Testing Satellite Snow Cover Observations Using Time-Lapse Camera Images in Mid-Latitude Mountain Ranges (Northern Spain)

Abstract

Reliable monitoring of snow cover in mountainous regions remains a challenge due to frequent cloud cover and the revisit limitations of optical satellites. This study compares satellite snow-cover records with >99,000 ground-based time-lapse camera observations across northern Spain (2003–2025). Cloud cover caused major data loss, with up to 57% of satellite images affected. Effective revisit intervals (the average time between usable images) diverge substantially from nominal values: 2.3 days for MODIS, 6.9 days for Sentinel-2, and over 21 days for Landsat. A hierarchical multisensor approach with 5-day gap-filling reduced this to just 1.3 days. On dates when cameras confirmed snow, satellites underestimated snow presence by 61.6% (Sentinel-2), 71.5% (Landsat), and 79.7% (MODIS), though gap-filling approaches reduced underestimation to 49.4%—deficits largely attributable to cloud-obscured scenes. When both satellite and camera provided cloud-free observations for the same date and location, classification agreement exceeded 85%. Despite this, satellites consistently failed to detect short-lived snow events and introduced temporal biases. On average, Snow Onset Dates were detected 13–52 days later, and Snow Melt-Out Dates differed by up to 40 days compared to camera-derived records. These results have implications for snow-cover monitoring using satellite images and highlight the need for integrating ground-based observations to compensate for satellite limitations and improve snow cover seasonality assessments in complex terrains.

1. Introduction

Monitoring snow cover in mountainous regions is essential for understanding hydrological cycles, water resource management, surface energy balance, ecosystem dynamics, and socioeconomic activities [1,2,3]. However, obtaining consistent and reliable snow cover data in these regions presents significant challenges due to complex topography, varying snow accumulation patterns, and atmospheric conditions that frequently obscure remote sensing observations [4,5].

Satellite remote sensing, a dominant tool for snow cover mapping, faces inherent limitations due to cloud cover obscuring surface observations, spatial resolution constraints dictated by sensor characteristics, and revisit times restricting the temporal frequency of data acquisition [6,7]. The presence of forest canopy further complicates the retrieval of snow cover information, as satellite sensors struggle to penetrate dense vegetation [8].

In this context, ground-based monitoring techniques, such as cameras including webcams and camera traps, have emerged as promising complementary tools for snow cover monitoring. These cameras can capture high-frequency imagery of snow cover conditions with fine spatial resolution. However, these methods are often sparse and do not offer broad spatial coverage [9]. Importantly, their implementation is particularly valuable in regions where there is no established meteorological station network dedicated to monitoring snowpack properties.

The use of time-lapse cameras thus arises as a practical alternative for obtaining detailed local observations of snow cover dynamics in areas lacking in situ instrumentation [10]. They can effectively estimate fractional snow cover, snow depth, and snow line elevation [10,11]. The high frequency of webcam imagery allows for near-real-time monitoring and can fill gaps in satellite data caused by cloud cover [11,12]. Webcams can capture snow dynamics under forest canopies, which is challenging for satellites [13]. Despite these advancements, integrating ground-based camera data with satellite imagery remains an underexplored avenue. However, webcams have limitations, including low spectral resolution and varying image quality due to uncontrolled radiometry and oblique viewpoints [14]. Although cameras provide high-resolution temporal and spatial data, they have limited spatial coverage and are susceptible to occlusions due to vegetation, terrain, and lighting variations [14].

Studies have demonstrated the utility of camera imagery in various related fields. For example, Ref. [15] created a webcam network and database for phenological studies, including snow cover. Time-lapse photography has been used to infer snow accumulation distribution [16], to estimate surface albedo using terrestrial photography [17], or even for lake ice monitoring [18,19]. Time-lapse cameras have also been used to measure snow accumulation and ablation dynamics during rain-on-snow events [18,20]. These camera-based observations offer advantages such as sub-meter spatial resolution, high temporal resolution (hourly or sub-hourly), and the ability to capture data under cloudy conditions that limit satellite-based measurements [21]. Several studies have explored the use of automated techniques to extract snow cover information from camera imagery. Ref. [8] applied an adaptive thresholding method to webcam images to obtain fractional snow cover information in northern boreal conditions. Ref. [22] used a camera network to observe snow processes and extract snow depth, albedo, and canopy interception. Ref. [23] evaluated Moderate-Resolution Imaging Spectroradiometer (MODIS) snow products globally using an extensive network of time-lapse cameras. Ref. [13] developed a semi-automatic procedure to retrieve snow accumulation and ablation in montane forests from time-lapse photography. Ref. [24] developed a method to estimate snow depth using virtual snow stakes in remote camera images. Ref. [10] developed a technique to produce high-resolution snow depth time series and used them for a case study of snow cover depletion in a High Arctic ecosystem. Ref. [25] used time-lapse cameras to monitor snow cover in Antarctica. The use of webcams and snow poles has enhanced validation of different physical parameterizations of the Weather Research and Forecasting Model (WRF) in the Cantabrian Mountains, where observational networks were previously sparse [26]. These studies highlight the potential of camera imagery for snow cover monitoring.

This study addresses this gap through a comprehensive 20-year comparative analysis of camera-derived snow observations with satellite data products in a mountainous region. Specifically, it aims to explore the differences in snow detection between both methods because of rapid changes in snow cover conditions, which may not be accurately detected using satellite imagery alone due to revisit periods or prolonged cloud cover episodes.

2. Data and Methods

2.1. Study Area

The study area focuses on the northern sector of the Iberian Peninsula, covering the Cantabrian Mountains and Sierra de la Demanda (Figure 1a). The study area spans over 400 km in length, and the geographical layout of both mountain ranges results in significant differences in snow cover during the same snowfall event [26], often leading to substantial contrasts between slopes or between the eastern and western sectors.

Snowfall is a key climatic feature in both the northern Iberian Range and the Cantabrian Mountains, with typical seasons spanning from November to May. The largest snow-covered extents generally occur between December and March, particularly in late January and February [27]. In both systems, spatial variability is controlled by elevation, topography, and exposure to dominant airflows. Sierra de la Demanda often exceeds 35 snow days annually above 1200 m [27], while in the Cantabrian Mountains, annual snow-cover days range from fewer than 6 below 500 m to over 130 above 2000 m, with values exceeding 300 days in shaded high-altitude areas [28]. Snowfall in both regions is linked to polar maritime advections under zonal and northwesterly flows, along with dynamic features like cut-off lows [29]. Orographic effects enhance accumulation and extend snow persistence, which in elevated zones may last into late spring and play a crucial role in regional hydrological regimes [6]. However, since 2000, the Cantabrian Mountains have shown a significant decline in snow duration, especially between 1500 and 2000 m, with losses surpassing −0.7 days/year [27].

