-
Created by
Michela Giusti on May 15, 2026
25 minute read
Contributors: G. Schwaizer (ENVEO IT GmbH), M. Heinrich (ENVEO IT GmbH), P. Malcher (ENVEO IT GmbH), U. Fasching (ENVEO IT GmbH), T. Nagler (ENVEO IT GmbH)
Issued by: ENVEO IT GmbH / Gabriele Schwaizer
Date:
Ref: C3S2_WP3-DDP-SCE-01_202506_PQAR_v1.0
Official reference number service contract: 2024/C3S2_313d_ENVEO/SC1
History of modifications
List of datasets covered by this document
Acronyms
General definitions
Breakthrough (B): An intermediate level between threshold and goal which, if achieved, would result in a significant improvement for the targeted application. The breakthrough value may also indicate the level at which specified uses within climate monitoring become possible. It may be appropriate to have different breakthrough values for different uses.
Climate Data Record (CDR): Defines a time series of measurements of sufficient length, consistency, and continuity to determine climate variability and change.
Commission Error: In the context of classification, a commission error (or false positive) occurs when a data point is incorrectly assigned to a particular class despite not belonging to it. This type of error reflects an overestimation by the classifier, whereby it identifies the presence of a feature or class that is, in reality, absent.
Essential Climate Variable (ECV): a physical, chemical or biological variable or a group of linked variables that critically contributes to the characterization of Earth's climate.
Goal (G): An ideal requirement above which further improvements are not necessary.
Interim Climate Data Record (ICDR): Defines a dataset that has been forward processed, using the baselined Climate Data Record algorithm and processing environment but whose consistency and continuity have not been verified. Eventually, it will be necessary to perform a new reprocessing of the CDR and ICDR parts together to guarantee consistency, and the new reprocessed data record will replace the old CDR.
Omission error: In the context of classification, an ommission error (also known as a false negative) occurs when a data point that truly belongs to a given class is incorrectly excluded from that class by the classifier. In other words, the classifier fails to detect or identify the presence of a class or feature that is actually present, leading to an underestimation.
Snow Cover Extent (SCE): Areal extent of snow-covered land, which can be expressed as binary or as a fraction.
Snow Cover Fraction (SCF): Fraction of snow covered area per pixel, given in per cent.
Snow Cover Fraction on Ground (SCFG): Fraction of snow covered area per pixel, given in per cent. In forested areas, a correction is applied to account for the shading effect of the forest canopy, estimating the snow cover fraction on the ground underneath the canopy. In non-forested areas, Snow Cover Fraction on Ground is the same as Snow Cover Fraction Viewable from above.
Snow Cover Fraction Viewable from above (SCFV): Fraction of snow covered area per pixel, given in per cent. In forested areas, the snow cover fraction viewable from above, on top of the forest canopy, is provided. In non-forested areas, Snow Cover Fraction Viewable from above is the same as Snow Cover Fraction on Ground.
Threshold (T): The minimum requirement to be met to ensure that data are useful.
L1TP: Landsat Processing Level Terrain Precision Correction: Radiometrically calibrated and orthorectified using ground control points (GCPs) and digital elevation model (DEM) data to correct for relief displacement.The highest quality Level-1 products suitable for pixel-level time series analysis. GCPs used for L1TP correction are derived from the Global Land Survey 2000 (GLS2000) data set.
Executive summary
This document presents the quality assessment results for global daily Snow Cover Extent (SCE) products derived from optical satellite observations. The evaluation focuses on daily SCE products generated from NOAA AVHRR and Sentinel-3 SLSTR data, using reference snow maps produced from high-resolution optical satellite imagery. The reference scenes are carefully selected to provide spatially and temporally representative coverage, ensuring a robust and comprehensive validation dataset. The quality assessment of Terra MODIS based SCE products is documented in the Product Validation and Intercomparison Report (PVIR) of ESA CCI Snow (Barella et al., 2024).
The AVHRR- and SLSTR-based SCE products are assessed through pixel-by-pixel comparisons with the corresponding reference snow maps. Performance metrics include the bias and the unbiased root mean square difference (ubRMSD), which together provide insights into the accuracy and precision of the products.
