Spatiotemporal Variability of Seasonal Snow Cover over 25 Years in the Romanian Carpathians: Insights from a MODIS CGF-Based Approach

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This manuscript presents a 25-year, pixel-based analysis of seasonal snow-cover phenology (SCD, SOD, SED) across the Romanian Carpathians using daily MOD10A1F (MODIS/Terra Cloud-Gap-Filled Snow Cover Daily L3 Global 500 m SIN Grid, Version 6.1) snow data at 500 m resolution, combined with FABDEM and Mann–Kendall/Theil–Sen trend analyses. The workflow description (data preparation, hydrological-year organization, derivation of SCD/SOD/SED, terrain stratification, trend mapping) is clear, facilitating replication and adoption. The topic is scientifically relevant and important, and the regional focus on an observation-sparse mountain system is timely. However, the current version has several methodological and statistical weaknesses that limit confidence in the inferred metrics and trends. Below are detailed comments.

1) Quantitative validation / uncertainty assessment

The manuscript does not provide sufficient accuracy assessment quantitatively for snow classification or for derived phenology dates (especially SOD/SED). In mountain terrain, MODIS snow retrieval errors (cloud/snow confusion, forest masking effects, terrain shadows, mixed pixels) can propagate into onset and melt timing. Please considering adding said validation, e.g.:

Validation against available in-situ snow observations (even if limited to certain elevations/years) reporting for snow-day counts and onset/end dates.

Cross-sensor comparison in representative sub-areas to benchmark MODIS-derived SCD/SOD/SED.

2) CGF “gap filling”

From what I understand, MOD10A1F CGF products replace cloudy pixels using temporally nearby clear-sky observations. This can introduce temporal lag (or persistence artifacts) during snow onset/melt periods—which is exactly where SOD/SED are most sensitive. The manuscript should state whether CGF quality layers and information were used. If not, please consider adding or provide a brief caveat statement.

Also, correct me if I am wrong, but the description of the MOD10A1F cloud-gap-filled (CGF) algorithm in the manuscript appears inaccurate. The manuscript states that CGF “substantially reduces [cloud] limitation through SPATIOTEMPORAL interpolation of adjacent clear-sky observations” (citing [33,54,55]), but the official MODIS Snow/Ice documentation describes MOD10A1F as a temporal carry-forward / retention approach: when a pixel is cloud on the current day, the product retains a previous-day non-cloud observation for that pixel, with quality information such as cloud persistence / observation age to indicate how “old” the substituted value is. In addition, Hall et al. (2019) explicitly frames the CGF procedure as using the current-day MOD10A1 and the previous-day MOD10A1F, replacing cloudy pixels with the most recent clear-sky observation and tracking persistence, rather than performing spatial interpolation.

3) NDSI

The paper uses NDSI ≥ 40 to classify snow while also providing the standard NDSI formula (−1 to 1). This can confuse readers unless you clearly explain the product scaling (e.g., 0–100 scaled NDSI) and confirm that you are thresholding the product’s scaled values rather than a raw computed NDSI. Please also provide a citation and/or sensitivity test supporting the choice of 40, especially for forested and mixed pixels.

4) Justification for using Terra

The manuscript argues Terra has more favorable illumination than Aqua due to earlier overpass time, but does not quantify the impact of excluding Aqua (or Terra+Aqua compositing) on cloud/shadow effects and phenology metrics. Consider adding a comparison of Terra-only vs Aqua-only vs combined approaches, or additional citation/discussion if feasible.

5) Consider increasing font size in Figure 8-10.

Author Response

Comment 1: The manuscript does not provide sufficient accuracy assessment quantitatively for snow classification or for derived phenology dates (especially SOD/SED). In mountain terrain, MODIS snow retrieval errors (cloud/snow confusion, forest masking effects, terrain shadows, mixed pixels) can propagate into onset and melt timing. Please considering adding said validation, e.g.: Validation against available in-situ snow observations (even if limited to certain elevations/years) reporting for snow-day counts and onset/end dates. Cross-sensor comparison in representative sub-areas to benchmark MODIS-derived SCD/SOD/SED.