Therefore, monitoring snow cover using satellite imagery, whether fully or partially covering the study area, is particularly useful in highlighting local variations. However, persistent cloud cover is common, often lasting for several consecutive days, with a mean duration of 3.3 days (for the period 2000–2023) according to a recent study [27]. This makes it necessary to explore new sources of ground-based information to avoid data gaps spanning multiple days, during which snow cover conditions can change rapidly within just a few hours.

2.2. Data

2.2.1. Time-Lapse Camera Images

A monitoring network has been established to track snow conditions using images from webcams and camera traps distributed throughout the study area. This network aims to ensure representative sampling across different altitudes (Figure 1b) and geographical extents (Figure 1a), while also accounting for variations between slopes. The selected snow cover monitoring sites are generally open areas, as far as possible from buildings and dense tree cover and are chosen to be representative of grassland environments.

In locations with available Wi-Fi, webcams have been installed with the participation of local communities, making this initiative a small-scale example of citizen science. In areas where internet access is unavailable, or in remote regions at higher altitudes beyond inhabited zones, camera traps have been deployed. These cameras are programmed to capture an image every six hours throughout the snow season, which begins in September and ends in August. The observation period recorded by the cameras covers from November 2003 to March 2025 (

Table 1. Detailed overview of camera image data.

Camera Type	Sampling Interval	Number of Sampling Sites	Number of Images (at 10:00)	Time Period Covered	Download Method
Public webcam	Every 6 h	23	92,755	2003–2025	SeqDownload v1.26 software (Automated)/Manual download
WiFi-Webcam	Every 6 h	4	4140	2021–2025	Manual download
Trap camera	Every 6 h	2	2517	2021–2025	Manual retrieval
TOTAL		29	99,412	2003–2025	-

).

Additionally, several public webcams have been used in this study, with programmed screenshots taken from their web addresses (URL) or IP cameras every six hours to ensure consistent time intervals across all three camera types. These images were automatically downloaded using the SeqDownload v1.26 software (https://www.nirsoft.net/utils/seqdownload.html, accessed on 1 August 2025). Images are captured at 04:00, 10:00, 16:00, and 22:00 (Spanish local time) to ensure at least two images with sunlight during the months with the shortest daylight duration. The 10:00 image is selected to represent the snow cover for that day, as it is the closest to the satellite overpass time (see table), which typically occurs between 10:45 and 12:30 Spanish local time.

2.2.2. Satellite Imagery

The image records captured by the cameras are compared with satellite images from the same dates within the 29 regions of interest (ROIs) chosen, allowing for a comparison of snow cover between both methods. The satellites used have different characteristics (

Table 2. Overview of satellite overpass times, revisit frequencies, and spatial resolutions.

Satellite	Local Overpass Time (CET/CEST)	Revisit Frequency	Data Availability	Spatial Resolution (m)	Number of Images
Sentinel-2A/2B	~11:30–12:30	Every 5 days (2–3 days combined)	2015 *–present (S2A)2017–present (S2B)	10–20	5553
Landsat 5	~10:45	Every 16 days	1984 **–2013	30–120	873
Landsat 8/9	~11:10	Every 16 days (8 days combined)	2013–present (L8), 2021–present (L9)	30–100	2243
MODIS (Terra)	~11:30	Daily	2000 **–present	250–500	9170
TOTAL					17,839

* Although Sentinel-2A acquisition started in 2015, the Harmonized Sentinel-2 MSI: MultiSpectral Instrument, Level-2A (Surface Reflectance) product available in Google Earth Engine (GEE) starts in March 2017. ** Although Landsat 5 acquisition began in 1984 and MODIS in 2000, only data from 2003 onwards were used to match the temporal range of available time-lapse camera records.

), as their spatial resolution varies, and not all of them provide continuous records throughout the entire period of interest.

2.3. Methods

2.3.1. Time-Lapse Cameras Snow Cover Extraction

To compare the area visible in the camera images with satellite records, one or more visible regions from the cameras were selected as ROI and compared with satellite data for the same location. The 29 ROIs (

Table A1. Geographic location and characteristics of the 29 regions of interest (ROIs) used in the study.

ID	Province	Mean Elevation (m)	Longitude (°)	Latitude (°)	Area (ha)
AS-01	Asturias	1544	−6.4154	42.9942	2.6
BU-01	Burgos	1111	−3.1770	42.1478	1.9
CA-02	Cantabria	1855	−4.8071	43.1551	3.5
CA-03	Cantabria	2083	−4.8195	43.1675	3.5
CA-04	Cantabria	2395	−4.8390	43.1801	2.9
LE-01	León	1242	−5.1783	43.0472	1.9
LE-02	León	1863	−5.1892	43.0368	3.5
LE-03	León	1528	−5.3797	43.0622	2.7
LE-04	León	1761	−5.3878	43.0399	2.7
LE-05	León	2002	−5.3990	43.0366	1.9
LE-06	León	1403	−4.7972	43.0516	0.8
LE-07	León	1852	−4.7874	43.0393	3.6
LE-08	León	2291	−4.7452	43.0300	3.3
LE-09	León	2089	−4.8694	43.1731	3.5
LE-10	León	2249	−4.8619	43.1744	3.1
LE-11	León	935	−4.9154	43.1504	2.8
LE-12	León	1027	−6.3152	42.9324	3.4
LE-13	León	1224	−5.4133	42.9739	0.2
LE-15	León	1789	−5.4199	42.9565	1.9
LR-01	La Rioja	1543	−2.8474	42.1744	0.8
LR-02	La Rioja	1713	−2.9665	42.2554	2.6
LR-03	La Rioja	2080	−2.9733	42.2458	2.4
LR-04	La Rioja	872	−3.0275	42.3234	1.8
LR-05	La Rioja	1637	−3.0784	42.2890	3.5
LU-01	Lugo	983	−7.1871	43.0321	2.7
LU-02	Lugo	1339	−6.9232	42.8202	3.4
LU-04	Lugo	1311	−7.0474	42.7074	3.2
OR-01	Ourense	1602	−6.7844	42.3749	0.7
PA-01	Palencia	2032	−4.3959	43.0345	2.1

) were manually delineated to ensure internal homogeneity, with areas ranging from 0.2 to 3.6 hectares and an average of 2.6 hectares. Each ROI was selected to represent relatively uniform snow distribution, avoiding zones with strong contrasts within the area. Images are manually processed and classified into four categories (

Table 3. Snow cover classification based on fractional snow cover thresholds in [6].

Category	Fractional Snow Cover (x)
Snow cover	x > 0.75
Partially snow cover	0.15 < x ≤ 0.75
No snow cover	x < 0.15
Not available	Lens obstruction, dense fog or camera offline

): total snow cover, partial snow cover, no snow cover, and image not available. Partial snow cover includes both patchy snow and thin snowfalls (<1 cm) that do not fully cover the grassland vegetation.

