In Greece, winter precipitation exhibits pronounced temporal and spatial variability due to the interaction between synoptic circulation patterns and complex topography. The Pindus Mountains form a major climatic divide, separating the wetter western mainland from the drier eastern regions. During meridional circulation, cold northeasterly flows – often triggered by Siberian or central European blocking anticyclones – bring precipitation to eastern Greece, particularly in mid to late winter (Bartzokas et al., 2003). Conversely, zonal circulation dominates in the west, producing longer and more frequent precipitation events, with limited spillover to the eastern Pindus (Bartzokas et al., 2003). Analysis of surface station data from 2010–2023, indicates that heavy winter precipitation events (>20 mmd-1) occur most frequently over western Greece (Kotroni et al., 2025). Snow cover typically persists from November to May, with isolated patches lasting into mid-summer (Masloumidis et al., 2025).

The study area was delineated using the hierarchical inventory of the world's mountains, Global Mountain Biodiversity Assessment (GMBA) v2 product (Snethlage et al., 2022a, b). The dataset was clipped to the Greek national borders, and ten mountain massifs across the mainland with maximum elevations >2000 m a.s.l. were selected for snow-cover reconstruction and climatological analysis (Fig. 1).

Water was masked using the Joint Research Centre's Global Surface Water dataset v1.4 (Pekel et al., 2016), based on the maximum extent observed over the dataset period, 1984–2022. Snow cover detection over forested areas presents a known challenge for optical remote sensing (Gascoin et al., 2024; Muhuri et al., 2021), and therefore a forest mask was applied using the 2015 Tree Cover Density 100 m dataset (European Environment Agency, 2018). Grid cells were masked when tree cover density was ≥50 % (Barrou Dumont et al., 2025). Lastly, an elevation mask was applied, discarding areas lower than 700 m a.s.l., based on the Shuttle Radar Topography Mission (SRTM) digital elevation dataset (DEM) v3 (Farr et al., 2007).

snowMapper is an iteratively-solved, PIML-driven model developed in this study for reconstructing daily snow cover (Fig. 2). The model is trained on in situ data, forced by reanalysis-derived meteorological conditions, and updated through direct insertion of binary snow cover from available high-resolution satellite imagery. Its modular structure comprises three main steps (Fig. 2):

Step 1: Preprocessing. This includes the preparation of (a) satellite imagery (Sect. 2.3.1), (b) meteorological forcing data (Sect. 2.3.2), (c) terrain data (Sect. 2.3.3), and (d) in situ training data (Sect. 2.3.4), followed by (e) training of the machine learning classifier (Sect. 2.3.5).

The entire process is performed using the cloud-based resources of Google Earth Engine through the Python API (Gorelick et al., 2017). The model input data and preprocessing, model development and snow reconstruction, and model validation and direct insertion of observations are detailed in the following subsections.