Response 1: To address the lack of quantitative accuracy assessment for MODIS snow classification and derived snow-phenology dates, we added a station-based validation and described the full procedure in the Methods, with the main accuracy outcomes summarized in the Discussion. Specifically, we validated MODIS/Terra MOD10A1F CGF (NDSI-based) snow metrics against daily station snow observations for July 2014–September 2025 using 26 stations spanning 586–2504 m a.s.l. (Table S1). Each station was paired with the nearest MODIS 500 m pixel and aggregated to hydrological years (1 Oct–30 Sep). We quantified accuracy for hydrological-year snow-cover duration (SCD) using Pearson correlation, MAE, RMSE, and bias (median and IQR), reported overall and by elevation class (Table 4; Fig. 13). To evaluate daily snow-occurrence classification accuracy during the snow season (Oct–May), we computed the F1 score (harmonic mean of precision and recall) for each station-year and summarized results by elevation class (Fig. 14). For phenology timing, station-based SOD and SED were derived using the same 5-day persistence rule as MODIS, and timing accuracy was summarized using bias and MAE by elevation class (Table 4).

Comment 2: CGF “gap filling”. From what I understand, MOD10A1F CGF products replace cloudy pixels using temporally nearby clear-sky observations. This can introduce temporal lag (or persistence artifacts) during snow onset/melt periods—which is exactly where SOD/SED are most sensitive. The manuscript should state whether CGF quality layers and information were used. If not, please consider adding or provide a brief caveat statement. Also, correct me if I am wrong, but the description of the MOD10A1F cloud-gap-filled (CGF) algorithm in the manuscript appears inaccurate. The manuscript states that CGF “substantially reduces [cloud] limitation through SPATIOTEMPORAL interpolation of adjacent clear-sky observations” (citing [33,54,55]), but the official MODIS Snow/Ice documentation describes MOD10A1F as a temporal carry-forward / retention approach: when a pixel is cloud on the current day, the product retains a previous-day non-cloud observation for that pixel, with quality information such as cloud persistence / observation age to indicate how “old” the substituted value is. In addition, Hall et al. (2019) explicitly frames the CGF procedure as using the current-day MOD10A1 and the previous-day MOD10A1F, replacing cloudy pixels with the most recent clear-sky observation and tracking persistence, rather than performing spatial interpolation.

Response 2: Thank you for this important clarification regarding the MOD10A1F cloud-gap-filled (CGF) procedure. We agree and have corrected the description of the MOD10A1F cloud-gap-filled (CGF) procedure. In the revised Methods (rows 187–201), we now describe CGF accurately as a temporal carry-forward/retention approach in which cloudy pixels are replaced by the most recent prior clear-sky observation for the same pixel, and we explicitly note the presence of the accompanying observation age/persistence quality information. We also clarified that, in this study, we used the CGF classification as provided (without additional filtering based on the persistence/age layer) and added a brief caveat that carry-forward during prolonged cloudy periods may introduce temporal lag that is most relevant for SOD/SED during rapid transition phases.

Comment 3: The paper uses NDSI ≥ 40 to classify snow while also providing the standard NDSI formula (−1 to 1). This can confuse readers unless you clearly explain the product scaling (e.g., 0–100 scaled NDSI) and confirm that you are thresholding the product’s scaled values rather than a raw computed NDSI. Please also provide a citation and/or sensitivity test supporting the choice of 40, especially for forested and mixed pixels.