Camera observations may be interrupted during certain seasons for various reasons. The most common issue is the presence of dense fog. Additionally, lens obstruction by snow for several hours during wind-driven snowfalls may occur, as well as internet connection interruptions during severe snowfall events caused by power outages (only in webcams).

2.3.2. Satellite Snow Cover Retrieval Using Google Earth Engine (GEE)

The daily FSC (fractional snow cover) records derived from time-lapse cameras were compared with satellite-based observations. For this purpose, platforms such as Google Earth Engine [30] enable large-scale processing and extraction of satellite information without the need to individually download each satellite image.

In this study, Google Earth Engine enabled the processing of 17,839 satellite images (Table 2), resulting in 494,970 records across the 29 ROIs analyzed between November 2003 and March 2025. The satellite products used in GEE included the following: MODIS Terra Snow Cover Daily Global 500 m (MOD10A1.061); USGS Landsat 5 Level 2, Collection 2, Tier 1; USGS Landsat 8 Level 2, Collection 2, Tier 1; USGS Landsat 9 Level 2, Collection 2, Tier 1; Harmonized Sentinel-2 MSI Level 2A (Sentinel-2A and 2B).

Subsequently, snow classification was performed for each image based on sensor-specific methods. For Landsat and Sentinel-2, the Normalized Difference Snow Index (NDSI) was calculated using the green and shortwave infrared bands, applying a threshold of NDSI > 0.4 to identify snow-covered pixels [31,32,33]. For MODIS, the ‘NDSI_Snow_Cover’ band was converted to FSC following [34], using the snow-covered threshold defined in Table 3. Each classified pixel was assigned to one of four categories—snow-covered, no snow, or partially snow-covered—according to thresholds of Table 3.

Following the extraction of snow cover masks from each image, daily integration of all satellite products was performed. The differing revisit frequencies of these satellites meant that on certain days some ROIs had only one available observation, often from MODIS (the only sensor providing daily coverage), while on other days multiple satellite observations were available for the same location. This allowed prioritization of the best available spatial resolution, assigning priority to Sentinel-2 over Landsat 5/8/9, and finally MODIS. A hierarchical method was applied, as described in [27], whereby for each day and pixel, the observation with the highest assigned weight was selected, generating a daily ImageCollection with variable spatial resolution.

To reduce the number of days with missing data due to persistent cloud cover, a gap-filling procedure was applied, allowing for interpolation of up to 5 consecutive days. This approach considers both prior and subsequent valid observations (up to 3 days before and after) and gap-filling according to the following criteria (adapted from [35]):

• If both adjacent observations (before and after) indicate snow, the missing days are assigned as snow.
• If both indicate snow-free conditions, the missing days are assigned as snow-free.
• If there is disagreement (one snow-covered, one snow-free or partially snow-covered), simple linear interpolation of FSC or NDSI values is applied to estimate the most probable classification.

For each ROI, daily FSC values were calculated as the proportion (from 0 to 1) of snow-covered pixels relative to total valid (non-cloudy) observations and subsequently reclassified into four categories (no snow, partially snow-covered, snow-covered, and cloudy), according again to the thresholds defined in Table 3. Cloud pixels were identified and masked in Google Earth Engine using the quality assurance bands provided with each satellite product, which flag pixels affected by clouds and cloud shadows. This masking process enables the selective use of only cloud-free pixels within each image and allows for the exclusion of entire ROIs on days when cloud-covered pixels exceed 50% of the area.

Figure 2 shows an example of the pixel-based classification in an ROI derived from both the time-lapse camera images and the satellite products, illustrating the calculation of FSC using Sentinel-2 (Figure 2a), Landsat 8 (Figure 2b), and MODIS (Figure 2c).

2.3.3. Snow Detection Metrics and Comparison Framework

Daily snow classification records obtained from both time-lapse cameras and satellite imagery were integrated into a unified dataset in RStudio (v. 2025.5.0.496), which enabled a detailed comparison of snow detection performance between satellite products and ground-based observations.

Firstly, data availability was examined across the full observation period, considering both valid and cloudy records of time-lapse cameras and satellite products to assess temporal coverage continuity (Figure 3).

Classification frequencies were computed separately for coincident camera–satellite observations and for the subset of days with camera-confirmed snow presence, enabling assessment of how cloud cover and revisit frequency influence satellite-based snow detection. The corresponding snow detection rates for each satellite product are presented in Figure 4.

To quantify agreement between satellite and camera observations, all coincident valid records (with no clouds in satellite observations) were compared, yielding match rates and estimates of cloud-induced data loss for each sensor. Effective revisit intervals were also calculated, integrating the impact of cloud cover into nominal acquisition frequencies. The definitions of the metrics listed in Table 4 are as follows:

• Total acquisitions refers to the total number of satellite observations processed for each product across all ROIs, including records affected by cloud cover.
• Overlapping observations indicates the number of dates with simultaneous valid records from both satellites and cameras.
• Cloud cover rate (%) represents the proportion of satellite acquisitions excluded due to cloud obstruction.
• Cloud-free comparable records correspond to the remaining valid satellite observations available for direct comparison with camera data.
• Agreement cases indicate the number of coincident records where satellite and camera classifications match.
• Misclassification rate (%) is defined as 100 minus the ratio of agreement cases to overlapping observations. This metric reflects the overall rate of disagreement, accounting not only for misclassification of snow cover but also for days when snow was present but not detected due to cloud obstruction.
• Match rate (%) is the proportion of agreement cases over the total cloud-free comparable records (snow-covered–snow-covered, partially covered–partially covered, or no snow–no snow).
• Effective revisit interval (days) denotes the average number of days between two valid (cloud-free) observations, reflecting both sensor revisit frequency and cloud cover effects.

Cloud cover dynamics were further characterized by examining both the frequency and duration of consecutive cloudy periods, providing insight into the temporal limitations affecting snow monitoring (Figure 5 and Figure 6).

Finally, two key seasonal metrics were derived for each ROI and snow season: the Snow Onset Date (SOD) is defined as the first day of snow presence—starting from September 1st—and the Snow Melt-Out Date (SMOD) is defined as the first snow-free day following the last snow cover period of the season. These metrics were computed only for those locations and seasons (N seasons in Table 5) with highly complete records (>95% of daily observations available) and sufficiently long snow cover durations (>100 days per season), in order to ensure robust and representative comparisons. Differences between satellite- and camera-derived dates were computed to evaluate the temporal accuracy of each observation platform in capturing snow cover seasonality (Figure 7 and Figure 8).

3. Results

3.1. Satellite and Time-Lapse Cameras Data Availability

An essential prerequisite for a robust comparison between satellite-derived and camera-based snow cover observations is the temporal availability of both data sources. Figure 3 shows the temporal distribution of records from cameras and satellites across the 29 ROIs, along with additional records generated in Google Earth Engine following the methodology described in [27].