After reconstructing each day's snow cover, but prior to the direct insertion of available Landsat- or Sentinel-derived observations, model performance was evaluated at the grid-cell-level against these clear-sky satellite observations on a day-to-day basis (Fig. 2). All available observations were used for validation, as they remained independent at this step and were only replaced after evaluation, following the process described in Sect. 2.4.3. Cells were classified as “true positive” (tp), “true negative” (tn), “false positive” (fp), or “false negative” (fn), based on the conditions outlined in Eq. (10): 10Evaluationdi=tp,if SCdiobserved=1 and SCdisimulated=1tn,if SCdiobserved=0 and SCdisimulated=0fp,if SCdiobserved=0 and SCdisimulated=1fn,if SCdiobserved=1 and SCdisimulated=0 At the end of the model run, we cumulate these daily grid cell-level evaluations across each month and massif, and calculate monthly performance metrics: accuracy (Eq. 11), snow overestimation (Eq. 12), and snow underestimation (Eq. 13) (Parajka and Blöschl, 2008): 11Accuracy(m,y)=∑i=1N(m,y)(tp(i,m,y)+tn(i,m,y))∑i=1N(m,y)(tp(i,m,y)+tn(i,m,y)+fp(i,m,y)+fn(i,m,y))12Overestimation(m,y)=∑i=1N(m,y)(fp(i,m,y))∑i=1N(m,y)(tp(i,m,y)+tn(i,m,y)+fp(i,m,y)+fn(i,m,y))13Underestimation(m,y)=∑i=1N(m,y)(fn(i,m,y))∑i=1N(m,y)(tp(i,m,y)+tn(i,m,y)+fp(i,m,y)+fn(i,m,y)) We bias-corrected the reconstructed monthly FSC values using an approach inspired by the trend-preserving delta-change method (Matiu and Hanzer, 2022). First, we calculated the representativeness index (RI) (Koehler et al., 2022), which quantifies the percentage of available observations in a given month for a given GMBA mountain area relative to the total number of grid cells in that month and area (Eq. 14): 14RI(m,y)=N(m,y)obsN(m,y)all Second, we removed all instances where the RI<0.015, as such a low proportion of observations was considered insufficient for reliable validation and subsequent bias correction. This threshold corresponds roughly to having a clear-sky view of half the area's grid cells on a single day within a month. Third, for each massif over the study period, we calculated the monthly mean underestimation (U‾m) and overestimation (O‾m) metrics. Fourth, these values were then used to correct the reconstructed FSC according to Eq. (15): 15FSC(m,y)corrected=FSC(m,y)+min⁡(U(m,y),U‾m)-min⁡(O(m,y),O‾m) Finally, we converted the corrected monthly FSC values to monthly snow-covered area (SCA) by multiplying each value by the masked area (km2) of the respective massif (Sect. 2.1).

We performed climatological analyses both across the entire study area and at the massif scale.

Ten-year moving averages of the bias-corrected SCA time-series were calculated for each calendar month of the snow season (November–May) to smooth interannual variability. For each month, we applied the Mann-Kendall test to identify trends and we estimated their slope using the Theil-Sen method, both for individual massifs and for all 10 massifs collectively. To facilitate inter-month comparison Sen's slopes were normalised by dividing by the mean SCA for that month, yielding the fractional annual change in SCA relative to the mean snow-covered area rather than to the total reconstructed area of each massif.

We calculated SCA anomalies relative to the 1984–2025 monthly means and identified extreme high and low snow-cover events based on the monthly 80th and 20th percentile thresholds. Extremes were grouped by snow season, and aggregated as means for magnitude and as counts for frequency. The autocorrelation-sensitive modified Mann-Kendall test was then applied (Hamed and Ramachandra Rao, 1998) to identify trends in both (Li et al., 2025).

Using the unsmoothed SCA time-series for each month, we examined correlations with climate and climate variability by performing Pearson's tests against: (a) mean air temperature, (b) cumulative precipitation, (c) North Atlantic Oscillation (NAO) index, (d) Arctic Oscillation (AO) index, and (e) Atlantic Multidecadal Oscillation (AMO) index.

Finally, each month's unsmoothed SCA time-series was split into two partially overlapping 21 year subperiods: 1984/85–2004/05 and 2004/05–2024/25. For each subperiod, we calculated the mean and standard deviation and applied Levene's test to assess differences in SCA variability.

For all statistical tests, we report statistical significance at the 95 % confidence level (p<0.05).

The 41-year daily snow cover reconstruction produced by snowMapper was based on 1.1 billion clear sky observations (image pixels) from Landsat and Sentinel-2 (Fig. 4), representing just 7.6 % of the final dataset. The remaining 92.9 % of snow cover values were reconstructed through the model's two-step gap-filling approach: the decision-tree algorithm (44.9 %) and the machine learning classifier (47.5 %).

A clear annual cycle is evident in the proportion of grid cells reconstructed by the decision-tree algorithm, which is higher during winter months, compared to the machine learning classifier. This pattern reflects seasonal variation in the reliability of the air temperature – snow cover relationship. In winter, high altitude temperatures are often ≤0 °C, enabling the decision-tree to confidently classify snow presence based on the previous day's conditions. However, during snow transition periods – late autumn/early winter and late spring – snow presence may not coincide with freezing temperatures preventing the first decision-tree criterion (Eq. 9) from being satisfied.

