Snow cover mapping with Meteosat third generation FCI: initial evaluation of the European Organisation for the Exploitation of Meteorological Satellites H SAF H43 snow mask product

H SAF snow cover extent products

The H10 product is a daily snow mask (i.e., SCE), created through snow detection using VIS/IR radiometry and multi-channel analysis with the Spinning Enhanced Visible and Infrared Imager (SEVIRI) instrument aboard Meteosat satellites (H-SAF_H10_PUM, 2018). It provides coverage across a pan-European region from 25°N to 75°N latitude (cf. Fig. 1). The SEVIRI instrument has an instantaneous field of view (IFOV) of 4.8 km at nadir, which increases to approximately 8 km in European coordinates, with a sampling rate of 3 km degrading to about 5 km. Each pixel is classified as either bare ground, cloud, snow, water, or no data. The H10 product consists of two main components: one for flat/forested regions, developed by the Finnish Meteorological Institute (FMI), and another for mountainous regions, developed by the Middle East Technical University (METU). Both parts were integrated into the operational environments of FMI and the Turkish State Meteorological Service (TSMS) by 2007. The data from both parts are merged at FMI using a mountain mask (H-SAF_H10_ATBD, 2011).

The H34 SCE product, which has been available since October 2017 as the successor to the H10, follows a similar generation chain. It utilizes the SEVIRI instrument on Meteosat satellites for daily snow detection via VIS/IR radiometry and multi-channel analysis. The product is produced using two independent algorithms—one for flat/forested areas (developed by FMI) and one for mountainous areas (developed by METU and executed at TSMS). The data from both algorithms are merged at FMI using a specific mask to combine information from both flat/forest and mountainous regions (H-SAF_H34_PUM, 2023). H34 offers an enhanced spatial domain, covering the full Meteosat Second Generation (MSG) disk with a horizontal resolution of ~5 km (cf. Fig. 1).

MODIS-derived reference snow maps

The validation of the H43 SCE product is conducted using reference snow cover maps derived from the MODIS MOD10A1 NDSI Collection 6.1 dataset (Riggs, Hall & Román, 2019). MOD10A1 is a daily global snow cover product generated from MODIS instrument onboard NASA’s Terra satellite. The MOD10A1 snow cover dataset has undergone extensive validation against both in-situ measurements and satellite-based reference datasets across a wide range of geographic settings, including complex mountainous terrains. Numerous studies have demonstrated its effectiveness, reporting accuracy levels between 77% and 100% (Gascoin et al., 2015; Klein, Barnett & Lee, 2003; Parajka & Blöschl, 2006; Tekeli et al., 2005). While some uncertainties arise due to cloud contamination, as well as minor accuracy reductions linked to snowpack properties and variations in illumination conditions (Brubaker, Pinker & Deviatova, 2005; Crawford, 2015), MOD10A1, with its 500 m spatial resolution, can be regarded as a well-established dataset for snow detection across diverse landscapes and climatic conditions. Given that this initial validation spans three winter months (December 2024–February 2025), MODIS provides the most suitable reference due to its daily temporal coverage and moderate 500 m spatial resolution, offering dense temporal sampling consistent with the period analyzed by the H SAF products.

In this study, reference snow cover maps are generated by applying a threshold of NDSI ≥ 0.4 to classify snow-covered pixels. This threshold has been widely adopted in the remote sensing community as a standard baseline for snow detection (Hall et al., 2002; Klein, Barnett & Lee, 2003). The threshold of 0.4 has demonstrated robust performance across diverse geographic and climatic settings and forms the foundation of the NASA MODIS snow algorithm. While several studies have explored the possibility of regionally optimized or dynamic NDSI thresholds to account for land cover heterogeneity, atmospheric conditions, or snow grain size (e.g., Gascoin et al. (2015), and Parajka & Blöschl (2006)), the universal threshold of 0.4 remains a practical and well-validated choice for large-scale, multi-region analyses. Given the wide spatial coverage of the current study and the need for consistency across three distinct mountainous regions, the use of the established 0.4 threshold ensures comparability and aligns with standard practices in operational and research-based snow mapping.