A total of 594,382 individual observations were obtained across the 29 ROIs for the period 2003 to 2025. In terms of data sources, 99,412 records were obtained from cameras; 225,214 from MODIS; 24,449 from Sentinel-2A/2B; and 19,136 from Landsat 5/8/9. Additionally, an improvement in data availability is observed in the product generated in GEE based on [27], which yielded 226,171 records by applying the pixel-level gap-filling algorithm.

As shown in Figure 3, the time-lapse camera network has provided nearly continuous ground-based observations since 2003 in several ROIs, covering full snow seasons with some discontinuous periods due to temporary interruptions in image acquisition. Unlike satellite observations, camera records are not affected by cloud cover, allowing for uninterrupted monitoring of snow presence and complementing satellite data. The MODIS platform provided the most continuous and extensive temporal coverage across the study period, with daily acquisitions from 2003 onwards. In contrast, Landsat satellites (5, 8, and 9) exhibited sparser temporal coverage due to their 16-day revisit interval, with significant data gaps during the 2011–2012 and 2012–2013 seasons caused by the exclusion of Landsat 7 imagery. A notable improvement in Landsat coverage occurred from 2021 onwards, when Landsat-8 and Landsat-9 began operating jointly, effectively reducing the revisit interval to 8 days and increasing the number of available observations. Sentinel-2 imagery has been available since mid-2015, progressively contributing higher-frequency acquisitions, particularly after the full operation of Sentinel-2B. Additionally, the gap-filled product generated in Google Earth Engine substantially improves data continuity by mitigating periods affected by persistent cloud cover, but since it relies on interpolation, it also presents certain limitations.

Figure 3. Temporal availability of daily snow cover classifications from time-lapse cameras, individual satellite sensors, and the GEE-derived gap-filled product (2000–2025).

Figure 3. Temporal availability of daily snow cover classifications from time-lapse cameras, individual satellite sensors, and the GEE-derived gap-filled product (2000–2025).

3.2. Snow Detection Frequencies for Comparable Camera-Satellite Records

Thus, while the distribution of satellite products remains variable compared to camera observations, the datasets are nonetheless comparable at the individual record level (via record joins). Based on this, satellite observations were evaluated whenever a corresponding camera record existed for the same date and location. The absolute number of observations (Figure 4a) exceeds 99,000 records for the sources with a daily revisit frequency (MODIS and time-lapse cameras), as well as for the product generated in GEE. In contrast, the number of Sentinel and Landsat observations matched with cameras is reduced to approximately 11,900 and 7800, respectively, due to their lower revisit frequency.

When considering relative frequencies (Figure 4b), a high proportion of invalid observations due to cloud cover is observed, particularly for MODIS (57.9%), followed by Sentinel-2A/2B (48.3%) and Landsat 5/8/9 (46%). In the case of the GEE-generated product, the percentage of invalid observations is reduced to 21.4%, thanks to the gap-filling algorithm (up to 5 days). This high cloud obstruction leads to an underestimation of the proportion of days classified as snow-covered (either partially or fully snow-covered). Overall, 35.5% of the camera observations recorded the presence of snow. For those same dates and locations, satellites consistently reported lower percentages, as a consequence of moderate cloud interference. For instance, Sentinel-2A/2B recorded 12% of observations with snow cover (representing a 65.2% reduction compared to camera data), Landsat 5/8/9 recorded 8.8% (75.3% reduction), MODIS recorded 7.4% (79.2% reduction), and the GEE product recorded 19.9% (43.8% underestimation).

A parallel analysis, considering only those dates where snow presence was confirmed by the cameras, shows a reduction in the absolute number of satellite observations across all sensors (Figure 4c), resulting, for instance, in ~35,000 observations where snow was present according to the cameras. In relative terms (Figure 4d), cloud cover percentages were consistently higher on these snow-confirmed days compared to the overall annual averages. For example, in Sentinel-2A/2B, cloud-covered observations increased to 51.1% on snow days (a rise of 2.8%). For Landsat 5/8/9, cloud cover increased to 56.1% (a rise of 10.1%), and for MODIS to 74.1% (a rise of 16.2%), reflecting increased cloudiness during snowfall events and subsequent days when snow cover persists.

Among the satellite observations corresponding to dates and locations where cameras confirmed snow presence, a portion was still classified as no snow, either due to detection mismatches or the lack of valid cloud-free acquisitions at the time of snow presence. This misclassification affected 5.6% of the snow-day observations in the case of MODIS and up to 15.4% for Landsat 5/8/9. Although snow detection rates increased considerably compared to the general assessment, none of the satellite products approached full detection, revealing substantial underestimation in actual snow cover classification. Specifically, Sentinel-2A/2B showed an underestimation of 61.6%, Landsat 5/8/9 of 71.5%, and MODIS of 79.7%. The GEE-generated product, supported by the gap-filling approach, reduced this underestimation to 49.4%.

Figure 4. Classification frequencies: (a) Absolute number of observations during the camera recording period. (b) Relative frequency of observations during the camera recording period. (c) Absolute number of observations on days with camera-confirmed snow presence. (d) Relative frequency of observations on days with camera-confirmed snow presence.

Figure 4. Classification frequencies: (a) Absolute number of observations during the camera recording period. (b) Relative frequency of observations during the camera recording period. (c) Absolute number of observations on days with camera-confirmed snow presence. (d) Relative frequency of observations on days with camera-confirmed snow presence.

3.3. Agreement and Complementarity Between Time-Lapse Camera and Satellite Snow Detection

The comparison of individual coincident records between the cameras and each satellite dataset reveals a consistently high degree of agreement (% matches). As presented in Table 4, the proportion of matching classifications (for example, snow-covered–snow-covered, partially covered–partially covered, or no snow–no snow between the camera and each satellite) exceeds 85% in all cases. Specifically, match rates reach 85.9% for MODIS, 88.6% for Landsat 5/8/9, and 91.1% for Sentinel 2A/2B, indicating that satellites with higher spatial resolution tend to yield more accurate classifications. The GEE-derived product, which incorporates a gap-filling procedure of up to 5 days, achieves a match rate of 89%. It also exhibits the lowest proportion of records discarded due to cloud cover, with only 20.9% of observations excluded. These excluded cases correspond to prolonged cloudy periods exceeding 5 days, which could not be filled and were thus retained as missing observations.

To complement the match rate, we also report the misclassification rate. This metric reflects the overall rate of disagreement, accounting not only for the misclassification of snow cover (e.g., false positives or negatives) but also for days when snow was present but not detected due to cloud obstruction. As shown in Table 4, the lowest misclassification rate was found in the hierarchical GEE product (30%), followed by Sentinel-2 (52.9%) and Landsat (52.1%), while MODIS exhibited the highest value (63.8%).