For the AVHRR-based SCE product, the validation yields a mean bias of -2.5%, indicating a slight underestimation, and an ubRMSD of approximately 18%. The Sentinel-3 SLSTR-based SCE product shows a mean bias of +2.7%, indicating a slight overestimation, with a lower ubRMSD of about 11%, reflecting improved consistency with the reference data.
Known limitations affecting the product performance are:
-
Geolocation inaccuracies, particularly for NOAA AVHRR data, though also observed for some Sentinel-3 SLSTR data.
-
Underestimation of fractional snow cover, due to the necessary pre-classification of snow-free areas aimed at minimizing false snow detection in other regions.
-
Challenges in distinguishing snow from cloud cover.
-
Underestimation of snow cover in shaded mountainous terrain and areas with low solar illumination.
- For Sentinel-3 SLSTR data, slight overestimations of snow cover fraction compared to the reference snow maps in areas with complex topography or variable snow density.
This document outlines the validation methodology, describes the reference datasets used, and summarizes the key validation results. Selected case examples are included to illustrate typical performance characteristics and known issues.
A dedicated section on climate change assessment will be included in a future update.
The compliance of the C3S SCE product with user requirements, particulary for climate analysis applications, is assessed by comparing the C3S SCE product specifications and quality to the 2025 update of the GCOS2-245 requirements for the ECV Snow quantity "Area covered by Snow".
1. Product validation methodology
1.1. Validation concept
The validation of the global Snow Cover Fraction (SCF) products is based on a pixel-by-pixel intercomparioson with reference snow maps generated from selected cloud free high resolution (HR) optical satellite data. The concept is based on the validation protocol developed and established in the ESA QA4EO Satellite Snow Product Intercomparison and Evaluation Exercise (SnowPEx) (https://snowpex.enveo.at). An adapted version of the validation procedure focusing on the validation of snow cover fraction products with reference snow maps from high resolution optical satellite data is presented in Figure 1.1.
Figure 1.1: Concept for validation of global Snow Cover Extent products, adapted from ESA QA4EO SnowPEx+.
1.2. Metrics used for the quality assessment
The pixel-by-pixel intercomparison between the global SCF product and the reference snow map results in a difference map expressed in percent, along with statistical metrics that quantify the quality of the SCF product.
The quality assessment employs the following metrics:
- Number of usable intercomparison pixels (Nui): the total number of pixels included in the evaluation, at the SCF product's grid spacing.
- bias: representing the systematic deviation of the product,
- unbiased Root Mean Square Difference (ubRMSD): a measure of the random error component, excluding systematic bias.
The bias quantifies the average offset between the SCF product and the reference, while the ubRMSD characterizes the magnitude of the unbiased differences.
The bias is calculated as the mean difference between the product and the reference in the same unit as the product.
\[ bias = \frac{1}{N_{ui}}\sum_{j=0}^{y} \sum_{i=0}^{x}\big(SCF(i,j) - SCF_{REF}(i,j)\big) \tag{Eq. 1.1} \]with
\[ \begin{align} SCF(i,j)&=\hbox{Product Snow Cover Fraction of the pixel i,j} \\ SCF_{REF}(i,j)&=\hbox{Reference Snow Cover Fraction of the pixel i,j} \\ N_{ui}&=\hbox{number of pixels}\\ \end{align} \]The ubRMSD quantifies the difference between the global SCF product and the reference snow map in the same unit as the product, so in per cent. The ubRMSD is calculated as:
\[ ubRMSD = \sqrt{\frac{1}{N_{ui}}\sum_{j=0}^{y} \sum_{i=0}^{x}\big(SCF(i,j) - \overline{SCF}) - (SCF_{REF}(i,j) - \overline{SCF_{REF}})\big)^2} \tag{Eq. 1.2} \]with
\[ \begin{align} SCF(i,j)&=\hbox{Product Snow Cover Fraction of the pixel i,j} \\ \overline{SCF}&=\hbox{Average Product Snow Cover Fraction} \\ SCF_{REF}(i,j)&=\hbox{Reference Snow Cover Fraction of the pixel i,j} \\ \overline{SCF_{REF}}&=\hbox{Average Reference Snow Cover Fraction} \\ N_{ui}&=\hbox{number of pixels}\\ \end{align} \]Since both the SCF product under evaluation and the reference snow map used for quality assessment may be subject to uncertainties, the term Root Mean Square Difference (RMSD) is used instead of Root Mean Square Error (RMSE). Although RMSD and RMSE are computed using the same mathematical formulation, RMSE implies that the reference represents the true value, which may not be the case here.