The fraction of values gap-filled by the decision-tree algorithm increased after 2000 due to the launch of MODIS Terra, while the availability of clear-sky observations has risen more than fivefold since the launch of Sentinel-2 during the last decade of the study period.

Overall, across the ten massifs, the daily snow cover reconstruction achieved an average accuracy of 93 % (Fig. 5). Of the remaining 7 % of misclassified cases, snow presence was underestimated in 4.7 % (snow present but classified as absent) and overestimated in 2.3 % (snow absent but classified as present). Model performance declines slightly during the transitional months of the snow season compared to mid-winter, primarily due to a systematic underestimation of snow cover.

Across the Greek mountains, SCA has exhibited a steep declining trend over the past four decades, both for the overall snow season (-8.64 km2yr-1) and for each month from November through to May (Fig. 6). Trends are statistically significant for all months except February (where p=0.1). The steepest absolute declines occur in December (-18.9 km2yr-1), followed by March (-13.7 km2yr-1), January and April (both -7.9 km2yr-1), February (-7.3 km2yr-1), November (-5.9 km2yr-1), and May (-0.6 km2yr-1). When these rates are normalised by each month's mean SCA (1984–2025), the greatest proportional losses emerge at the beginning and end of the snow season: November (-3.0 %yr-1), December (-1.9 %yr-1), April (-1.6 %yr-1), May (-1.3 %yr-1), and March (-1.1 %yr-1). In contrast, the mid-winter months show markedly smaller fractional declines, with January at -0.5 %yr-1 and February at -0.4 %yr-1.

Beyond the overall downward trends, the monthly evolution of SCA exhibits three distinct temporal patterns (Fig. 7):

Steady decline. November shows a consistent, approximately linear decrease in SCA throughout the past four decades.

Trend analysis of the monthly SCA anomalies across the study area and time period reveals a significant decline in both the magnitude and frequency of extremely high snow-cover instances, while no significant trend is observed for extremely low snow cover instances (Fig. 8).

Pearson's correlation analyses show that, across the entire study area, SCA is more strongly associated with temperature and precipitation than with regional climate indices (Fig. 9). Air temperature displays strong and statistically significant anticorrelations with SCA in all months, whereas precipitation shows moderate to strong positive correlations during November–December and February–March. Among the climate indices, a significant anticorrelation with NAO is found only in February, while both the AO and AMO show significant anticorrelations in February and March. These patterns are broadly consistent at the individual massif scale, with only minor shifts (±1 month) in the timing of significant correlations with precipitation and the climate indices (Fig. A7). At the massif scale, February is the only month in which all ten massifs display a significant positive correlation between SCA and precipitation, while March is the only month in which all ten massifs display a significant negative correlation between SCA and AMO. Among the three climate indices, the AO exhibits the clearest and most coherent behaviour, with positive correlations across all ten massifs in early winter (December–January) and negative correlations in the later months (February–April), significant only for the massifs located in the northern parts of mainland Greece (Fig. A7).

Levene's tests, applied to the two 21 year sub-periods (1984/85–2004/05 and 2004/05–2024/25) for each month, revealed no significant differences in the variability of total SCA between the two sub-periods. This result is consistent both across the entire study area and at the scale of the individual massifs (Fig. A6).

Over the past four decades, ten of the highest massifs in Greece have undergone a rapid decline in SCA throughout the snow season, amounting to an average reduction of 57.5 % (±1.4 %) at the end of the study period relative to the 1984–2025 mean. Rates of decline are greatest at both the beginning and end of the season, consistent with global evidence for a shortening of snow-cover duration driven by later onset and earlier melt-out dates (Notarnicola, 2020). Snow cover declined steadily in November, December and March across most of the record, whereas January, April and May show a pronounced drop around the turn of the century (±10 years). In contrast, the decline in February becomes apparent only within the past decade. Corresponding patterns are observed in the extremely positive SCA anomalies, which display a marked shift – both in frequency and magnitude – toward milder conditions, consistent with the reported decline in winter cold spell frequency (Tringa et al., 2022) and the overall warming trends (Fig. A8).