Across the satellite datasets, the proportion of cloud-affected observations ranges from 45.3% for Landsat to 56.6% for MODIS. The comparatively coarser spatial resolution of MODIS likely contributes to a higher proportion of cloud-related exclusions, as larger pixels are less suited to capturing the small-scale spatial variability of the ROIs under analysis.

The effective revisit period (defined here as the mean interval between valid, cloud-free observations) provides an integrated assessment of both the nominal satellite revisit cycles and the impact of cloud cover on data availability. This metric reveals relatively lower effective revisit times for MODIS, averaging 2.3 days, compared with 6.9 days for Sentinel 2A/2B. The GEE product, by integrating data from all three satellite sources and applying a gap-filling approach, substantially reduces the effective revisit period to an average of just 1.3 days. The effective revisit interval is 21.4 days for Landsat 5/8/9. It is worth noting that the inclusion of Landsat 9, operational since October 2021, complements the acquisitions from Landsat 8 and contributes to improve temporal coverage. If we differentiate both Landsat operational periods, the effective revisit period reached 27.1 days during the earlier phase when Landsat 5 and Landsat 8 operated independently. Since the joint operation of Landsat 8 and Landsat 9, the effective revisit period has improved significantly, reducing to 10.2 days. The GEE product, by integrating data from all three satellite sources and applying a gap-filling approach, substantially reduces the effective revisit period to an average of just 1.3 days (Table 4).

Table 4. Summary of satellite snow cover observations and their agreement with time-lapse camera records.

Source	Total Acquisitions	Overlapping Observations	Cloud Cover Rate (%)	Cloud-Free Comparable Records	Agreement Cases	Misclassification Rate (%)	Match Rate (%)	Effective Revisit Interval (Days)
Hierarchical (Sentinel 2 > Landsat > MODIS) + Gap filling 5 Days [27]	226,171	99,412	20.9	78,130	69,540	30	89	1.3
Landsat 5/8/9	19,136	7764	45.3	4193	3717	52.1	88.6	21.4
Landsat 5 and Landsat 8 (before 31 October 2021)								27.1
Landsat 8/9 (after 31 October 2021)								10.2
MODIS	225,214	99,082	56.6	41,688	35,834	63.8	85.9	2.3
Sentinel 2A/2B	24,449	11,876	50.2	6140	5594	52.9	91.1	6.9
Time-lapse cameras	99,412	99,412	-	-	-		-	-

Table 4. Summary of satellite snow cover observations and their agreement with time-lapse camera records.

Source	Total Acquisitions	Overlapping Observations	Cloud Cover Rate (%)	Cloud-Free Comparable Records	Agreement Cases	Misclassification Rate (%)	Match Rate (%)	Effective Revisit Interval (Days)
Hierarchical (Sentinel 2 > Landsat > MODIS) + Gap filling 5 Days [27]	226,171	99,412	20.9	78,130	69,540	30	89	1.3
Landsat 5/8/9	19,136	7764	45.3	4193	3717	52.1	88.6	21.4
Landsat 5 and Landsat 8 (before 31 October 2021)								27.1
Landsat 8/9 (after 31 October 2021)								10.2
MODIS	225,214	99,082	56.6	41,688	35,834	63.8	85.9	2.3
Sentinel 2A/2B	24,449	11,876	50.2	6140	5594	52.9	91.1	6.9
Time-lapse cameras	99,412	99,412	-	-	-		-	-

3.4. Cloud-Induced Observation Gaps

Cloud cover is a major limitation in the use of optical satellite imagery, particularly in mountain environments. In this study, cloud cover dynamics were further analyzed on a monthly scale (Figure 5). The seasonal pattern highlights the highest frequency of cloudy days occurring from November to May, according to MODIS records—the only product with daily observations—exceeding 60% of invalid observations during this period due to cloudiness. This coincides with the peak snow accumulation season in the study area. The percentages identified with Sentinel-2A/2B and Landsat 5/8/9 are similar during this period, ranging between 40% and 60%, considering that they cover different time spans (Sentinel data are only available since 2017) and that Landsat has a longer revisit interval. The integration of these satellites into the GEE-derived product, which applies a gap-filling method of up to 5 days per pixel, allows these values to be halved, resulting in a data loss of only ~25–30% due to cloud cover from November to May.

Figure 5. Monthly cloud cover patterns and observation gaps across the ROIs. (a) Percentage of days affected by cloud cover. (b) Mean duration of cloudy periods for each satellite sensor, calculated as the average number of consecutive days between the first cloud-covered observation and the next valid (cloud-free) acquisition.

Figure 5. Monthly cloud cover patterns and observation gaps across the ROIs. (a) Percentage of days affected by cloud cover. (b) Mean duration of cloudy periods for each satellite sensor, calculated as the average number of consecutive days between the first cloud-covered observation and the next valid (cloud-free) acquisition.

Cloudiness decreases considerably in the regions of interest during summer, remaining around 30%. Regarding the average consecutive duration of cloudy events—measured as the period between two valid observations for each satellite—MODIS and Sentinel show similarly low mean values of 2.6 and 4.3 days, respectively. The longer revisit interval of Landsat results in higher values, with an average of 11.6 days, peaking in November at 18.9 days, and remaining around 12 days during the rest of the main snow season. The GEE gap-filling product reduces this interval to an average of just 2.8 days.

The cumulative number of observations (Figure 6) highlights how satellite data availability is shaped by cloud cover patterns. Cloud-covered periods were characterized based on their duration for each satellite, defined by the time between two valid (cloud-free) observations. The MODIS curve appears as a nearly perfect and smooth cumulative distribution, in contrast to the more stepped profiles observed for the other satellite products. This is due to the high temporal resolution of MODIS, with near-daily acquisitions, which generates a dense and continuous distribution of observation intervals. In contrast, sensors with a lower revisit frequency or more sporadic data availability result in more discrete and irregular patterns. A notable discontinuity is observed in the GEE-derived product, associated with the 5-day gap-filling period applied during processing.

By definition, all analyzed events (gaps) consist of at least one consecutive day without valid observations; therefore, 100% of cases correspond to cloudy periods of at least one day. As the length of these consecutive cloudy periods increases, the cumulative proportion of cases decreases progressively. For periods lasting at least 2 consecutive days without valid observations, the proportions drop to 71% for the GEE product, 62% for MODIS, 53% for Sentinel-2, and 51% for Landsat 5/8/9. From approximately 3 consecutive cloudy days onwards, the differences between satellite products become increasingly marked. The GEE product, benefitting from its gap-filling strategy, maintains better temporal continuity, while lower-revisit-frequency products are progressively more affected. For gaps of 5 or more consecutive days, the proportions fall to 19% (GEE), 25% (MODIS), 37% (Sentinel-2), and 47% (Landsat).