The ubRMSD captures the random error component, while the bias represents the systematic deviation between the SCF product and the reference map. Together, these two components can be used to define the total deviation, expressed as the Root Mean Square Difference (RMSD).
\[ RMSD^2 = ubRMSD^2 + bias^2 \tag{Eq. 1.3} \]1.3. Reference data
For the quality assessment of the SCF product, reference snow maps are generated from selected high-resolution optical satellite data acquired by the Landsat missions. To ensure a representative reference dataset, the selection of Landsat scenes considers both spatial and temporal coverage. The temporal distribution spans different seasons and years.
For the validation of AVHRR-based SCF products, 44 Landsat scenes acquired between 1985 and 1999 were selected. These scenes cover all months of the year and are spatially distributed across the globe (Figure 1.2). However, areas east of 50°E are underrepresented in the dataset due to various limitations, including cloud cover in either the SCF product or the Landsat imagery, absence of snow, missing satellite data or polar night in the AVHRR based SCF product, or geolocation mismatches between the AVHRR-based SCF product and the Landsat-derived snow maps.
| a) |
|
| b) |
|
| c) |
|
Figure 1.2: a) Spatial, b) monthly, and c) annual distribution of selected high resolution optical satellite data used for the generation of reference snow maps for the validation of AVHRR based SCE products.
For the evaluation of SLSTR-based SCF products, 25 Landsat scenes from the year 2023 are selected. These scenes are distributed globally and were acquired primarily during the main winter and melt seasons. The Southern Hemisphere is represented by only two scenes from South America, where both SLSTR- and Landsat-based snow cover maps indicate the presence of seasonal snow outside glaciated areas. An overview of the selected Landsat scenes is provided in Figure 1.3.
| a) |
|
| b) |
|
Figure 1.3: a) Spatial and b) monthly distribution of selected high resolution optical satellite data used for the generation of reference snow maps used for the validation of SLSTR based SCE products. All reference data are of the year 2023.
After downloading, the selected Landsat (L1TP) scenes are reprojected from their native UTM/WGS84 coordinate system to a geographic (latitude–longitude) map projection on the WGS84 ellipsoid, consistent with the coordinate reference system of the global SCF products. The data are resampled to a spatial resolution of 0.00025° × 0.00025°. At this resolution, SCF maps are derived from all selected Landsat scenes using the Locally Adaptive Multi-Spectral Unmixing (LAMSU) algorithm (Keuris et al., 2023). Pixels classified as water bodies, glacier, and cloud cover are masked.
The high-resolution SCF maps are subsequently aggregated to match the grid spacing of the SCF products under validation. Specifically, a running average method (Welford, 1962) is applied to generate SCF maps at 0.05° × 0.05° resolution for AVHRR-based products and at 0.01° × 0.01° for SLSTR-based products.
The resulting aggregated SCF maps, aligned to the resolution of the respective SCF product under validation, serve as the reference snow map for the quality assessment.
As the LAMSU algorithm used for the generation of the reference snow maps provides information on the snow cover fraction viewable from above, the focus of the validation activities is on non-forested or sparsely forested areas. Thus, all validation results are valid for the products snow cover fraction on ground (SCFG) and snow cover fraction viewable from above (SCFV).
2. Validation results
In this section, the quality assessment results for SCF products from AVHRR and SLSTR data are reported. The quality assessment of MODIS based SCF products is reported in the Product Validation and Intercomparison Report of ESA CCI Snow (Barella et al., 2024).
2.1. Quality assessment of AVHRR based SCE products
The quality assessment of the AVHRR-based SCF products, performed through a pixel-by-pixel intercomparison with 44 aggregated reference SCF maps (see Section 1.3), reveals a generally consistent performance with a slight overall underestimation of snow cover. Figure 2.1 presents the distribution of bias and unbiased RMSD values from all intercomparisons, while Table 2.1 summarizes the corresponding statistics, including the mean, median, minimum, and maximum values. The mean and median bias values were found to be -2.5% and -2.6%, respectively, indicating a small negative bias in the AVHRR-derived SCF estimates. However, individual bias values exhibit a broader range, spanning from approximately -18% to +17%, with the majority falling between -11% and +6%, suggesting localized variations in retrieval accuracy. The unbiased RMSD further supports these findings, with both the mean and median values around 18%. The unbiased RMSD values range from near 0%, indicating near-perfect agreement in some cases, up to about 34%, highlighting instances of significant deviation. These results underscore the importance of considering both bias and variability in the interpretation and application of AVHRR-based SCF products, particularly in climate monitoring and trend analysis.
Figure 2.1: Overall validation results for AVHRR based SCE products from evaluation with 44 reference snow maps.
Table 2.1: Statistical metrics resulting from the validation of AVHRR based SCF products with 44 reference snow maps from 1985 to 1999.
| Measure | Bias (%) | Unbiased RMSD (%) |
|---|---|---|
| Mean | -2.5 | 17.8 |
| Median | -2.6 | 17.5 |
| Minimum | -17.9 | 0.1 |
| Maximum | 16.8 | 34.4 |
To gain deeper insights into the overall quality assessment results of the AVHRR-based SCF product, four representative validation cases are examined in detail. Each case is selected from a different climatic region and features distinct surface and snow conditions. For each validation example, four visual components are presented: (a) the AVHRR-derived SCF subset, (b) the reference snow map derived from a Landsat scene acquired on the same date, (c) a false-colour composite of the original-resolution Landsat image, and (d) a difference map showing the pixel-wise SCF difference in percentage between the AVHRR and the reference snow map.
2.1.1. Case 1: The Alps, Europe – 25 February 1990
Reference snow map from Landsat scene: LT05_L1TP_192027_19900225_20200916_02_T1
This validation example represents a mountainous region in the European Alps in the main winter season. The AVHRR-based SCF extent shows good overall spatial alignment with the reference snow map, with substantial agreement in areas of high snow cover. However, cloud cover is significantly overestimated in the AVHRR based SCF product, resulting in a reduction of snow detection. In patchy snow areas, the AVHRR product underestimates SCF, and the product lacks a smooth transition between snow-free and fully snow-covered conditions. Instead, most pixels are classified as either completely snow-free or nearly 100% snow-covered, with very few pixels exhibiting low fractional values (Figure 2.2). Most of the snow free pixels are resulting from the pre-classification based on the Normalized Difference Snow Index and the Brightness Temperature centered at about 11 µm. Details are provided in the ATBD.
Figure 2.2: Validation example for the Alps, Europe, on 25 February 1990: a) Subset of the global SCF product derived from AVHRR data, b) reference snow map from Landsat 5 TM scene (path 192, row 027), c) false-colour composite of the Landsat scene at original grid spacing, d) difference map showing AVHRR-based SCF minus reference snow map in per cent.
2.1.2. Case 2: The Andes, South America – 20 October 1992
Reference snow map from Landsat scene : LT05_L1TP_232084_19921020_20200914_02_T1
In this high-altitude example from the Andes in the melt season, the AVHRR-based SCF product shows again an overestimation of cloud cover and an underestimation of snow extent, particularly in heterogeneous terrain with patchy snow distribution. While there is good agreement in pixels with high SCF values, transitional snow zones are not well captured. The SCF retrieval tends to shift abruptly between 0% and nearly 100%, failing to reflect the fractional nature of snow cover in the transition from snow covered to snow free areas (Figure 2.3). Similar as for the example in the Alps, most of the snow free classifications result from the pre-classification step. An overall tendency of understimated SCF from AVHRR is observed in the Andes.
Figure 2.3: Validation example for the Andes, South America, on 20 October 1992: a) Subset of the global SCF product derived from AVHRR data, b) reference snow map from Landsat 5 TM scene (path 232, row 084), c) false-colour composite of the Landsat scene at original grid spacing, d) difference map showing AVHRR-based SCF minus reference snow map in per cent.
2.1.3. Case 3: Maloti Mountains, Lesotho / South Africa – 18 July 1996
Reference snow map from Landsat scene : LT05_L1TP_169080_19960718_20200911_02_T1
This case study from Southern Africa in the main winter season reveals strong performance in identifying areas with extensive snow cover. In regions where snow is continuous, the AVHRR-derived SCF closely matches the reference. Cloud detection also performs well, with minimal false cloud classification. However, in areas of fragmented or patchy snow, the retrieval continues to struggle, exhibiting an unrealistic step change from no snow to full snow coverage, with few intermediate SCF values detected (Figure 2.4).