Although our analysis incorporates satellite-derived snow cover data, these observations are used within snowMapper primarily to constrain the simulated snow cover and limit the temporal propagation of gap-filling errors. Therefore, the reported trends are unlikely to be biased by the increasing availability of satellite observations in recent decades (Bayle et al., 2024). This is because the analysis is based not on irregularly captured satellite imagery, but on monthly aggregates derived from reconstructed daily snow cover maps.

To compare our results with existing literature on Greece and with other, better-studied mountain ranges, we repeated the trend analysis workflow described in Sect. 2.6 using the snow-cover phenology dataset of Notarnicola (2024a, b). Because that dataset is based on MODIS and begins only in 2000, we filtered our own dataset to the same time period, and re-ran the trend analysis. This revealed, once again, a significant decline in SCA, with a Sen's slope, normalised to the mean snow cover area, of -2.3 %yr-1. The snow loss estimated from the Notarnicola (2024a) dataset was of comparable magnitude (-1.8 %yr-1), with a slight underestimation likely due to methodological differences: our analysis is restricted to the snow season (November–May), whereas Notarnicola (2024b) considers the full hydrological year.

Using the same dataset, comparable rates of snow loss over 2000–2023 are evident for other mountains with Mediterranean or Mediterranean-like climates and marginal snowpacks (Table 1). In contrast, major mountain ranges with more persistent snowpacks exhibit far slower declines (e.g., Rockies: -0.02 %yr-1; Hindu Kush Himalaya: -0.04 %yr-1; Cordillera Central: -0.9 %yr-1) (Notarnicola, 2024a). These comparisons highlight that marginal snowpacks, especially those of the Greek mountains, are diminishing at some of the fastest rates observed globally.

The strong anticorrelation between SCA and air temperature – along with the moderate correlation with precipitation – across the ten massifs is consistent with findings from other Mediterranean mountain regions (Alonso-González et al., 2020; López-Moreno et al., 2025). These relationships indicate that temperature is the dominant control on SCA throughout the snow season, reflecting its dual influence on both atmospheric processes (e.g., the partitioning of precipitation into rain or snow) and land-surface snowpack processes, including accumulation, compaction, sublimation, and melt (Fayad et al., 2017). The role of precipitation, significant mainly in early (November–December) and late winter (February–March), aligns with the region's seasonal dynamics, during which zonal and meridional circulation patterns respectively contribute most to snow accumulation (Bartzokas et al., 2003). However, the lack of a persistent drying trend in recent decades (Fig. A8; Lagouvardos et al., 2024; Vicente-Serrano et al., 2025) suggests that declining snowfall, and therefore decreasing SCA, is primarily temperature-driven rather than the result of reduced precipitation (Abbas et al., 2024; Barnett et al., 2005; Giorgi et al., 1997; Kad et al., 2023; Li et al., 2025; Monteiro and Morin, 2023; Pepin et al., 2015, 2022; Rottler et al., 2019).

In contrast to the strong anticorrelation between SCA and temperature, the negative correlations with the NAO, AO, and AMO are generally weak and become significant only during the peak of the snow season (February–March), with small spatial variations among the massifs. The concurrent timing of the NAO and AO anticorrelations likely reflects large-scale synoptic patterns that influence the northeastern Mediterranean in late winter, particularly the formation of a blocking Central European anticyclone or the expansion of the Siberian anticyclone, and their downstream effects over the Balkans and eastern Mediterranean (Bartzokas et al., 2003; Tayanç et al., 1998). While the NAO exerts a strong influence on winter precipitation and temperature over the western Mediterranean (Lionello et al., 2006; López-Moreno et al., 2011), thereby modulating SCA in regions such as the Atlas Mountains (Marchane et al., 2016), our results indicate that this influence weakens markedly toward the eastern Mediterranean.