Extended gaps of 20 or more days are rare for all products except Landsat, where they still account for 16% of the cases. This is directly related to Landsat’s longer revisit time. Remarkably, for Landsat, the cumulative proportion of cases drops below 5% only after 49 consecutive cloudy days. The most extreme case documented in the dataset shows an observational gap of 185 consecutive days without a single valid Landsat image in one ROI.

Figure 6. Cumulative distribution of cloud and revisit-period-related gaps, expressed as the percentage of events with a given number of consecutive days until the next usable observation.

Figure 6. Cumulative distribution of cloud and revisit-period-related gaps, expressed as the percentage of events with a given number of consecutive days until the next usable observation.

3.5. Snow Cover Seasonality Characterization

The ability to accurately characterize snow cover seasonality depends strongly on the temporal continuity and observational reliability of satellite products. In this context, we compared the SOD and SMOD derived from different satellite sensors against reference observations obtained from ground-based cameras.

A comparative summary of the differences between satellite-derived and camera-based snow cover dates is presented in Table 5. For the SOD, the combined product of GEE shows the lowest mean difference (13.7 days) and the smallest maximum offset (67 days), indicating better alignment with ground-based observations. MODIS performs moderately well (mean difference of 19.8 days), followed by Landsat 5/8/9 (21.1 days), while Landsat 5/8/9 exhibits the largest average difference (51.6 days). This is partly explained by its longer revisit interval.

Table 5. Differences in SOD and SMOD between satellite-derived estimates and camera-based observations for each satellite product.

	Product/Satellite	Min Diff. (in Days)	Mean Diff. (in Days)	Mean Absolute Difference ∣Δ∣ (Days)	Max Diff. (in Days)	N (Number of Seasons)
SOD differences (in days)	Hierarchical (Sentinel 2 > Landsat > MODIS) + Gap filling 5 Days [27]	−2	13.7	13.8	67	93
	MODIS	0	19.8	19.8	78	100
	Landsat 5/8/9	2	51.6	51.6	95	59
	Sentinel 2A/2B	0	21.1	21.1	81	27
SMOD differences (in days)	Hierarchical (Sentinel 2 > Landsat > MODIS) + Gap filling 5 Days [27]	−70	−14.4	21.4	64	90
	MODIS	−73	−23.4	25.8	61	98
	Landsat 5/8/9	−98	−38.5	40.7	53	55
	Sentinel 2A/2B	−69	−12.1	23	37	25

Table 5. Differences in SOD and SMOD between satellite-derived estimates and camera-based observations for each satellite product.

	Product/Satellite	Min Diff. (in Days)	Mean Diff. (in Days)	Mean Absolute Difference ∣Δ∣ (Days)	Max Diff. (in Days)	N (Number of Seasons)
SOD differences (in days)	Hierarchical (Sentinel 2 > Landsat > MODIS) + Gap filling 5 Days [27]	−2	13.7	13.8	67	93
	MODIS	0	19.8	19.8	78	100
	Landsat 5/8/9	2	51.6	51.6	95	59
	Sentinel 2A/2B	0	21.1	21.1	81	27
SMOD differences (in days)	Hierarchical (Sentinel 2 > Landsat > MODIS) + Gap filling 5 Days [27]	−70	−14.4	21.4	64	90
	MODIS	−73	−23.4	25.8	61	98
	Landsat 5/8/9	−98	−38.5	40.7	53	55
	Sentinel 2A/2B	−69	−12.1	23	37	25

The largest SOD differences occur in cases where the camera detects a brief early-season snowfall episode that is not captured by the satellites, either due to revisit limitations or cloud cover. Instead, the satellites typically detect the second, more persistent snowfall event several weeks later, once conditions allow. As a result, all mean differences in SOD are positive, since the first snowfall is always detected earlier by the camera. A few negative values (−1 and −2 days) are present only in the combined product of GEE, caused by the pixel-level gap-filling which can anticipate the detection by up to 5 days in the absence of direct observations.

In the case of the SMOD, differences are more variable and can be both negative and positive. Negative differences occur when the satellite detects the first snow-free day after the last snow event, but subsequent short snowfall episodes—missed by the satellite due to cloudiness or revisiting gaps—are still captured by the camera, extending the snow season. Positive differences, on the other hand, arise when the snow has already melted according to the camera, but persistent cloud cover prevents satellites from recording the melt until the next cloud-free and snow-free observation becomes available.

The absolute values of the mean differences show that products with higher temporal and spatial resolution tend to better match the reference dates, while those with longer revisit times (such as Landsat) show larger discrepancies. This supports the use of multisensor approaches to improve temporal consistency in snow season monitoring.

As illustrated in Figure 7 and Figure 8, MODIS stands out for providing a temporally dense and consistent sequence of detections, resulting in smooth and nearly complete timelines for both SOD and SMOD. In contrast, Landsat 5/8/9 and Sentinel-2A/B exhibit more fragmented and discontinuous detection patterns, with longer observational gaps and reduced capacity to identify key seasonal transitions, particularly in periods with frequent cloud cover. The reduced number of observations available for Sentinel-2 is also explained by its later launch and operational availability, which limits its time series to seasons starting from 2017 onwards.

The combined SMOD and SOD analyses confirm that snow seasonality derived solely from satellite records is sensitive to cloud-related artefacts and revisit constraints. The integration of camera-based observations provides critical validation and allows more accurate reconstruction of snow cover dynamics at local scales.

Figure 7. Differences between satellite-derived and camera-derived SOD across all locations and seasons.

Figure 7. Differences between satellite-derived and camera-derived SOD across all locations and seasons.

Figure 8. Differences between satellite-derived and camera-derived SMOD across all locations and seasons.

Figure 8. Differences between satellite-derived and camera-derived SMOD across all locations and seasons.

4. Discussion

The results of this study provide valuable insights into the comparative capabilities and limitations of satellite-derived and ground-based camera observations in monitoring snow cover dynamics across a mid-latitude mountain region. Through an extensive dataset spanning over two decades and encompassing more than half a million individual records, we have been able to identify systematic differences in snow detection rates, spatial–temporal coverage, and the influence of cloud cover among different satellite platforms. These discrepancies are not only of methodological relevance but also carry significant implications for the accurate characterization of snow seasonality metrics. The following discussion is structured to critically assess these findings by examining the strengths and weaknesses of the different observation sources, quantifying snow cover underestimation due to clouds, evaluating the impact of missing observations on snow variability indicators, and considering the broader relevance of these outcomes in the context of snow monitoring efforts worldwide.