Figure 2.4: Validation example for the Maloti mountains, Lesotho / South Africa, on 18 July 1996: a) Subset of the global SCF product derived from AVHRR data, b) reference snow map from Landsat 5 TM scene (path 169, row 080), c) false-colour composite of the Landsat scene at original grid spacing, d) difference map showing AVHRR-based SCF minus reference snow map in per cent.
2.1.4. Case 4: Alaska, North America – 23 May 1998
Reference snow map from Landsat scene : LT05_L1TP_046013_19980523_20200909_02_T1
The Alaskan case acquired in the melt season shows high performance, with good spatial correspondence between the AVHRR-based SCF and the reference map, including a reasonable depiction of areas with both high and low snow cover. The SCF product demonstrates sensitivity to fractional cover, capturing a broader range of snow cover percentages. A small spatial misalignment of approximately half a pixel is observed, which may introduce some localized discrepancies. Cloud detection errors are minimal, and the SCF values generally reflect realistic gradients in snow distribution. Notably, some overestimation occurs in areas where the actual snow cover is high but not complete, for instance, pixels with ~90% snow cover are sometimes classified as 100%, suggesting a too rapid transition to full coverage.
Figure 2.5: Validation example for Alaska, North America, on 23 May 1998: a) Subset of the global SCF product derived from AVHRR data, b) reference snow map from Landsat 5 TM scene (path 046, row 013), c) false-colour composite of the Landsat scene at original grid spacing, d) difference map showing AVHRR-based SCF minus reference snow map in per cent.
2.1.5. Summary
The four regional case studies highlight both strengths and limitations of the AVHRR-based SCF product. Across all regions, areas with extensive snow cover are generally well identified, and cloud detection is effective in most cases. However, persistent challenges are observed in capturing fractional snow cover, particularly in patchy snow environments. While the SCF retrieval approach tends to simplify snow cover patterns in heterogeneous landscapes, particularly by limiting the representation of gradual transitions between snow-free and snow-covered areas, it offers a robust classification in more homogeneous conditions. The observed underrepresentation of snow in complex terrain is partly a result of necessary pre-classification constraints, which are needed to reduce misclassification in other regions. Overall, the product performs reliably in areas with dense snow cover, making it a valuable resource for large-scale snow monitoring and climate applications.
2.2. Quality assessment of MODIS based SCE products
The MODIS based SCE products are evaluated within the ESA CCI Snow project. In open land, the validation of MODIS-based SCF products from the Climate Research Data Package (CRDP) v3.0 with reference snow maps from high resolution optical satellite data resulted in a mean bias of -2.5%, and a mean unbiased RMSD of 11%.
A detailed documentation of the product validation and intercomparison results performed in the ESA CCI Snow project for the Climate Research Data Package (CRDP) v3.0 is provided in the Product Validation and Intercomparison Report (PVIR, Barella et al., 2024).
2.3. Quality assessment of SLSTR based SCE products
The validation results of SLSTR based SCF products with reference snow maps from Landsat data presented in Figure 2.6 and Table 2.2 demonstrate a strong overall agreement. The mean bias of 2.7% and median bias of 2.1% indicate only a slight positive deviation, reflecting a generally accurate estimation of SCF. Nevertheless, there are cases with a slight negative bias indicating understimation of snow cover, with an absolute minimum bias value of -3.6%. Additionally, the mean and median unbiased RMSD values of 10.6% highlight a robust consistency in pixel-level agreement across the dataset. Notably, a minimum unbiased RMSD of 2.7% illustrate that certain regions exhibit very close correspondence to the reference, showcasing the product’s capability to capture snow cover with high precision. While maximum bias and unbiased RMSD values of 13.8% and 22.2%, respectively, suggest localized challenges, these remain within acceptable ranges for large-scale snow cover mapping. Overall, these validation metrics affirm the reliability and accuracy of the product for climate applications.
Figure 2.6: Overall validation results for SLSTR based SCE products from evaluation with 25 reference snow maps.