To generate the reference snow cover maps for validating H43, a systematic re-gridding and aggregation approach is employed to align the higher-resolution MODIS data with the coarser H43 (~2 km) spatial resolution. The methodology follows these steps: Each MODIS pixel with an NDSI ≥ 0.4 is classified as snow, while pixels with NDSI < 0.4 are labeled as non-snow. MODIS snow classifications are then aggregated over the exact footprint of each corresponding H43 (~2 km) pixel. The most frequently occurring class within the footprint (snow, non-snow, cloud, water, unclassified, no data) is assigned as the label for the corresponding H43 pixel. In cases where equal numbers of snow pixels and pixels with labels other than snow exist within the footprint, the pixel is always labeled as snow to ensure that snow presence is not underestimated.

This method enables a robust pixel-wise validation framework, ensuring that the reference dataset accurately represents the ground truth snow conditions for evaluating the performance of the H43 product. In addition to validating H43, reference snow cover maps are also produced for H34 to facilitate a direct intercomparison between these two products as applying the same approach in H43, yet taking the H34’s footprint into account (i.e., (~5 km). This intercomparison provides further insights into the improvements introduced in H43 and allows for a more comprehensive evaluation of the evolution of snow cover detection capabilities within the H SAF project. A set of SCE reference images for both H34 and H43 over the European Alps is illustrated in Fig. 4.

WMO In-situ snow depth dataset

In-situ datasets consist of the observations from 176 synoptic stations (Alps: 125, Turkey: 48, Georgia: 3) from the Regional Basic Synoptic Networks (RBSNs) of WMO. The stations within the RBSNs are strategically distributed to ensure comprehensive coverage and timely data exchange on a global scale. SD observations from synoptic stations were recorded at standard times—00:00, 06:00, 12:00, and 18:00 UTC—covering the period from 2010 to 2025. The daily mean SD for each station was determined by averaging the recorded SD values. The spatial distribution of the stations with respect to 500 m elevation ranges is given in Fig. 5. For the in-situ validation, each station record was classified as snow-covered when the measured snow depth (SD) was equal to or greater than 5 cm, and snow-free when SD < 5 cm. This threshold provides a conservative and scale-consistent definition of snow presence, minimizing the influence of shallow or patchy snow that may not be reliably detected at the ~2 km spatial resolution of the H43 product. The selected 5 cm cut-off is consistent with the WMO Guide to Instruments and Methods of Observation, Volume II–Measurement of Cryospheric Variables (WMO-No._8, 2024) and with recent validation studies (Avanzi et al., 2023; Dong, 2018; Hall et al., 2019; Piazzi et al., 2019). The number of stations according to 500 m elevation zones are given in Table 3.

Results from the WMO snow depth dataset

This interpretation summarizes the snow detection performance of the H43 and H34 products based on in-situ observations from synoptic ground stations. The comparison includes the European Alps, Turkey, and the Georgia–Caucasus region. It should be noted that the number of in-situ observations is limited in each region and does not match the spatial and temporal coverage of MODIS-derived reference snow maps. Additionally, it should be noted that each in-situ snow depth observation was matched with the corresponding daily H43 and H34 pixel value at the same geographic location, ensuring a direct day-by-day comparison consistent with the temporal resolution of the satellite products. No temporal averaging or multi-day aggregation was applied, as the validation aimed to preserve the daily dynamics and reflect the true temporal behavior of the H43 and H34 products.

Over the Alps region, the in-situ validation dataset consisted of approximately 2,500 snow depth (SD) observations gathered during the three winter months from up to 125 meteorological stations. These observations were used to generate pixel-level categorical comparisons between satellite-derived snow classifications and ground-based snow presence. While the detailed A, B, C, and D statistics are not tabulated, they underpinned the derivation of the performance metrics shown in Fig. 7. The number of satellite–in-situ comparisons (i.e., pixel–day pairs) per product and month varied between 550 and 1,056, providing a robust basis for evaluating detection performance.

The in-situ validation results over the Alps reveal that H43 delivers more reliable snow detection performance than H34 across all three winter months. The improvement is most evident in the POD, which was consistently higher for H43 (0.77–0.79) compared to H34 (0.67–0.75). In contrast, H34 exhibited higher miss rates (C) and FAR, particularly in December (H43 = 0.18, H34 = 0.16). Both products achieved similar overall accuracy (ACC) in December (H43 = 0.68, H34 = 0.69), but H43 demonstrated clearer improvements in the following months—by January, ACC increased to 0.75 for H43 vs. 0.64 for H34, supported by a reduced FAR (0.08 vs. 0.13). This trend continued into February, where H43 reached its highest performance (POD = 0.79, FAR = 0.09, ACC = 0.75), while H34 remained lower (POD = 0.68, FAR = 0.12, ACC = 0.64).

Overall, these results confirm that H43 provides more consistent and accurate snow detection over complex Alpine terrain, particularly during mid- and late-winter, when surface heterogeneity and persistent cloud cover pose challenges for satellite-based retrievals.

In Turkey, the in-situ SD observations contributed a total of ~650 satellite-to-ground match-ups spanning the three winter months. These were derived from measurements at up to 48 synoptic stations, with individual monthly station counts ranging from 14 to 47. Pixel-level classification comparisons were constructed based on the A, B, C, and D categories, reflecting agreements, false detections, omissions, and no-snow confirmations. While the full breakdown of these counts is not shown, they served as the basis for deriving the FAR, POD, and ACC metrics presented in Fig. 7.

The monthly validation metrics over Turkey indicate that both H34 and H43 demonstrate strong performance in detecting snow cover, but H43 shows overall improvements in key accuracy indicators, especially in the later months of the season. H34 already shows a very high POD across all 3 months (above 0.98), indicating a strong ability to correctly identify snow-covered areas. However, H43 exhibits even higher or comparable POD values in each month. Notably, in February 2025, H43 achieves a near-perfect POD of 0.9957, confirming its excellent snow detection capability. While H34 had a consistently high hit rate (A), H43 maintained similarly high hit rates and slightly improved detection performance, especially in months with increased snow cover extent.

H43 presents a modest reduction in false alarms compared to H34 in February, with a FAR of 0.1673 vs. H34’s 0.1480. However, in December and January, H43’s FAR is slightly higher, which may relate to localized cloud-snow confusion or terrain-induced variability in reflectance. These false alarms (B) were slightly more frequent in H43 early in the season but remained within acceptable limits and did not overshadow the gains in detection performance. In terms of overall classification accuracy, H43 demonstrates improved or comparable ACC values for each month. In January and February, H43 surpasses H34, achieving 0.7945 and 0.8423, respectively, compared to H34’s 0.7474 and 0.8571. The slight drop for H43 in February is negligible, particularly given its stronger POD and comparable FAR. The overall agreement with ground truth is clearly higher for H43 in January, which appears to be a transitional month in terms of snow conditions.

In the Georgia–Caucasus region, in-situ snow depth observations provided a limited but valuable dataset for validation, with nearly 65 match-ups collected from up to 3 stations. The resulting A, B, C, and D counts enabled basic agreement assessments despite the relatively small sample size. These values underpinned the monthly calculation of the FAR, POD, and ACC metrics (cf. Fig. 7). Due to data scarcity, the metrics derived for some months may reflect limited representativeness and should be interpreted with caution.

For H34, hit rates remained consistently high, with perfect POD values across all 3 months. However, the extremely small sample sizes magnified the impact of false alarms. In December 2024, H34 showed a high false alarm rate (FAR) of 0.75. January 2025 had zero hits and only one false alarm, resulting in undefined (NaN) POD. February showed the most reliable detection performance, with a lower FAR (0.20), supported by a more balanced distribution of hits and correct negatives.

For H43, POD remained 1.0 in all months, indicating successful identification of all snow cases in the limited sample. However, similar to H34, FAR values were elevated, particularly in December (0.80) and February (0.56), reflecting challenges in discriminating snow from snow-free conditions in such a small dataset. January again lacked any true positives, making it difficult to draw meaningful conclusions for that month.

The corresponding confidence intervals of ACC, FAR and POD metrics for H34 and H43 are given in Tables 6 and 7, respectively. For H34, high detection consistency was observed over Turkey, where POD values exceeded 0.98 with narrow CIs (±0.007–0.009), while lower and more variable POD values (0.68–0.75) were found over the Alps. FAR values of H34 decreased from December to February (0.25→0.15) in Turkey and remained low (0.12–0.16) in the Alps. ACC ranged between 0.74–0.86 in Turkey and 0.64–0.69 in the Alps, with CI widths typically below ±0.03.

The H43 product demonstrated slightly improved performance, with POD values reaching 0.95–0.99 over Turkey and 0.77–0.79 over the Alps, accompanied by narrower CIs (±0.015–0.018). FAR values were lower than those of H34, particularly over the Alps (0.09–0.18), while ACC values increased modestly to 0.74–0.84 in Turkey and 0.68–0.75 in the Alps.

Overall, both products exhibited high sensitivity (POD), but H43 showed slightly higher false alarm rates than H34, especially in February. The scarcity of in-situ observations in the region and the disproportionate effect of individual misclassifications contribute to erratic month-to-month variations in accuracy metrics. Despite these limitations, H43 consistently maintained 100% detection capability when snow was present, suggesting robustness in hit rate performance under data-constrained conditions. Future assessments in this region would benefit greatly from denser and more consistent ground station coverage to improve the statistical reliability of validation results.

Across Alps and Turkey regions, the analysis quantifying the percentage of snow conditions that could not be detected due to cloud cover under all-sky conditions (cf. Statistical Metrics used in the Validation) demonstrate the clear benefit of the higher temporal resolution provided by the Meteosat missions. Over the Alps, the percentage of snow observations missed due to cloud contamination was 28.4% for H34 and 30.4% for H43, compared with 47.1% for MODIS. In Turkey, similar trends were observed, with missed snow rates of 25.2% for H34, 25.7% for H43, and 48.9% for MODIS. These findings confirm that the frequent observations from MSG (every 15 min) and MTG (every 10 min) substantially reduce the snow data gaps due to cloud contamination, improving the temporal continuity of snow cover monitoring. The results are consistent with those reported by Sürer & Akyürek (2012) and Sürer, Parajka & Akyürek (2014), who demonstrated that merging multiple SEVIRI acquisitions per day significantly improves snow detection in mountainous areas by mitigating the effects of cloud and illumination variability. Overall, the present analysis shows that H43, derived from the MTG-FCI sensor, preserves the operational strengths of its predecessor while offering enhanced capability for near-real-time snow detection under variable atmospheric and illumination conditions.

In summary, the in-situ validation reinforced the improved capabilities of H43 over H34 across a range of geographic and climatic conditions, particularly in complex mountainous terrains and during periods of peak snow accumulation. While the results from WMO synoptic stations confirmed H43’s enhanced accuracy and detection reliability, the limited spatial and temporal density of in-situ observations—especially across under-monitored regions such as the Georgia–Caucasus—posed constraints on comprehensive performance assessment. To overcome these limitations, this study employed a cross-sensor validation framework by leveraging MODIS MOD10A1 NDSI-based reference snow cover maps, which offer superior temporal continuity, higher spatial coverage, and consistent data availability. Consequently, the combined use of satellite-based reference datasets and ground observations ensures a more complete and statistically meaningful assessment of snow detection performance.

Results from the MOD10A1 NDSI-derived reference dataset

The cross-sensor validation using MODIS-derived reference datasets provides a more robust and spatially comprehensive evaluation framework, as it incorporates a significantly larger number of daily observations in the error matrix. Both H43 and H34 were validated against daily MODIS MOD10A1 snow-cover maps using corresponding daily composite images, ensuring full temporal consistency and comparability between datasets. The high spatial resolution and daily temporal coverage of MODIS ensure broader geographic representation and improved sampling across diverse snow and land surface conditions. This comprehensive coverage enhances the statistical reliability of accuracy metrics, supporting a more meaningful comparison of snow detection performance between H34 and H43 products.

The minimum, maximum and mean values of the statistical metrics over the European Alps is provided in Fig. 8, together with the cloud cover percentage. The results show that H43 consistently performs better than H34, with higher POD and ACC values, and a slightly lower FAR on most days. The lowest PODs were recorded in early December, likely due to increased cloud cover. Both products maintained high ACC overall, but H43 showed stronger alignment with the MOD10A1 NDSI reference dataset. Monthly analysis indicates that in December, both products had similar POD values near 0.77. In January, H43 improved significantly to 0.92, while H34 remained steady. By February, H43 reached its peak POD at 0.95, reflecting enhanced snow detection. FAR values were lower in December, suggesting fewer false positives early in the season. However, they rose slightly in January and February, likely due to changing snow conditions and melting events. Accuracy stayed consistently high, with H43 achieving the highest ACC of 0.97 in February. Cloud cover strongly influenced both products, lowering POD and ACC on heavily clouded days, though H43 proved more robust. The most cloud-affected periods were early December and mid-February, during which accuracy occasionally declined.

Overall, H43 showed better snow detection performance than H34, especially in terms of POD and ACC. H34 had more variability and a slightly higher FAR. Seasonal trends revealed that POD improved over time, with a clearer increase in H43. Although cloud cover remained a limiting factor, its impact was less noticeable in H43, suggesting that the updated algorithm handles cloud interference more effectively. These results underline the improved ability of H43 to detect snow cover in complex terrain such as the European Alps.

Over Turkey, H43 consistently outperformed H34 in detecting snow cover. The POD for H34 ranged from 0.615 to 0.976, with an average of 0.88. In comparison, H43 showed a higher and more stable range between 0.80 and 1.00, averaging 0.94 (cf. Fig. 9). The FAR for H34 varied between 0.0009 and 0.1701, while H43 ranged from 0.007 to 0.237. Despite the broader range, H43’s mean FAR was 0.070, only slightly higher than H34’s 0.060, indicating that the improved detection did not come at the cost of significantly more false alarms. In terms of ACC, H43 maintained stronger consistency, ranging from 0.857 to 0.996, compared to 0.892 to 0.987 for H34. The mean ACC was 0.96 for H43 and 0.95 for H34, highlighting H43’s improved performance in capturing snow distribution across Turkey.

Monthly analysis shows that snow detection improved from December to February. In December, the average POD was 0.858 for H34 and 0.931 for H43. By February, detection increased to 0.896 for H34 and 0.959 for H43. False alarm ratios stayed relatively stable, with minor changes in January likely due to changing snow conditions. Accuracy was highest in February when snow classification became more consistent. Cloud cover had a strong impact, especially in early December and mid-February, lowering accuracy for both products. However, H43 handled cloud interference better than H34, maintaining steadier performance. Overall, the results over Turkey confirm that H43 performs better than H34, offering higher detection rates and better accuracy, with only a small increase in false alarms. Seasonal changes affected both products, but H43 proved more effective, especially in cloud-affected conditions.

H43 generally showed better snow detection than H34 over the Georgia-Caucasus region, with higher POD and ACC values and only a slightly higher FAR. Both products performed better in the later months, with seasonal effects influencing detection. While cloud cover continued to impact accuracy, H43 proved more effective in handling it. Over the full period, H34 had a POD between 0.674 and 0.939, averaging 0.854, while H43 ranged from 0.712 to 0.955, with a slightly higher average of 0.858 (cf. Fig. 10). FAR values were similar overall: H34 ranged from 0.028 to 1.000, and H43 from 0.011 to 0.372, with mean values of 0.102 and 0.128, respectively, indicating a small increase in false positives for H43 under certain conditions.

H43 consistently achieved higher overall accuracy than H34, with ACC values ranging from 0.817 to 0.991, compared to 0.902 to 0.985 for H34. The mean ACC was 0.932 for H43 and 0.91 for H34, confirming H43’s stronger agreement with the reference data. Monthly trends show increasing POD for both products from December to February, with H43 improving more significantly. In December, average POD values were 0.854 for H34 and 0.923 for H43, rising in February to 0.858 and 0.937, respectively. FAR remained generally stable, though slightly higher in January—possibly due to transitional snow conditions. ACC followed a similar pattern, with the highest values in February when snow cover detection was more stable. Cloud contamination, especially in early December and late February, caused noticeable drops in POD and ACC. Still, H43 showed better resistance to cloud-related impacts.

The H43 product is based on blending 60 consecutive images per day, which is foreseen as an alternative to different filtering methods used for cloud reduction in optical remote sensing products. The results indicate that the blending of multiple observations during the day allows a significant cloud reduction over the three study regions.

The confidence interval (CI) analysis derived from MODIS-based validation provided a detailed assessment of the reliability and temporal stability of the H34 and H43 snow products (cf. Tables 6 and 7, respectively). For H34, POD values were consistently high across all regions, ranging from 0.79 to 0.92 with narrow confidence intervals (±0.001–0.002). The highest detection performance was observed over Turkey and Georgia, while slightly lower but stable POD values were recorded over the Alps. FAR remained low throughout the winter period, decreasing from approximately 0.06 to 0.04 in Turkey and maintaining comparable levels in Georgia and the Alps. ACC values were uniformly high for H34, typically exceeding 0.90 across all regions with very narrow confidence intervals (±0.0003–0.0008), confirming the statistical robustness of the validation results.

The H43 product demonstrated slightly improved detection and overall performance compared to H34. POD values increased to 0.92–0.97 in Turkey and Georgia, and to 0.81–0.88 over the Alps, accompanied by similarly narrow CIs (±0.001). FAR values for H43 remained low, generally between 0.04 and 0.07, indicating consistent discrimination of snow-free areas. Accuracy values exceeded 0.93 for all regions, with minimal seasonal variation and CI widths below ±0.0005. These results highlight the strong statistical stability of both products, with H43 exhibiting marginally higher accuracy and lower uncertainty than H34.

The comparison between MODIS-based and in-situ validations clearly demonstrates the effect of sample size on the statistical confidence of accuracy metrics. Owing to the substantially larger number of match-up observations in the MODIS-based validation, the CIs of all performance metrics (ACC, POD, and FAR) became significantly narrower compared to those derived from the limited in-situ dataset. This reflects the expected statistical behavior of binomial estimators, where larger sample sizes reduce the uncertainty of proportion-based metrics. Consequently, the narrower CIs obtained from the MODIS-based analysis indicate a higher level of statistical precision and reliability in the estimated accuracy values, whereas the wider intervals observed in the in-situ validation mainly arise from the smaller number of collocated snow observations.

Performance with respect to elevation and aspect

To complement the overall performance assessment, the spatial distribution of classification errors—specifically false alarms (i.e., B) and missed detections (i.e., C)—was further analyzed with respect to elevation zones and terrain aspect across the three validation regions. This terrain-based evaluation provides insight into the topographic dependencies of the H34 and H43 products, helping to identify systematic retrieval challenges associated with snow detection in varying altitudinal belts and slope orientations. Such analysis is crucial for improving algorithm robustness in complex mountainous environments.

All evaluations were performed using percentage values rather than raw pixel counts. This approach was necessary because the H34 and H43 products differ in spatial resolution. Direct comparison of raw pixel counts could lead to biased interpretations, as the total number of classified pixels per unit area is not the same between the two datasets. It was preferred to focus on false alarms (B) and misses (C) since they highlight two key sources of uncertainty: overestimation (false detection of snow) and underestimation (failure to detect existing snow). These two errors directly affect the usability of satellite snow products in hydrology, modelling, and snowpack monitoring. Additionally, the analysis focuses on elevation zones starting from 1,000–1,500 m, as snow cover below this threshold is typically rare and short-lived. Including lower zones could distort results due to limited observations and higher variability. By focusing on mid- to high-elevation ranges, the assessment better reflects product behavior under consistent snow conditions relevant for cryosphere monitoring.

For the European Alps, within the 1,000–1,500 m elevation band—where snow cover typically becomes more intermittent and sensitive to short-term meteorological variability—H34 and H43 snow products exhibit comparable behavior in terms of classification uncertainty, though with nuanced differences (cf. Fig. 11). Misses remain slightly higher in H34 (~1.58%) relative to H43 (~1.46%), indicating that H34 may occasionally overlook snow patches in areas with transitional snow cover or complex surface conditions. On the other hand, false alarms are marginally more frequent in H43 (~0.45%) than in H34 (~0.34%), suggesting a greater likelihood of overestimating snow presence under fluctuating radiometric or terrain-driven reflectance patterns. This elevation range, often characterized by fragmented snow fields and active melt cycles, appears to challenge both products, though H43 adopts a slightly more inclusive snow-mapping approach. Looking across directional slopes, north- and east-facing aspects tend to enhance these differences. H34 shows a moderate level of misses over shaded and potentially snow-retentive orientations, while H43 records somewhat increased false alarm levels in the same sectors. These distinctions highlight the varying sensitivities of the two products to topographic modulation of snow visibility—whether through early morning illumination, slope shading, or terrain-forced heterogeneity.

Both H34 and H43 snow cover products exhibit moderate levels of classification uncertainty in the 1,500–2,000 m elevation range, yet they differ in how that uncertainty manifests. H34 shows a higher rate of misses (~1.71%) compared to H43 (~1.24%), suggesting that it may apply more selective snow labeling in areas where snow conditions are variable or less clearly distinguishable. This could be influenced by partial snow coverage, shadow effects, or mixed-pixel complexities in mid-elevation terrain, where snowpack tends to be less stable than at higher altitudes. Conversely, false alarm rates are higher in H43 (~0.37%) than in H34 (~0.21%), indicating that H43 is more likely to misclassify snow-free surfaces as snow-covered in this elevation band. This behavior may reflect increased surface brightness, terrain reflectance anisotropy, or temporary snow events that lead to misclassification under certain illumination conditions. Aspect-based analysis reveals that north- and east-facing slopes show relatively higher rates of both misses and false alarms in H43, suggesting sensitivity to terrain-induced illumination effects or early morning solar exposure. H34, while showing lower false alarm percentages overall, tends to miss more snow presence across these aspects.

In the 2,000–2,500 m elevation range, H34 exhibits a higher average rate of misses (~1.16%) compared to H43 (~0.60%), indicating a more selective detection of snow-covered areas within this elevation band. This behavior may reflect the influence of residual cloud contamination, terrain-induced shading, or spectral confusion between snow and other high-reflectance surfaces such as rocks and glacier ice. In contrast, H43 demonstrates a more complete snow retrieval pattern, likely supported by its conservative thresholding strategy and possibly more consistent classification over mountainous terrain under clear-sky conditions. At the same time, H34 shows a higher false alarm rate (~0.74%) relative to H43 (~0.12%), suggesting that H34 is more prone to assigning snow labels to snow-free pixels, especially in radiometrically complex or geometrically challenging areas. This tendency aligns with H34’s generally more sensitive detection approach, which may prioritize completeness over strict precision. Aspect-wise, north-, northeast-, and southeast-facing slopes—which exhibit diverse snow retention and melting dynamics—tend to show higher rates of both misses and false alarms in H34, particularly where topographic shading or variable solar input can complicate snow identification. In contrast, H43 maintains consistently low percentages of both misses and false alarms across these directional classes, suggesting a more robust and stable snow classification under varying topographic and illumination conditions.

In the ≥2,500 m elevation range—the highest and most snow-dominated terrain of the Alps—the H34 and H43 snow cover products display distinct behaviors in terms of classification performance, particularly regarding misses and false alarms. H34 exhibits a higher rate of misses (~1.16%) compared to H43 (~0.60%), suggesting that H34 applies a more selective snow classification approach under conditions of spectral and topographic complexity. Meanwhile, H43 reports a slightly higher false alarm rate (~0.08%) than H34 (~0.03%). When aspect is considered, north-, northeast-, and northwest-facing slopes, which typically favor longer snow retention, show relatively high miss rates in H34 compared to H43. This suggests that H34 may underrepresent snow extent in shaded or cold-facing slopes. Likewise, east- and southeast-facing aspects, where snowmelt tends to occur earlier due to increased solar exposure, also exhibit higher misses in H34, possibly due to greater variability in snow cover or mixed-pixel effects. H43, in contrast, maintains a more stable distribution of both misses and false alarms across aspect classes, indicating consistent snow detection performance despite the ruggedness of the terrain.

In the 1,000–1,500 m elevation range in Turkey, H34 and H43 snow products show differences in how they identify and miss snow-covered areas (cf. Fig. 12). Based on normalized values, H34 has a lower false alarm rate (~0.23%) than H43 (~0.29%), meaning it is slightly more cautious when classifying snow presence. This may help reduce overestimation of snow in areas where snow is thin or transient. On the other hand, H34 reports a higher rate of misses (~0.36%) compared to H43 (~0.23%), which shows that H34 may not capture some snow-covered areas in this elevation zone. These could be areas with patchy snow, terrain shadowing, or mixed surfaces, where snow detection is more difficult. When aspect is considered, north- and northeast-facing slopes—which typically hold snow longer—tend to show more misses in H34, while east-facing slopes contribute to slightly higher false alarms in H43. These differences reflect how each product responds to topographic and radiometric conditions that affect snow visibility.

In the 1,500–2,000 m elevation range in Turkey, H34 and H43 snow products show noticeable differences in both false alarm and miss rates. H34 has a lower false alarm rate (~0.39%) compared to H43 (~0.50%), meaning H34 is less likely to label snow-free areas as snow-covered. This may help reduce overestimation in areas where terrain or lighting conditions could cause confusion. At the same time, H34 has a higher miss rate (~0.79%) than H43 (~0.48%), suggesting that it may not detect snow in some locations where H43 does. These misses in H34 could be linked to shaded areas, early melt zones, or areas with thin or patchy snow cover. Regarding aspect, north- and east-facing slopes tend to show more misses in H34, possibly because of their longer snow retention and more challenging radiometric conditions. H43, while more inclusive in snow detection, tends to produce slightly more false alarms, especially on slopes where surface brightness varies.