4.1. Strengths and Limitations of Optical Satellite Imagery and Time-Lapse Cameras

The characterization of snow cover in mountain environments depends on the availability, continuity, and quality of the observations. MODIS clearly stands out for offering the densest and most continuous record among the satellite products used in this work. Its daily revisit time and long temporal archive allow for more complete timelines, especially for detecting seasonal patterns. However, its coarse spatial resolution (500 m) poses a relevant limitation in mountainous environments [33], where snow cover often shows high spatial heterogeneity and where subpixel variations can be substantial. Previous studies have emphasized that snow cover patterns in mountain environments typically fluctuate over distances of less than 100 m, which implies that spatial resolutions above this threshold may fail to capture important subgrid variability [36]. To overcome the spatial limitations of MODIS data in mountainous terrain, strategies as spatial downscaling have been proposed in recent years. One effective approach involves the probabilistic redistribution of MODIS snow cover using higher-resolution Sentinel-2 data. Ref. [6] demonstrated that applying a Sentinel-2-based snow probability model significantly improves the spatial resolution of MODIS snow-derived products in the Iberian Peninsula.

Landsat sensors (5, 8, 9) offer improved spatial resolution (30 m) and a relatively long historical archive, with a 16-day revisit time that can be reduced to 8 days by combining data from Landsat 8 and 9 [37]. Sentinel-2 (10–20 m) provides even higher spatial detail and has the potential for a 5-day revisit (or 2–3 days when both Sentinel-2A and 2B are considered), although its temporal coverage is limited, as it has only been available since 2015. Notably, the nominal revisit times often overestimate the true observation frequency in cloud-prone regions such as the Cantabrian Mountains and Sierra de la Demanda. In our study, the effective revisit time, defined as the average number of days between valid (cloud-free) snow observations, revealed significant disparities between sensors: MODIS achieved an effective revisit period of 2.3 days, Sentinel-2A/2B of 6.9 days, and Landsat 5/8/9 of 21.4 days. These values highlight that, despite its coarser resolution, MODIS remains the most temporally reliable source in persistently cloudy regions. The best performance overall was obtained by the combining all satellites using a multisensor product with 5-day gap-filling [27], which reduced the effective revisit interval to just 1.3 days while maintaining a high match rate (89%) between the product and camera records; that is, only considering days when both sources provided valid data.

By contrast, time-lapse cameras offer near-continuous daily (or sub-daily) snow cover observations [10], unaffected by satellite-specific constraints such as orbital revisit cycles or cloud cover. Their high temporal resolution allows the capture of ephemeral snowfall events. They have proven effective in validating the presence and duration of snow cover in remote and cloudy regions [9,10].

Moreover, time-lapse imagery enables the retrieval of not only snow presence [8,12] but also related features such as snow depth [10], spatio-temporal variability of vegetation and snow cover [15], or even regional snowline elevation [11]. Cameras can support operational snow monitoring networks by providing low-cost and flexible systems [22], particularly in data-sparse or difficult-to-access regions [9,38] or even at the continental scale [39]. However, limitations persist. These systems generally have restricted spatial representativeness, as their field of view is limited and fixed. Environmental conditions such as fog, low light during polar night in high latitudes, or the presence of snowflakes and frost on the lens can degrade image quality or interrupt data acquisition [10]. Additionally, technical issues such as power outages, battery failures, or connectivity interruptions may lead to temporal gaps in the record.

Overall, time-lapse cameras complement satellite records by providing fine-scale, ground-based validation data, especially valuable for studies in high-latitude or environments with steep topography, where cloud cover or coarse resolution may compromise satellite-based snow monitoring.

4.2. Matching Records and Undetected Snow Cover Using Optical Satellite Data

The analysis of coincident records between satellite imagery and camera observations allows a more objective assessment of satellite performance under real conditions. When comparing only dates on which both methods provided valid observations, we found that a large proportion (79%) of snow-covered scenes visible in the camera images were not classified as “snow” in the satellite products. This misdetection is primarily related to the frequent cloud cover present during and after snowfalls, which limits surface visibility in all optical satellite systems. Although the proportion of cloudy images is similar across all sensors, differences in spatial resolution influence their ability to correctly classify snow-covered and partially snow-covered pixels. In this sense, MODIS, with its 500 m resolution, shows the highest rate of snow underdetection (79.2%), followed by Landsat 5/8/9 (71.3%) and Sentinel-2A/2B (65.2%). These values reflect a substantial underrepresentation of snow cover, which can be particularly critical in regions with frequent or prolonged cloudiness. The use of sensors with higher spatial resolution, such as Sentinel-2, helps to mitigate this issue by improving snow detection under partially clear-sky conditions. Similar spatial limitations have been reported even for very-high-resolution optical satellite platforms when mapping snow depth in complex alpine terrain [40,41]. Findings have been reported by [42], who demonstrated that even high-resolution Sentinel-2 imagery may face detection limitations in mountainous regions under persistent cloud cover and patchy snow conditions.

In contrast, the GEE-derived hierarchical product, which integrates Sentinel-2, Landsat, and MODIS observations and applies a 5-day temporal gap-filling strategy [27], shows a notably lower underestimation rate of 43.8%. This indicates that combining data sources with a rule-based priority scheme, and allowing short temporal interpolation, can significantly improve snow detection, especially in fragmented or cloudy periods. This methodological enhancement is crucial for achieving more continuous and reliable records in snow-covered mountain regions.

Consistent with the underestimation rates, the hierarchical GEE product exhibits the lowest misclassification rate, highlighting its improved accuracy through multisensor integration and temporal gap-filling. Conversely, MODIS shows the highest misclassification rate, attributed to its coarser spatial resolution and challenges detecting heterogeneous snow patterns.

Focusing on those days when snow presence on the ground was confirmed by the cameras (Figure 4c,d), an increase in the proportion of cloudy observations is evident across all satellite products relative to general conditions. This is an expected outcome, as snowfall events are often followed by several days of persistent cloud cover, which prevents optical satellites from acquiring valid surface observations. Although the proportion of snow-covered pixels increases during these periods relative to overall averages, it still remains well below full detection due to cloud obstruction. In these snow-confirmed records, substantial underestimations were observed: MODIS failed to detect snow cover in 79.7% of cases, Landsat 5/8/9 in 71.5%, and Sentinel-2A/2B in 61.6%. These figures illustrate the significant limitations of optical satellite systems to fully capture snow cover, even when snow is known to be present. The GEE-derived gap-filled product substantially improved detection, reducing underestimation to 49.4%, yet considerable gaps persist under prolonged cloud cover conditions. The benefits of multisensor integration for improving snow detection accuracy have also been demonstrated for operational products such as the Interactive Multisensor Snow and Ice Mapping System (IMS) 1 km [43].

Such underrepresentation of snow is consistent with findings from other studies. Ref. [13] reported similar patterns in montane forests, where even relatively deep snow cover went undetected due to canopy or cloud interference. Similar underestimation issues have also been reported for satellite-based snowfall retrievals in mountainous regions, with biases reaching up to 65% compared to ground observations [44].

These mismatches and limitations may have significant implications for the reliability of snow cover fraction estimates derived solely from satellite data, particularly when used in hydrological or climatic models.