Table 2.2: Statistical metrics resulting from the validation of SLSTR based SCF products with 25 reference snow maps of 2023.
| Measure | Bias | Unbiased RMSD |
|---|---|---|
| Mean | 2.7 | 10.6 |
| Median | 2.1 | 10.6 |
| Minimum | -3.6 | 2.7 |
| Maximum | 13.8 | 22.2 |
To better understand the spatial behavior and limitations of the SCF product derived from Sentinel-3 SLSTR data, four representative case studies are presented in more detail. Each case captures different terrain, illumination, and snow distribution conditions. Comparisons were made against high-resolution reference snow maps generated from Landsat 8 or 9 using the LAMSU algorithm (see Section 1.3).
2.3.1. Case 1: The Andes, South America – 25 April 2023
Reference snow map from Landsat scene: LC09_L1TP_232082_20230425_20230426_02_T1
During the melt season in the high-altitude Andes, the SLSTR-based SCF product shows a strong ability to capture overall snow distribution, even in a region characterized by complex topography and challenging observational conditions. Comparison with the reference snow map reveals that the SLSTR-based SCF generally performs well in identifying broad snow patterns and continuous snow cover. In fully snow-covered areas, the product frequently reaches values close to 100%, highlighting its strong sensitivity and effectiveness in detecting extensive snow coverage.
Some localized differences occur. For example, in shaded areas—particularly around 69.05°W / 31.59°S—the SCF is underestimated due to reduced spectral reflectance in shadowed terrain, which limits accurate snow detection in optical satellite observations. Conversely, the slightly higher SCF values from SLSTR compared to the reference map in snow-covered areas likely reflect its coarser spatial resolution, which smooths subpixel variability better resolved by the reference map.
Despite these expected differences, the SLSTR-based SCF product demonstrates robust performance in detecting seasonal snow cover dynamics, offering a valuable resource for large-scale snow monitoring in mountainous regions. Its ability to effectively capture snow extent in complex terrain underscores its strong potential for operational use in climate studies.
Figure 2.7: Validation example for the Andes, South America, on 25 April 2023: a) Subset of the global SCF product derived from SLSTR data, b) reference snow map from Landsat 9 OLI-2 scene (path 232, row 082), c) false-colour composite of the Landsat scene at original grid spacing, d) difference map showing SLSTR-based SCF minus reference snow map in per cent.
2.3.2. Case 2: Xinjiang, China – 20 April 2023
Reference snow map from Landsat scene: LC08_L1TP_141030_20230420_20230429_02_T1
Validation of the SLSTR-based Snow Cover Fraction (SCF) product over the semi-arid mountainous terrain of Xinjiang during the melt season demonstrates its capacity to broadly capture snow cover patterns under challenging environmental conditions. The SLSTR data show good sensitivity in detecting snow presence, particularly in fully snow-covered regions. However, the product tends to overestimate cloud cover, especially in areas of continuous snow cover and at transition zones between snow-covered and snow-free surfaces. This likely results from conservative cloud masking and spectral confusion in mixed pixels.
A notable characteristic of the SLSTR-based SCF is its limited representation of lower fractional snow cover values along the snow line, caused by a somewhat conservative pre-classification of snow-free areas. Similar to observations in the Andes, snow-covered regions often exhibit slightly elevated SCF values, frequently plateauing at 100%. While this reflects strong detection capability in homogeneous snow-covered zones, it can obscure finer-scale variability captured by the reference snow map, especially in transitional areas.
Despite these limitations, the SLSTR-based SCF product remains effective in identifying large-scale snow cover patterns within the semi-arid climate region. Its consistent performance in detecting extensive snow coverage across diverse terrain types reinforces its value for climate studies.
Figure 2.8: Validation example for Xinjiang, China, on 20 April 2023: a) Subset of the global SCF product derived from SLSTR data, b) reference snow map from Landsat 8 OLI scene (path 141, row 030), c) false-colour composite of the Landsat scene at original grid spacing, d) difference map showing SLSTR-based SCF minus reference snow map in per cent.