In the 2,000–2,500 m elevation range, H34 and H43 show different behaviors in terms of snow classification. H34 has a lower false alarm rate (~0.17%) than H43 (~0.23%), which means H34 is more selective in identifying snow and is less likely to wrongly label snow-free areas as snow-covered. However, H34 reports a higher miss rate (~0.55%) compared to H43 (~0.29%), suggesting that H34 may not fully capture snow presence in certain parts of this elevation zone. These misses can occur in areas with complex terrain, partial snow coverage, or shaded slopes, where snow is more difficult to detect. When aspect is considered, north- and northeast-facing slopes contribute to slightly higher miss rates in H34, while H43 provides more consistent detection across aspects, though with a small increase in false alarms. The overall trend indicates that H43 is more inclusive in snow detection, while H34 continues to prioritize avoiding over-detection.

At the highest elevation band (≥2,500 m), snow is expected to be widespread and persistent. In this zone, H34 maintains a low false alarm rate (~0.06%), slightly lower than H43 (~0.08%), indicating that it rarely labels snow-free areas as snow-covered. However, misses are more common in H34 (~0.50%), while H43 shows a lower miss rate (~0.24%), suggesting that H43 is more successful in capturing snow presence under high-elevation conditions. The higher omission in H34 may result from the effects of slope, shading, or spectral confusion with bright surfaces like rock or glacier edges. Aspect analysis shows that H34 tends to miss more snow on north- and northwest-facing slopes, which generally retain snow longer. H43 performs more evenly across directional slopes, with only a slight increase in false alarms, pointing to its consistent classification behavior even in complex topography.

In Georgia, for the 1,000–1,500 m elevation range, both H34 and H43 products show relatively close values for false alarms and misses, though with some differences (cf. Fig. 13). H34 reports a slightly higher miss rate (~1.08%) than H43 (~0.92%), which means that H34 leaves out more snow-covered areas in this elevation zone. These misses could result from partial snow cover, terrain shadows, or radiometric complexity. False alarm rates between the two products are quite similar. H34 shows a false alarm rate of ~0.78%, while H43 is at ~0.77%, indicating that both products have a comparable tendency to overestimate snow presence in this elevation class. Aspect-based trends reveal that H34 has somewhat more frequent misses on north-facing and northeast-facing slopes, while H43 performs more evenly across aspect classes. Small variations in terrain exposure and snow retention may explain these differences.

In the 1,500–2,000 m elevation zone, both H34 and H43 show moderate levels of omission and commission errors, but with slight differences in magnitude. H34 has a miss rate of ~1.30%, while H43 is slightly lower at ~1.08%. This suggests that H43 detects more snow-covered areas in this elevation range, possibly due to a less restrictive detection threshold or better sensitivity to transitional snow conditions. For false alarms, the two products are again close in value. H34 records a rate of ~0.81%, while H43 is at ~0.79%. This indicates that both products show similar levels of over-detection in areas where snow may be patchy, melting, or confused with other bright surfaces. When examining aspects, north-facing and east-facing slopes contribute to the differences in omission rates, especially in H34. In these directions, snow tends to persist longer, which can challenge detection under certain lighting and viewing conditions.

Within 2,000–2,500 m, snow cover is more consistent, but detection performance still varies between the products. H34 shows a higher miss rate (~0.81%) compared to H43 (~0.61%), meaning H43 captures a slightly larger portion of the snow-covered area in this elevation band. This can be helpful in ensuring completeness in mountainous snow monitoring. False alarms are also somewhat different. H34 records a rate of ~0.45%, while H43 is slightly lower at ~0.40%, indicating both products are relatively stable in avoiding over-detection, with a small advantage for H43. In terms of terrain orientation, north- and northeast-facing slopes again show a larger share of H34’s misses.

And finally, in the highest elevation class (≥2,500 m), H34 shows a higher miss rate (~0.72%) than H43 (~0.46%), meaning it is more likely to leave out snow-covered areas in this range. This could be due to mixed terrain conditions or the presence of persistent snow that is harder to separate from other bright surfaces. False alarm rates are low for both products, but H34 reports a slightly higher value (~0.18%) than H43 (~0.14%). Both products perform well in avoiding over-detection at these altitudes, where snow presence is typically stable. Aspect-wise, the pattern continues: H34 tends to miss more snow on cold-facing slopes, especially in complex terrain, while H43 maintains more balanced performance across aspect classes.