4.3. Cloud Cover as a Limiting Factor for Satellite-Based Snow Detection

Our analysis of satellite records across the Cantabrian Mountains and Sierra de la Demanda shows that the periods with snow cover—as confirmed by camera observations—are also those with the highest cloud frequency, particularly between November and May, ranging from 40% to 60% of cloudy days, depending on the area. The effect of persistent cloudiness extends beyond the number of valid observations—it directly influences the temporal continuity and quality of snow cover records. For instance, the average duration between two valid satellite acquisitions reveals important differences across platforms: while MODIS and Sentinel-2 show relatively short average gaps (2.6 and 4.3 days, respectively), Landsat imagery exhibits much longer average observational gaps (~11.6 days), with peaks of up to 18.9 days during November. These long gaps are particularly problematic during the snow accumulation and melt periods, when snow cover can vary rapidly.

Our cumulative gap analysis further illustrates how cloud cover modulates data availability. For instance, ≥5-day gaps occur in 19% of cases for that GEE-derived product, compared to 25% for MODIS, 37% for Sentinel-2, and 47% for Landsat. These findings reinforce the broader conclusion that persistent cloud cover poses a substantial and recurrent challenge for snow monitoring using optical satellite data. This effect has also been observed in vegetation studies: Ref. [45] showed that Landsat-based greening trends in alpine environments were inflated due to declining visibility in snow-covered months, leading to an artificial extension of the growing season. A similar mechanism could lead to snow cover trends being underestimated in satellite products due to limited observation opportunities during peak snow periods.

These limitations underline the importance of using high-temporal-frequency products and incorporating complementary strategies—such as gap-filling algorithms, fusion of multi-source satellite data, and ground-based validation tools like time-lapse cameras—to ensure more complete and unbiased snow cover variability records. Additionally, the use of satellite platforms with afternoon overpass times, such as Aqua MODIS or Suomi-NPP VIIRS, can be useful in regions frequently affected by morning fog or low stratus clouds, as they may capture snow conditions once these layers dissipate. However, in our study area, where cloud cover is typically associated with persistent synoptic systems, their inclusion would not significantly improve data availability.

4.4. Implications for Snow Seasonality

Accurately characterizing snow seasonality in mountainous environments requires not only spatially representative data, but also high temporal continuity and observation reliability. Our results demonstrate that satellite-derived seasonal metrics such as the SOD and SMOD differ substantially from those obtained using time-lapse camera observations. These discrepancies highlight the limitations of current satellite products—especially when observations are interrupted by prolonged cloud cover or constrained by long revisit intervals.

For the SOD, all satellite products showed a positive bias compared to the camera data, indicating a consistent delay in the detection of the first snow event of the season. The mean delays ranged from 13.7 days in the GEE-derived hierarchical product to over 50 days in Landsat 5/8/9. The largest differences occurred in early-season situations where the camera detected ephemeral snow covers that were missed by the satellites due to cloud cover or insufficient acquisition frequency. This effect is particularly relevant in our study area, where the first snowfall events—typically occurring in October or November—are brief but often cover the entire ROI. These early snowfalls are well recorded by the cameras but are frequently missed by satellites due to the lack of valid observations caused by persistent cloudiness. Since the next snowfall may occur weeks or months later, the first snow detected by satellites often corresponds to a later and more persistent event. In such cases, spatial resolution is less critical than revisit frequency, which explains why MODIS, despite its coarse resolution, performs relatively well in detecting SOD.

SMOD results were more variable and often dependent on cloud dynamics during spring. In several cases, satellites anticipated the melt-out date compared to the camera, particularly when short snowfall episodes late in the season were obscured by cloud cover or occurred between acquisitions. This led to an underestimation of the snow season duration. In addition, the spatial resolution of the satellite sensor plays a key role at this stage: small snow patches, which are still visible in camera imagery, may not be captured by coarser-resolution sensors like MODIS, resulting in a premature estimation of melt-out. Although sensors like Sentinel-2 and Landsat offer improved spatial detail, they are still limited by revisit frequency. Conversely, positive differences in SMOD—when the satellite detects the first snow-free day after the ground-based record shows melt-out—are generally linked to persistent clouds that prevented timely snow disappearance detection from space. These mixed patterns were more frequent in products with lower temporal resolution, especially Landsat, which showed the largest absolute differences (mean |Δ| = 40.7 days).

As illustrated in Figure 7 and Figure 8, MODIS provided the most consistent temporal records, resulting in smooth SOD and SMOD timelines. This advantage stems from its near-daily acquisitions, which reduce the chance of missing key transitions. In contrast, Sentinel-2 and Landsat, though offering higher spatial resolution, displayed more fragmented and discontinuous timelines due to their sparser temporal sampling and increased cloud-related data loss.

Importantly, our results underscore how these mismatches in SOD and SMOD can translate into critical biases when estimating snow cover duration and its trends over time. An error of several weeks in identifying the start or end of the snow season could significantly distort conclusions about hydrological timing or vegetation phenology. Similar pixel-level uncertainties in SMOD estimates have been reported in the French Alps and Pyrenees even in long-term multisensor analyses combining high-resolution satellite records, with RMSE values around 28 days [46]. This reinforces the need for caution when interpreting satellite-derived phenological metrics in isolation and advocates for their validation using complementary ground-based observations.

5. Conclusions

This study assesses the performance of optical satellite products in detecting snow cover dynamics in mid-latitude mountain environments by using time-lapse cameras as a reference. By integrating ground-based and remotely sensed observations, we critically examine the timing and reliability of snow seasonality metrics such as the Snow Onset Date (SOD) and Snow Melt-Out Date (SMOD). The main conclusions are as follows:

• Cloud cover is the most critical constraint in satellite-based snow monitoring, particularly during accumulation and melt periods. It leads to substantial observation gaps and seasonality biases, especially in satellites with low revisit frequencies.
• Time-lapse cameras consistently detect early and short snow cover events missed by satellites. Their high temporal resolution makes them especially valuable for capturing ephemeral snowfall, verifying snow persistence, and identifying snowmelt transitions.
• Satellite-derived snow cover is systematically underestimated when snow presence is confirmed by cameras. During confirmed snow cover periods, underestimation reached 61.6% for Sentinel-2, 71.5% for Landsat, and up to 79.7% for MODIS, highlighting the extent of snow detection gaps under cloudy conditions. The gap-filled multisensor product significantly reduced these mismatches, lowering underestimation to 49.4%.
• Each method has limitations: optical satellites observations are affected by cloud-induced data gaps and spatial resolution constraints, while cameras face issues such as limited spatial coverage.
• Integrated approaches that fuse satellite data offer the most robust solution for monitoring snow dynamics. These strategies enhance temporal continuity, reduce uncertainties in snow season metrics, and are essential for snow cover monitoring.

Future efforts should promote the deployment of camera networks, automate snow detection from imagery, and operationalize the fusion of satellite and near-surface datasets to enhance real-time snow monitoring in complex terrains.