At Satelligence, we are committed to a zero-deforestation future, and we support companies in their journey towards carbon neutrality. We do so by being able to process thousands of satellite images at scale, in order to provide near real-time insights about illegal deforestation in their supply chains.
Equally important to our change detection algorithm’s performance and scalability, is the accuracy of our forest layer. This enables us to clearly distinguish between plantation clear cutting and illegal agricultural expansion into forest environments, which companies need to be aware of in order to comply with international regulations and legislations, such as the EUDR.
In this article, we report on the accuracy of our forest layer, and describe the methodology that we used for the assessment.
Key insights:
The Satelligence Forest Baseline Layer (FBL) distinguishes between several forest types, namely primary forest, disturbed moist forest, regrowth forest, dry and seasonal forest, and native vegetation. The FBL is composed of data processed in-house, as well as several third-party layers from different providers, covering different countries. The data is stacked together at high spatial resolution (10m) for various years, including 2020 (relevant for the EUDR), which is used for this assessment.
In this exercise, we focus on estimating the Overall Accuracy (OA) of the binary classification forest vs non-forest, without looking at forest classes individually. With the aim of achieving less than 2% error on the OA score, we used the stratified random sampling procedure described in Olofsson et al. (2014) to estimate the number of samples needed per class.
In an ideal scenario, we would obtain independently generated labels and use them to validate the FBL. However, such a dataset does not exist, unless one were to use third-party forest maps (typically also based on remote sensing data) as validation data. Therefore we selected samples from our FBL, and engaged the labelling company Enlabeler to provide hundreds of labels for several countries, using very high resolution (VHR) open layers such as Google Maps and Bing Maps for reference, as well as Planet NICFI monthly data. While the latter does not have fine enough spatial resolution to distinguish forest from plantations, it offers the opportunity to check for (multi-)yearly replanting of a given commodity. Furthermore, we also exploited the Timeline feature of Google Earth Pro when needed, e.g. for selected samples where the VHR layers disagreed.
Labelling samples requires considerable expertise, especially in areas with high density of smallholders’ farms and with the presence of shading trees next to forest patches. Furthermore, contextual knowledge is very important; while the footprint to label is only 30 m in diameter, one has to also gauge whether that footprint is part of a forested landscape, as opposed to an agricultural or urban environment, whether the forest patch is over a determined spatial size and height, etc. Acknowledging that the assessment of these samples is such a challenging task, we additionally engaged remote sensing experts within our team to look at a subset of the samples’ datasets. Whenever there was disagreement between the individuals, a majority vote was used to determine the final sample class.
A total of 462 samples were selected from our baseline, of which 48 samples (10.4 %) are forest. The non-forest samples are split across all relevant EUDR commodities, but were grouped together, as we are specifically assessing the accuracy of the forest / non-forest mapping.
Comparing the FBL to the labels, we achieve a 93 % OA. For comparison, we validated three other datasets against the same labels, namely the JRC-EUDR-V2, JRC-TMF, and Côte d’Ivoire’s national land cover map. The OA of the two JRC datasets is 77% and 76 %, respectively, while Côte d’Ivoire’s own national landcover map achieves a score of 78 %.
These results are in agreement with CIAT and WCF (2025), an independent research paper that evaluated the accuracy of a broad selection of datasets, including the ones used in this assessment, over both Côte d’Ivoire and Ghana. According to their analysis, Satelligence achieves the highest precision (87%) over forest in Côte d’Ivoire, followed by the national land cover dataset (73%). Note that while CIAT and WCF (2025) report on “precision”, we report on “overall accuracy”. These metrics are not identical, but they are related to each other.
The differences are likely due to Côte d’Ivoire’s complex agricultural landscape, where all EUDR commodities are grown, from relatively large oil palm plantations to a high number of smallholder cocoa and rubber plots. Our commodity maps allow us to distinguish between forest and plantations and likely are driving the higher score in the overall accuracy of our forest baseline.
| Overall Accuracy % | |
| Satelligence forest baseline | 93.3 |
| JRC EUDR v2 | 77.1 |
| JRC TMF | 75.5 |
| Côte d'Ivoire national landcover | 77.5 |
Comparison of forest maps, Côte d'Ivoire
A total of 409 samples were selected from our baseline, of which 126 samples (31 %) are forest. As for Côte d’Ivoire, the non-forest samples are split across all relevant EUDR commodities, but were grouped together to assess the accuracy of the forest / non-forest mapping. Note that while Brazil’s land area (8.5 M km2) is about 26 times larger than that of Côte d’Ivoire (322,000 km2), the number of samples is comparable (462 vs 409). This is because the stratified random sampling is based on the fractional areas of the different land classes, and the samples’ number itself is mostly driven by the uncertainty that we want to achieve on the overall accuracy (2%). However, since forest covers a higher fraction of the country’s land mass in Brazil, more samples were taken for this land class compared to Côte d’Ivoire (126 vs 48).
When compared to the labels, we achieve 87.5 % OA, 8.5 percentage points higher than the JRC EUDR v2 global layer (79 % OA). This is half of the 16.2 percentage points difference between these two datasets in Côte d’Ivoire. Indeed the two forest maps show higher similarity over Brazil. However, large differences are still visible in specific areas, especially in the central and southern parts (see inset in the figures), where crops are likely mistaken for forest.
| Overall Accuracy % | |
| Satelligence forest baseline | 87.5 |
| JRC EUDR v2 | 79.0 |
Comparison of forest maps, Brazil
A total of 423 samples were selected from our baseline, of which 104 samples (24 %) are forest. Again, non-forest samples split across all relevant EUDR commodities were grouped together to assess the accuracy of the forest / non-forest mapping. As in Brazil, forests cover a large portion of the country’s land area in Indonesia. Therefore the forest samples cover a high proportion of the samples’ dataset (24%), comparable to Brazil (31%) and much higher than in Côte d’Ivoire (10.4%), despite the total number of samples being similar in all three countries.
Over Indonesia, our forest classification scores 86 % OA. While still higher than the OA of the JRC EUDR v2 classification (80.1 %), the difference between the two layers in Indonesia is less pronounced than in the previous two countries (5.9 points).
| Overall Accuracy % | |
| Satelligence forest baseline | 86.0 |
| JRC EUDR v2 | 80.1 |
Comparison of forest maps, Indonesia
In a regulatory landscape that increasingly demands traceability, transparency, and accountability, the quality and credibility of data behind deforestation(-free) claims are evermore important. As is highlighted in CIAT’s 2025 report, Satelligence provides forest baseline and carbon dataset that combine highly accurate satellite imagery with local, on-the-ground data. This offers a reliable foundation for EUDR compliance, Scope 3 reporting, and beyond.
In its own assessment of the CIAT 2025 report, the WCF also acknowledged that developing standardised frameworks and methods will ensure stronger and more efficient collaboration across commodity industries. Being in clear alignment with Satelligence goals, our team strives to contribute to what the WCF refers to as a “shared language around reporting” with our clear, highly accurate, and coherent data insights.
Report produced by Satelligence Data Team, April 2025
CIAT and WCF, Establishing best practices and evaluating data and methods for forest monitoring, 2025
Global / regional layers:
National layers: