**3. Methods**

After introducing the retrieval algorithms implemented for the generation of the satellite snow products investigated in this study, the processing of in-situ data and imagery used to validate S-2 observations is described. The methodologies for the S-2-based assessment of H-SAF snow products and the assessment of S-2 imagery against in-situ data are presented and discussed, as well as the evaluation metrics.

#### *3.1. Satellite Retrieval Algorithms*

Snow detection can be retrieved from optical imagery through different algorithms, which generally rely on thresholding methods based on channel differences and ratios to exploit the different spectral properties of snow-covered areas with respect to snow-free surfaces and clouds. One of the most commonly-used indices is the normalized difference snow index (NDSI), which is defined as the difference of reflectance observed in a visible band and a shortwave infrared one, divided by the sum if the two reflectance values [72]. Indeed, since snow reflectance is high in the visible wavelengths and

low in the shortwave infrared ones, this method enables distinguishing snow from clouds and other non-snow-covered conditions [72,73]. However, it is noteworthy that the suitability of each retrieval algorithm necessarily depends on the main features of the satellite data to be processed [55,74,75].

## 3.1.1. Sen2Cor Algorithm

S-2 L1C data are downloaded from the Copernicus open access hub. The L1C image product consists of a series of 100 × 100 km2-tiles, each of which is made of thirteen compressed JPEG-2000 images, one for every single band. The MSI TOA reflectance images are processed through the Sen2Cor version 2.5.5, namely the last version of Sentinel-2 L2A prototype processor provided by ESA. Sen2Cor consists of ten main modules and it can perform the tasks of atmospheric, terrain and cirrus correction of L1C input data to generate optimally corrected BOA reflectance images. In this study the L2A\_SceneClass (SC) module is used to perform the classification of the input images and to generate Scene Classification (SCL) maps at a spatial resolution of 20 m. The SC algorithm allows the detection of clouds, snow and cloud shadows, and the generation of a classification map consisting of four different classes for clouds (including cirrus), together with six different classifications for shadows, cloud shadows, vegetation, soils/deserts, water and snow (Table 6). The SC module consists of the cloud/snow algorithm, the cirrus detection algorithm, and the cloud showdown detection algorithm to generate the classification map [76]. Each algorithm processes the TOA reflectance input data through a sequence of thresholding filters, which are applied to S-2 spectral bands, band ratios, and indexes. In the cloud/snow detection algorithm, each test provides a cloud probability, which is recursively updated at each step. After thresholding the brightness in the red region of the solar spectrum (band 4), all potentially cloudy pixels are filtered by thresholding the NDSI [72], which is evaluated from spectral bands 3 and 11. Snow confidence map is generated by detecting snow pixels, according to four successive filters using spectral bands 2, 3, 8, 11. Ancillary information on yearly snow climatology is used to define the monthly snow probability of each pixel and to discard possible false snow detections. All potentially snow pixels are then filtered by sequentially thresholding the reflectance in band 8 (NIR) and band 2 (blue), and the ratio between band 2 and band 4 to identify main water bodies. Lastly, possible false cloud detection at the boundaries of snowy regions is removed by performing a brightness test on band 12. Once cloud and snow confidence masks are generated, an optional spatial filter can be applied to reduce possible false cloud detection. The cirrus detection algorithm mainly relies on the reflectance thresholding of band 10, because of the high water vapor absorption in this region [77], and an additional cross check is performed against the probabilistic cloud mask.


**Table 6.** Classes of Sen2Cor SCL map.

#### 3.1.2. H-SAF H10 Algorithms

The distribution of snow cover and the non-uniformity of snow properties are significantly different over mountainous and flat/forested areas, since they strongly depend on the local topography and vegetation cover. Therefore, two distinct stand-alone algorithms are implemented within the product generation and applied according to a mountain mask defined in [56]. The algorithm for flat/forested areas has been developed by Finnish Meteorological Institute (FMI) [70,78]. The algorithm utilizes TOA radiances of six SEVIRI channels (0.635, 0.81, 1.64, 3.90, 10.80, 12.00 μm), the brightness temperatures of three channels (3.90, 10.80, and 12.00 μm), sun and satellite zenith and azimuth angles, the International Geosphere-Biosphere Programme (IGBP) land-cover type by the U.S. Geological Survey (USGS), and the land surface temperature (LST) classification produced by the EUMETSAT's Satellite Application Facility on Land Surface Analysis (LSA SAF) [77]. While information from channels around 0.635, 0.81, and 1.64 μm are used to classify different surface types [72,79,80], the algorithm exploits the radiance ratio of SEVIRI channels 2 (0.81 μm) and 3 (1.64 μm), and the brightness temperature difference of channels 10 (12.00 μm) and 4 (3.90 μm) to properly detect clouds [81,82]. The Middle East Technical University (METU) developed the algorithm for mountainous areas [56], which exploits 4 SEVIRI spectral channels (0.635 μm, 1.64 μm, 3.90 μm, 10.80 μm). Cloud discrimination is preliminary performed to identify cloud-free pixels by jointly using Cloud Mask (CMa) and Cloud Type (CT) products of the EUMETSAT's Nowcasting Satellite Application Facility (NWC SAF) [83]. Firstly, pixels having reflectance values higher than 0.35 are collected, because of the high visible reflectance of snow. Secondly, since snow cover has a low reflectance in the middle infrared and a high reflectance in the visible, pixels having snow index (SI) value lower than 0.6 are collected, which is evaluated by dividing channel 3 (1.64 μm) to channel 1 (0.635 μm) [84,85]. Lastly, pixels having temperature lower than 288 K on channel 9 (10.80 μm) are accepted, considering that the temperature of snow cannot exceed the freezing point [86]. It is noteworthy that sun zenith angle (SZA) thresholds are applied, SZA > 80 in the FMI algorithm and SZA > 85 in the METU algorithm are used for discarding the low-illuminated areas. No atmospheric correction is included in both algorithms. The final snow recognition product results from the merging of the products for flat/forested and mountainous areas over the full H-SAF spatial domain.

#### 3.1.3. H-SAF H12 Algorithms

Consistently with H-SAF H10 product, two different retrieval algorithms are separately applied for flat/forested and mountainous areas. Since the observing cycle of satellites over Europe is about 3 h, the scenes are multi-temporally analyzed to search for time instants of cloud-free conditions in 24 h. The retrieval algorithm of FSC in forested/flat areas has been developed at FMI. The method is based on a semi-empirical reflectance model [87], which evaluates the reflectance as a function of the snow-covered area by using visible and near-infrared data (visible band 1) [88]. Since forest transmissivity is of critical importance to estimate the snow-covered area in all conditions of forest coverage, the algorithm relies on the transmissivity map generated from reflectance data acquired at full dry snow cover conditions to guarantee a proper contrast between forest canopy and ground. However, a priori information on forests is not needed, because the effective average forest transmissivity is estimated from Earth observation reflectance data. Conversely, the retrieval algorithm for mountainous areas involves the thresholding of NDSI at 0.4, since it allows the derivation of the resulting average fraction of snow-covered area by retrieving snow and snow-free ground from satellite data according to the reflectance values [55]. Because in mountainous regions the sun zenith and azimuth angles, as well as direction of observation relative to these are the most limiting factors [89], possible terrain effects are properly removed from the measured radiance through a statistical-empirical correction method [90]. Once the visible channel is corrected due to topographic effect, the average reflectance values are determined from pixels of pure snow-covered area and pure bare ground. In defining the pure snow-covered area and snow-free bare ground, the NDSI threshold greater than 0.4 and lower than 0 are used, respectively. The original model has previously been developed by [91]. According to this approach, the pixel reflectance is modelled as a linear mixture of snow, individual tree species and snow-free bare ground (e.g., rock, soil, low vegetation). For mountainous areas the equation considers snow and bare ground, because of the general lack of trees at high altitudes:

$$R\_G = A\_{SW} R\_{SW} + A\_{BG} R\_{BG} \tag{1}$$

where *RG* is the modelled pixel reflectance for a given wavelength, *A* represents area fractions of a pixel (with *ASW* + *ABG* = 1), R is the reflectance, subscripts *SW* and *BG* refer to snow and bare ground, respectively.

It is noteworthy that both atmospheric and topographic corrections are implemented in H12 algorithms. The products for flat/forested and mountainous areas are merged over the full H-SAF area, according to the mountain mask defined in [56]. The merging algorithm is properly designed to minimize the projection errors [92].

#### *3.2. Validation of Sentinel-2 Imagery with In-Situ Data*

#### 3.2.1. Validation of Sentinel-2 Imagery by In-Situ Webcams

The validation is based on the comparison of single daily FSC values derived from camera observations to the corresponding single values obtained from S-2-derived FSC maps over the observed area. FSC values have been estimated by experts through the visual inspection of camera images. Visual inspections have been limited over area of interests (AOIs) selected according to both camera properties and the local topographic features, so that the snow cover is clearly visible, and the relative surface area can be estimated as accurately as possible. FSC values have been estimated in 10%-intervals (i.e., 0%, 10%, ... , 90%, 100%). The visual inspection of each image has been performed by 4 expert observers. With the aim of assessing the subjective error, the resulting RMSE of FSC estimates has been also calculated, as shown in Table 7. The study of Arslan et al. (2017) [33] has estimated that the subjective error is within 10%, in terms of FSC. For the comparison, average values of the visual estimates have been used to minimize the subjective error.


**Table 7.** AOI Sizes, corresponding number of S-2 pixels and subjective error of webcam data observers.

To properly validate the mapping of S-2-derived FSC, a mask for each selected webcam has been created by drawing polygons over the approximate AOIs using Google Earth. This has been done by visually comparing the landmarks in webcam images and Google Earth optical satellite data overlay. AOIs have been modified according to the landmarks so that the polygons were as accurate as possible. The polygons have been then converted into GeoTIFF files to be used in the FSC evaluation over each AOI from S-2-based snow cover maps (S-2-derived FSC). Area sizes of those polygons and the corresponding number of S-2 pixels are shown in Table 7. The comparison has been performed also over the three AOIs corresponding to relatively low numbers of pixels in S-2 grid, since in those sites most of the images were available either in full snow cover or snowless conditions, which make this analysis feasible. Along with FSC, cloud cover fraction is also calculated for each AOI.

As an example, AOI for the Kenttärova canopy camera and S-2 derived snow cover map over the AOI for 18 February 2018 are shown in Figure 9. In Figure 9c are reported the approximate camera FOV (white polygon) and the selected AOI (yellow polygon) in the Google Earth view. AOI from the camera FOV is reported in Figure 9a,b. The snow cover map derived from the S-2 image over the AOI polygon is shown in Figure 9d. In this example, while experts have estimated FSC as full snow cover (100%) from camera images (at 09:00 and 11:00), S-2-derived FSC (at 10:10) has been estimated as 82%. However, it is noteworthy that 38% of the analysed pixels over the AOI are cloudy or unclassified.

**Figure 9.** AOI for Kenttärova canopy camera and S-2 snow cover map over the AOI for 18 February 2018–(**a**) AOI marked on the camera image at 09:00 (yellow polygon); (**b**) AOI marked on the camera image at 11:00 (yellow polygon); (**c**) Approximate FOV of the camera (white polygon) and AOI (yellow polygon) in Google Earth; (**d**) Extracted snow cover map from S-2 image at 10:10 (white: snow; brown: no-snow; black: clouds and unclassified; red: camera location). AOI is approximately 0.25 km2, corresponding to 633 S-2 pixels.

(**d**)

After obtaining the value pairs for the comparison, the ones having cloud cover fraction over 50% have been filtered out. The value pairs of FSC have been compared and the resulting RMSE values have been evaluated.

#### 3.2.2. Validation of Sentinel-2 Imagery against Ground-Based Snow Measurements

The procedure to validate S-2 snow mapping against ground-based data relies on the thresholding of SD measurements to properly define snow and snowless conditions. According to the in-situ measures, the presence of snow is detected whenever a threshold of 5 cm is exceeded. This threshold has been set due to the expected uncertainty in measuring devices [93].

#### *3.3. Procedures of Cross-Sensor Comparison between Satellitesnow Products*

(**c**)

After processing S-2 data through Sen2Cor SC module, a quality check of the satellite data series has been performed through random visual inspections to prevent possible systematic inconsistencies. The comparison between observations sensed by different sensors on the same day is performed at the scale of S-2 tile. Indeed, the consistency assessment of H-SAF products against S-2 data is carried out individually over each single tile. In order to properly perform the comparison analysis, all the satellite snow products have been preliminarily re-projected to the same common image projection, namely WGS84/UTM. Since the analysis is tile-based, maps over the geographic extension of each selected S-2 tile are derived from the original full images of both H-SAF products. The selection of the data subset limited over the domain of each tile is carried out by considering the local coordinates of tiles borders, in order to properly guarantee the intersection of the satellite products over each S-2 tile.

#### 3.3.1. Comparison between Sentinel-Based Snow Masks and H-SAF H10

Figure 10 shows the comparison procedure. Firstly, binary snow masks (presence/absence of snow cover) are derived from both H10 and S-2 SCL maps. According to the original classification of SCL map (Table 6), vegetation, not vegetated, and water (Table 6, classes 4, 5, 6) pixels have been classified as no-snow pixels. Unclassified (Table 6, classes 0, 1, 2, 7) and cloud-contaminated (Table 6, classes 3, 8, 9, 10) pixels are flagged and neglected in the comparison of snow maps, with the aim of preventing possible cloud cover affecting the snow detection [57]. Consistently, H10-based snow masks are derived by considering snow and bare-soil pixels. Since the satellite products are differently gridded, the comparison is performed at the coarser spatial resolution of the H-SAF H10 [55]. For each H10 grid cell, the percentage of snow cover is determined according to the S-2 observations by counting the number of S-2 snow pixels versus the total number of S-2 pixels in the coarser cell [55]. This computation results into an S-2-based FSC map (S-2-derived FSC). In order to restore a binary snow mask, each resulting S-2-based coarse cell is then classified as snow if FSC is higher than 50%, otherwise it is classified as soil [57,94]. S-2-based coarse cell where more than the 50% of fine S-2 pixels are classified as cloud or unclassified are neglected in order not to compromise the analysis results. A preliminary analysis has been performed by testing threshold values equal to 25%, 50%, 75% to properly assess the impact of the thresholding of cloud cover at pixel scale. The results have revealed a poor sensitivity of the comparison procedure to the threshold value.

**Figure 10.** Flowchart of comparison procedure of H10 product.

#### 3.3.2. Comparison between Sentinel-Derived FSC Maps and H-SAF H12

Consistently with the assessment of H-SAF H10 product, the analysis of H12 data relies on the same procedure and assumptions. The only main difference is the lack of the thresholding of FSC derived from S-2 imagery over the coarser H-SAF grid (S-2-derived FSC), since it is directly comparable with H12 product. Indeed, this analysis compares FSC maps of H12 product, which are generated through retrieval algorithms (Section 3.1.3), with the S-2-based FSC maps, which are derived by counting the number of S-2 snow pixels versus the total number of S-2 pixels in the coarser H12 cell. The comparison scheme is reported in Figure 11.

**Figure 11.** Flowchart of comparison procedure for H12 product.

It is noteworthy that when mapping the snow cover through remotely-sensed optical imagery, forests constitute a challenge, since the canopy (1) partially obscures the signal along the path from ground to sensor and, (2) alters the observed reflectance [87]. Therefore, the accuracy of snow mapping generally decreases in forested areas with respect to non-forested regions [72,95,96]. Even though several methodologies have been proposed for the detection of snow under forest canopy [87,91,96–100], this issue still remains a critical research topic. Subpixel classification methods [87,101,102] are used to generate FSC maps with the aim of overcoming the limitations related to mixed-pixels problem affecting coarse-resolution imagery, namely possible mixtures of land cover classes (i.e., snow, soil, rock, vegetation, water, etc.) and area fractions of different cover classes within a pixel.

When assessing the consistency between S-2 and H12 data, the impact of vegetation on snow detection need to be investigated. Indeed, since the H12 retrieval algorithm in flat/forested area involves transmissivity maps, this product results from snow detection at ground level. Conversely, S-2 snow mask is derived from snow detection on canopy and thus it can be hindered by the presence of vegetation, mainly where forest cover is present. It is noteworthy that this difference in retrieval algorithms is supposed not to affect the analysis shortly after snowfall events, when forested areas are likely to be classified as snow pixels because of canopy interception [103]. On the other hand, during periods when no snowfall event occurs, the comparison between the two snow products can be more challenging. Indeed, especially in dense forests, the lack of intercepted snow can lead to a misleading S-2 classification as snow-free surface despite the presence of snow cover under canopy. Therefore, with the aim of addressing this critical issue, a further analysis has been performed according to the information on different vegetation types supplied by ESA GlobCover 2009 land cover map. The impact of the vegetation cover has been investigated throughout the whole analysis period by considering a sample of one tile in each country. The comparison between S-2-derived FSC and H12 product has been performed after preliminarily filtering out of the vegetated pixels in S-2 data by using the information supplied by GlobCover data for flat areas and without filtering in mountainous regions. For the filtering, S-2 derived FSC maps have been collocated with the GlobCover map. The pixels corresponding to flat regions in the mountain mask and belonging to the vegetation class V1 in GlobCover (described in Section 2.2) have been discarded from S-2 derived FSC maps. After that, the comparison has been performed through the same algorithm previously described. This procedure has also been applied for the vegetation class V2 to properly assess the impact of different vegetation types on snow detection. Results are presented separately for both classes.

## *3.4. Evaluation Metrics*

For the consistency assessment of the mapping of snow extent, a contingency table is evaluated (Table 8).


 (a), number of FALSE ALARMS (b), number of


From these classification results, different scores for dichotomous statistics are evaluated:

> POD = *a*/(*a* + *c*)

• Probability of detection:

> FAR = *b*/(*a* + *b*)

 (2)

 (3)


• Accuracy:

• False alarm ratio:

$$\text{A.C.C} = (a+d)/(a+b+c+d) \tag{5}$$

• Critical success index:

$$\text{CSI} = a/(a+b+c) \tag{6}$$

*Geosciences* **2019**, *9*, 129

• Heidke skill score:

$$\text{HSS} = 2(ad - bc) / \left[ (a + c)(c + d) + (a + b)(b + d) \right] \tag{7}$$

FSC is assessed through the evaluation of Root Mean Square Error (RMSE) with respect to the reference dataset -*FSCref* .

$$\text{RMSE} = \sqrt{\frac{\sum\_{i=1}^{n} \left( FSC\_{ref,i} - FSC\_i \right)^2}{n}} \tag{8}$$

#### **4. Results and Discussion**

#### *4.1. Validation of Sentinel-2 Imagery*

When validating S-2 imagery against in-situ observations, it is noteworthy that the evaluation metrics have been evaluated by considering ground-based datasets as the reference ones.

#### 4.1.1. In-Situ Digital Imagery

The comparison relies on a total of 50 pairs of FSC values, resulting from the analysis of matching webcam and S-2 images, both in Finland and in Italy. The evaluation has revealed a total RMSE value of 12.22%. In overall, S-2 snow mapping reveals a general FSC overestimation. When neglecting full-snow and bare-soil classification (i.e., FSC equal to 0% and 100%), the total RMSE value increases up to 19.82% for 19 value pairs. It is noteworthy that no outlier affects the distribution and the analyzed data boast a high correlation. Figure 12 shows the data scatterplot resulting from the comparison.

**Figure 12.** Distribution of compared value pairs.

The undesired cases resulting in high errors have been investigated in more detail. In Torgnon, where the AOI is selected over the mountainous area, the three scenes having the highest error are those affected by the largest cloud cover fraction, greater than 29%. Indeed, in the presence of patchy snow cover, partial cloud cover over the area is likely to unavoidably affect the FSC derived from S-2 data. Under conditions of patchy snow cover, S-2 data are affected by overestimation in Lompolojankka. Two occurrences are during the melting period, where the ground is mostly cover by meltwater. One further occurrence is in early winter, where snow cover is still not full and sparse vegetation is likely to hinder the visual inspection.

4.1.2. Ground-Based Snow Measurements

The validation of S-2 snow mapping against ground-based SD measurements in Turkey has revealed a significant consistency of satellite imagery, as evidenced by the highest number of hits and lower values of false alarms and misses (Table 9).

**Table 9.** Contingency table of ground-based validation of S-2 binary snow maps in Turkey for winter season 2017/18.


As shown in Table 10 reporting the resulting evaluation metrics, the remotely-sensed highresolution observations properly succeed in detecting the presence of snow cover.

**Table 10.** Evaluation metrics of ground-based validation of S-2 binary snow maps in Turkey for winter season 2017/18.


#### *4.2. Cross-Sensor Comparison of Snow Extent Products*

For each S-2 tile a pixel-to-pixel analysis has been performed to evaluate the consistency between the S-2-based maps of snow extent and H-SAF H10 product.

Figures 13–15 show the comparison results in terms of POD, FAR and ACC for each analyzed tile in Italy, Finland and Turkey, respectively. When assessing the evaluation metrics, it is noteworthy that H10 product generally reveals higher performances over flat areas (i.e., Finland), rather than over mountainous regions (i.e., Italian Alps and Turkey). This issue is mainly due to the impact of the local complex topography affecting the sensors capability to detect snow. Indeed, mapping the high spatial variability of snow cover distribution over mountain sides is a challenging task at the coarser satellite resolution. Conversely, when considering the vegetation cover of each pixel (Table 3), even the presence of the vegetation species supposed to hinder the snow detection (e.g., evergreen needle-leaved forest) results in a lesser impact than topographic factors. However, both in Italy and Turkey, H10 product reveals a slightly weaker reliability over the most vegetated tiles, namely T32TQS and T33TUM, and T36TWL, respectively. This result suggests that the vegetation has a greater impact where the local topography is complex, due to overlapping effects. Nevertheless, H10 product generally ensures an accuracy greater than 0.8, except for tiles T32TNS and T33TPS over Italian Alps.

With the aim of investigating whether product performances are affected by the seasonality of snow cover, the evaluation metrics have been assessed under different snow cover conditions. Indeed, three different periods have been individually assessed, namely early winter (i.e., October and November), winter (i.e., December-March), melting period (i.e., April and May) (Figure 16).

Consistently with the tile-scale analysis, the agreemen<sup>t</sup> between S-2 and H10 data is higher over flat areas (i.e., in Finland). The analyses show that in early winter the lower accuracy and higher FAR values of H10 product over flat areas are mainly due to frequent cloudiness, which is likely to affect the snow detection. However, it is noteworthy that the presence of canopy is likely to have a higher impact during early winter and melting period. Indeed, in those months, the presence of patchy snow cover and the lower frequency of snowfall events intercepted by canopy make the snow mapping more challenging, especially where dense forests are present. Furthermore, it is important to consider that the 50%-thresholding of FSC derived from S-2 data (Section 3.2) is likely to affect the analysis mainly during the transition periods, when patchy snow cover is present.

**Figure 13.** Evaluation metrics at tile scale in Italy.

**Figure 14.** Evaluation metrics at tile scale in Finland.

**Figure 15.** Evaluation metrics at tile scale in Turkey.

**Figure 16.** Seasonal evaluation metrics in Italy, Finland, and Turkey.

The higher consistency between H10 product and S-2 snow masks over flat areas is confirmed when assessing the evaluation metrics throughout the whole analysis period (Table 11). While H10 product reveals satisfying evaluation metrics of the same ranges in Finland and in Turkey, the Alpine complex topography strongly affects the snow mapping over this region, as proved by the poorer scores.


**Table 11.** Comparison between S-2-based snow masks and H10 product—Median values of evaluation metrics throughout the whole analysis period. H10 product requirements for both flat/forested (i.e., Finland) and mountainous areas (i.e., Italian Alps, Turkey) (PODthr and FARthr) are reported [104].

#### *4.3. Cross-Sensor Comparison of Effective Snow Cover Products*

The agreemen<sup>t</sup> between the mapping of the FSC derived from S-2 imagery and H12 product has been assessed according to the same pixel-to-pixel approach.

Like in the assessment of H10 data, the complex topography in mountainous areas affects the consistency between the two analyzed datasets, especially over the Italian Alps, where RMSE values are higher than the other case studies (Figure 17). However, RMSE scores are generally lower than 0.4, except for the same tiles in Italy, in compliance with the product requirements [105].

**Figure 17.** RMSE at tile scale in Italy, Finland, and Turkey.

Figure 18 shows that during the winter period RMSE values are generally higher than in other seasons. This issue is mainly due to overestimated classifications as full snow cover over the coarser spatial resolution of H12 product with respect to that derived from 20-m S-2 imagery, especially in mountainous regions.

The RMSE assessment over the whole analysis period confirms higher performances of H12 product over flat areas (i.e., Finland) than in mountainous regions (i.e., Italian Alps and Turkey) (Table 12). However, the product target requirements [105] are generally satisfied over both mountainous (RMSE ~ 30%) and flat (RMSE ~ 20%) areas.

**Figure 18.** Seasonal RMSE in Italy, Finland, and Turkey.

**Table 12.** Overall RMSE values for the comparison between S-2-derived FSC maps and H12 product. H12 product requirement for both flat (i.e., Finland) and mountainous areas (i.e., Italian Alps, Turkey) (RMSEthr) is reported [105].


Impact of Vegetation on Snow Detection

In order to properly assess the impact of the vegetation cover within the assessment of H12 product, the results obtained by filtering out the main classes V1 and V2 are individually evaluated against the reference ones relying on all S-2 pixels. As expected, the vegetation cover has a negligible impact on the comparison results in the Alpine region, since H12 and S-2 snow retrieval algorithms are consistent over the mountain mask (Section 3.2.2). Conversely, in Finland and Turkey the comparison procedure reveals a slight sensitivity to the different vegetation classes. While the snow detection is more affected by V1-vegetation class (i.e., needle-leaved evergreen forest) in Finland, V2-vegetation class (i.e., broadleaved deciduous forest) has a higher impact in Turkey. Figure 19 shows that by filtering out V1-vegetation pixels, the RMSE value in Finland increases with respect to the all-pixels analysis. The reduction in consistency between H12 and S-2 snow mapping suggests a good agreemen<sup>t</sup> of the two datasets over dense forests mainly due to the long-lasting snow interception over canopy during the winter season. Consistently, in Turkey the resulting RMSE slightly increases when filtering V2-vegetation pixels.

**Figure 19.** Impact of vegetation types on the comparison procedure of H12 product.