Although climate change is often associated with increased variability (Nordling et al., 2025; Pendergrass et al., 2017), our comparison of two 21 year subperiods (1984/85–2004/05 and 2004/05–2024/25) revealed no consistent changes in SCA variability across the ten massifs. This likely reflects the magnitude and spatial scale at which snow cover has declined over the past four decades. Furthermore, the rapid warming of the eastern Mediterranean and Greece (Lagouvardos et al., 2024), combined with the strong influence of temperature on SCA, means that even in years with exceptional snowfall, such as February 2022 (Patlakas et al., 2024), snowpack conditions remain unfavourable for persistent cover, leading to accelerated melt.

These findings are particularly relevant for marginal snowpacks in mountains with Mediterranean and Mediterranean-like climates, which are shifting towards increasingly ephemeral snow cover (López-Moreno et al., 2025; Sturm and Liston, 2021). They also carry implications for more continental mountain ranges that currently maintain seasonal snowpacks but may be transitioning toward marginal conditions. Underlined by the decreasing magnitude and frequency of positive SCA anomalies, our results add to a growing body of evidence warning of escalating snow-drought risk in Mediterranean and global mountain regions (Avanzi et al., 2024, 2025; Baba et al., 2025; Gottlieb and Mankin, 2025; Han et al., 2025; Pokharel et al., 2024; Singh et al., 2025). Taken together, these conclusions highlight the urgent need for modernising and optimising water-resource management strategies in Greece.

snowMapper demonstrated strong skill in reconstructing a 41-year snow-cover climatology across the temperate mountains of mainland Greece. The framework provides a modular, medium-complexity solution for gap-filling spatially and temporally patchy satellite observations, using a novel modelling approach that integrates area masking, preprocessing, gap-filling, post-processing, and validation. Furthermore, the high performance of the model's PIML component – achieving high accuracy despite being trained on in situ data from outside the study area – highlights its potential for broader application in other data-sparse or unmonitored mountain regions (e.g., Pritchard, 2021). In addition to offering a standalone model for directly reconstructing snow-cover time series, snowMapper could also be used as part of more complex hydrological modelling workflows, through the use of daily high-resolution snow-cover maps for deriving boundary conditions and even calculating snow depletion curves, ultimately replacing coarser-resolution MODIS-based datasets (Tekeli et al., 2005).

The slightly lower performance of snowMapper during the transitional months of the snow season reflects a well-documented challenge in snow-cover and snow-hydrological modelling (Chereque et al., 2025; Krinner et al., 2018; Toure et al., 2018). These months are characterised by high meteorological variability (e.g., air temperatures fluctuating around 0 °C that complicate rain-snow partitioning and melt-freeze cycles) and strong pre-conditioning of land-surface conditions (e.g., above-freezing ground temperatures at the start of the snow season that produce short-lived snowpacks).

We emphasise that the careful selection of a reanalysis product is essential. Because the snow reconstruction relies on reanalysis-derived air temperature and precipitation, the regional accuracy and native spatial resolution of the chosen dataset exert a strong influence on overall model quality. These factors propagate through every stage of the workflow, from MicroMet-based downscaling and the decision-tree gap-filling algorithm to the performance of the machine learning classifier. In this study, we used ECWMF's ERA5-Land, due to its comparatively high spatial resolution. Despite this, its land mask may include mountain grid cells that overlap coastal areas, potentially introducing a warm bias for mountainous terrain on islands or narrow peninsulas. In such cases, MicroMet's lapse-rate downscaling can produce positively biased air temperatures, which then propagate through the gap-filling steps and result in earlier modelled snowmelt or, in some cases, a complete absence of snow accumulation when precipitation falls under conditions perceived as too warm for snowfall.

Future versions of snowMapper will incorporate downscaling of additional meteorological variables, including incoming shortwave radiation, relative humidity, and wind speed. These additions would allow the calculation of wet-bulb temperature, providing a more physically robust alternative to the constant Tbase used for rain-snow partitioning. Because such variables are commonly available in widely used reanalysis products, they could improve the performance of the machine-learning classifier where corresponding in situ observations exist. Furthermore, alternative machine learning approaches, such as support vector machines or regression tree boosting, will be integrated to provide additional methods for training the classifier, other than random-forest.