If the basic check returns Non-negligible, an advanced check powered by Satelligence can be run for the product. The advanced result then replaces the basic result for every origin it covers. The basic result stays in the history but no longer determines the status.
There is no tolerance threshold: an origin is Non-negligible as soon as forest loss after the cut-off date is detected on it, and Negligible when none is detected. The result is decided per origin. A product takes the result of its weakest origin.
Data analysis
Basic check
For the basic check, Prewave overlays the Joint Research Centre (JRC) primary forest map layer with the Hansen forest loss dataset. A plot is Non-negligible only when both conditions are met: the location is classified as primary forest in the JRC dataset, and that forest has been lost according to the Hansen forest loss dataset.
Advanced check
The advanced check combines a forest baseline with change detection algorithms to identify forest loss.
The forest baseline uses a hybrid approach. It merges curated and harmonised open datasets, such as the JRC Tropical Moist Forest layers and the University of Maryland (UMD) primary forest maps, with Satelligence's own forest classification models. The models apply the forest definition of the Food and Agriculture Organization of the United Nations (FAO) and national forest definitions, and sort forest into primary forest, disturbed forest, regrowth, dry forest and, in South America, other native vegetation. These are Satelligence's own classes, not the forest categories defined in the EUDR. Open layers are quality-checked, standardised, and corrected to remove areas deforested before the cut-off date.
Satelligence adds its own commodity maps of the EUDR commodities and of crops such as sugar cane and coconut, created from Landsat, Sentinel-1 and Sentinel-2 imagery. Areas mapped as a commodity are removed from the forest layer, which prevents crops from being counted as forest in farming landscapes.
Change is detected with two algorithms:
SpatioTemporalAdaptiveBareness (STAB) works on optical imagery from Landsat and Sentinel-2. It compares the bare-soil level of each pixel with the statistics of the surrounding forest. A pixel with much more bare soil than the forest around it is flagged as deforested. The threshold adapts to the landscape, so it can account for seasons in dry tropical forests and savannas such as the Cerrado and the Chaco.
Bayesian Iterative Updating works on radar imagery from Sentinel-1. It compares current observations with historical observations of the same place to calculate the probability of change.