2.3.3. Case 3: North America – 29 December 2023
Reference snow map from Landsat scene: LC08_L1TP_031033_20231229_20240108_02_T1
In the North American validation scene during the peak winter season, the SLSTR-based SCF product demonstrates strong detection of snow-covered areas, especially in regions with widespread snow presence. However, some challenges related to cloud masking arise, with cloud contamination observed in both fully snow-covered regions and transitional zones, particularly over bright surfaces. Some of these cloud classifications result from post-processing using temperature and precipitation data, which helps reduce SCF commission errors in other areas. In certain regions, the SLSTR-based SCF tends to report higher snow cover values compared to the reference snow map, which presents more moderate fractional values. Additionally, the SCF retrieval shows limited gradation between snow-free and snow-covered surfaces, often reaching 100% in snow-detected areas. This behavior indicates a more categorical approach to snow detection, which, while effective at identifying snow presence, may oversimplify subtle transitions during snow accumulation or melt. Despite these limitations, the SLSTR-based SCF performs reliably in consistently snow-covered environments.
Figure 2.9: Validation example for North America, on 29 December 2023: a) Subset of the global SCF product derived from SLSTR data, b) reference snow map from Landsat 8 OLI scene (path 031, row 033), c) false-colour composite of the Landsat scene at original grid spacing, d) difference map showing SLSTR-based SCF minus reference snow map in per cent.
2.3.4. Case 4: Russia and Azerbaijan – 3 January 2023
Reference snow map from Landsat scene: LC09_L1TP_168031_20230103_20230315_02_T1
In this winter scene from the mountainous region between Russia and Azerbaijan, the SLSTR-based SCF product exhibited patterns consistent with those observed in other complex terrains. Snow cover tended to be underestimated in shaded areas and regions with intricate illumination conditions, highlighting the challenges of accurately capturing snow presence in low-reflectance environments. Some areas are also pre-classified as snow-free in the SLSTR-based SCF product, whereas the reference snow map captures small snow patches within these regions. Conversely, in areas of continuous snow cover, the product often showed slightly higher SCF values compared to the reference snow map. This overestimation was most evident where the reference map detected high, but not maximal, snow fractions—likely due to its finer sensitivity to variations in topography and snow distribution. Despite these differences, the SLSTR-based SCF demonstrated a strong ability to detect snow presence across diverse terrain, contributing valuable insights also into regional snow patterns.
Figure 2.10: Validation example for an area between Russia and Azerbaijan , on 3 January 2023: a) Subset of the global SCF product derived from SLSTR data, b) reference snow map from Landsat 9 OLI-2 scene (path 168, row 031), c) false-colour composite of the Landsat scene at original grid spacing, d) difference map showing SLSTR-based SCF minus reference snow map in per cent.
2.3.5. Summary
Across all four case studies, the SLSTR-based SCF product demonstrates a consistent ability to capture broad snow cover patterns across diverse and complex terrains. In fully snow-covered regions, the product consistently identifies snow presence, often assigning SCF values close to 100%. While this highlights its strong detection capability in homogeneous snow-covered areas, it also leads to slight over-estimations compared to the reference snow maps, which capture more nuanced fractional values, especially in areas with complex topography or variable snow density.
The product commonly underestimates snow cover in shaded terrain and regions with low reflectance, where snow detection becomes more challenging due to limited surface illumination. Additionally, lower fractional snow cover values are underrepresented, indicating reduced sensitivity to patchy or marginal snow cover. In transitional zones between snow-covered and snow-free surfaces, the product tends to respond in a more binary manner, which limits its ability to reflect gradual changes during snow accumulation or melt.
Cloud contamination also poses a challenge, especially over bright snow-covered areas and at snow/no-snow boundaries, occasionally affecting the accuracy of the SCF retrieval. The cloud masking is partly caused by the applied post-processing using ERA5 temperature and precipitation reanalysis data, intended to reduce false positives. While this step helps limit overestimation of snow cover, it also tends to increase cloud detection, particularly in transition zones. In some regions, snow-free classifications appear despite the presence of small snow patches identified by the reference maps.
Geolocation shifts in the SLSTR-based SCF product are occasionally observed during the product quality assessment. Such spatial offsets can cause localized inconsistencies in snow cover classification from one day to the next.
Overall, the SLSTR-based SCF product delivers robust and reliable information on large-scale snow dynamics, offering valuable contributions to climate research.
3. Climate Change Assessment
This section will be added at a later date.
4. Compliance with user requirements concerning data quality
In this section, the validation results are compared with the GCOS-245 requirements for the ECV quantity "Area Covered by Snow" of the 2022 edition, updated in 2025 ( The 2022 GCOS ECVs Requirements ). Table 4.1 presents the spatial and temporal resolutions and the measurement uncertainty of the C3S SCE products, evaluated against the GCOS requirements updated in 2025 for the ECV quantity "Area Cover by Snow".
For each requirement in Table 4.1, a Goal (G), Breakthrough (B), and Target (T) value is specified (see General definitions). Bolded GCOS requirements indicate the compliance level achieved by the reported values for the C3S Snow Cover Extent products. Colours in the column "Reported value" indicate the threshold level of the GCOS-245 requirements (updated in 2025) currently met by the C3S SCE products:
- Green (not yet applicable) = reported value is within Goal.
- Light yellow = reported value is within Breakthrough.
- Orange = reported value is within Target.
- Red = reported value does not meet Target threshold.
- Grey = requirement is not applicable (N/A) for product.
All C3S SCE products meet the requirement for temporal resolution to be useful for climate applications (breakthrough threshold) and the minimum required measurement uncertainty, indicated by the target threshold. Regarding horizontal resolution, daily global satellite data available before February 2000 only support SCE retrieval at a coarser resolution of 0.05° (approximately 5 km). Since late February 2000, the SCE products have met the minimum requirement (target threshold) for horizontal resolution in non-mountain regions. However, the minimum requirement for horizontal resolution in mountainous areas of 0.5 km is not yet achieved by the C3S SCE products. The stability of the C3S SCE products remains to be assessed, as explained below.
Table 4.1: Compliance with GCOS-245 Requirements, 2022 edition - updated in 2025 for the ECV quantity "Area Covered by Snow". Details about the meaning of the styles and colours are provided in the text.
|
Requirement |
|
GCOS Requirement |
Reported value |
||
|
|
G |
B |
T |
||
|
Horizontal Resolution (km) |
Non-mountain |
0.1 |
0.5 |
1 |
01/1982 - 02/2000: A1 = 0.05° (ca 5 km) (larger than Target value for non-mountains and mountains before the year 2000) |
|
02/2000 - present: A1 = 0.01° (ca 1 km) (within Threshold) |
|||||
|
Mountain |
0.01 |
0.1 |
0.5 |
A1 > 0.5 km (does not meet Target value for mountains) |
|
|
Vertical Resolution |
|
- |
- |
- |
N/A |
|
Temporal Resolution |
|
3h |
daily |
monthly |
A2 = daily (within Breakthrough) |
|
Required Measurement Uncertainty (%) |
|
5 |
10 |
20 |
A3 <= 20% ubRMSD (within Threshold) |
|
Stability (%/decade) |
|
1 |
5* |
10* |
A4 **) |
*) Associated notes from GCOS-245 Requirements, 2022 edition - updated in 2025: "These values still lack justification in the scientific literature and need to be critically assessed."
**) The decadal stability of Snow Cover Fraction (SCF) estimates remains to be systematically evaluated in conjunction with climate change assessments. The stability of daily SCF products exhibits considerable variability per pixel, ranging from 0% to 100%. The daily variability per pixel is influenced by several factors, including the geolocation accuracy of the satellite observations, the illumination conditions per pixel, the presence of undetected or misclassified clouds, and potential SCF misclassification at the pixel level.
References
Barella, R., Notarnicola, C., C. Marin, V. Premier, N. Ciapponi, C. Mortimer, C. Derksen, and G. Schwaizer (2024). ESA CCI+ Snow ECV: Product Validation and Intercomparison Report, version 4.0, May 2024 (https://climate.esa.int/media/documents/Snow_cci_D4.1_PVIR_v4.0.pdf).
Keuris, L., M. Hetzenecker, T. Nagler, N. Mölg, and G. Schwaizer (2023). An Adaptive Method for the Estimation of Snow-Covered Fraction with Error Propagation for Applications from Local to Global Scales. Remote Sensing, 15:1231. doi:10.3390/rs15051231.
Welford B.P. (1962). Note on a method for calculating corrected sums of squares and products. Technometrics. 1962 Aug 1;4(3):419-20.