Response 3: Thank you for this important comment. We have revised the manuscript to explicitly clarify the scaling of the MOD10A1F NDSI Snow Cover layer and to confirm that our snow classification applies the threshold to the product’s scaled values (0–100) rather than to a raw, independently computed NDSI in the conventional −1 to 1 formulation. Specifically, we now state that in MOD10A1F the NDSI Snow Cover layer is provided as a scaled integer (0–100), and we therefore apply NDSI ≥ 40 to the scaled product values, which is equivalent to approximately NDSI ≥ 0.4 in the conventional −1 to 1 range. We also added supporting citations for using NDSI ≥ 0.4 as a widely adopted threshold in MODIS snow mapping. In addition, to support the choice of T40—particularly relevant to transitional and partly vegetated conditions—we performed a dedicated sensitivity analysis by repeating the full workflow using alternative scaled thresholds T30, T35, and T45, and quantified resulting differences in SOD, SED, and SCD relative to the T40 baseline across five elevation classes (≤500, 500–1000, 1000–1500, 1500–2000, >2000 m). The analysis shows that threshold-related effects are primarily concentrated in day-level timing (SOD/SED) and in transitional snow conditions, while seasonal aggregates and long-term behavior (spatial patterns, interannual variability, and trend diagnostics) remain comparatively robust. We summarize median offsets in the main text (Fig. 17) and provide additional robustness diagnostics in the Supplementary Materials (Fig. S5 and Tables S2abc).

Comment 4: Justification for using Terra. The manuscript argues Terra has more favorable illumination than Aqua due to earlier overpass time, but does not quantify the impact of excluding Aqua (or Terra+Aqua compositing) on cloud/shadow effects and phenology metrics. Consider adding a comparison of Terra-only vs Aqua-only vs combined approaches, or additional citation/discussion if feasible.

Response 4: We addressed this point by adding a dedicated Terra–Aqua–composite sensitivity test over the full Aqua-available period (hydrological years HY2003–HY2025). Using the same NDSI threshold and processing workflow, we derived SCD, SOD and SED from (i) Terra-only (MOD10A1F), (ii) Aqua-only (MYD10A1F), and (iii) a same-day Terra+Aqua composite, and summarized agreement by elevation bands (Fig. 18). Results show that SCD is highly consistent across approaches (r ≈ 0.91–0.996), indicating that excluding Aqua does not materially affect the duration metric used in the manuscript. While systematic offsets exist (Aqua–Terra: −13.9 to −5.8 days; Composite–Terra: +5.5 to +25.4 days), the dominant elevational patterns remain unchanged. For phenology timing, compositing generally improves pixel-level agreement relative to Terra–Aqua alone, whereas SED agreement is lower than SOD, consistent with spatially patchy late-season snow and differences in overpass geometry/illumination. Overall, this sensitivity analysis supports our use of Terra-only as a robust, internally consistent record for the study objectives, and indicates that including Aqua or compositing would not alter the main conclusions.

Comment 5: Consider increasing font size in Figure 8-10.

Response 5: We carefully checked Figures 8–10 and increased the font size (axis labels, tick labels, and legend/text annotations where applicable) to improve readability and visual clarity in both the main manuscript and the exported final-resolution figures.

Reviewer 2 Report

Comments and Suggestions for Authors

The manuscript presents a comprehensive study that provides the first detailed, long-term (2000-2025) assessment of snow cover phenology across the entire Romanian Carpathians using MODIS data. It effectively documents a clear shortening of the snow season, primarily driven by earlier melt, with significant altitudinal and regional variations. The following points aim to strengthen the discussion and contextualize the findings further.

1. The study applies a fixed NDSI ≥ 0.4 threshold to identify snow cover across all elevations and land cover types. However, snow detectability using NDSI can vary substantially with vegetation cover, terrain shading, and elevation. How sensitive are the derived snow phenology metrics (SCD, SOD, SED) to the chosen NDSI threshold, particularly in low elevation forested areas versus high-elevation alpine zones? Have the authors conducted any validation or uncertainty assessment (e.g., comparison with in situ observations or alternative snow products) to quantify potential biases?
2. The authors conclude that snow cover at elevations ≥2000 m is relatively stable and resilient to recent climate change, based largely on non-significant Mann-Kendall trends. How do the authors distinguish between true climatic stability and potential limitations related to sample size, reduced spatial coverage, or limited statistical power at high elevations?
3. While the manuscript attributes the observed shortening of snow cover mainly to earlier snowmelt, the analysis does not explicitly link snow changes to temperature or precipitation variability.
4. The methodological framework (MODIS data, Mann-Kendall test, Sen’s slope) is widely used in snow-cover trend analyses. Beyond providing regional coverage for the Romanian Carpathians, what are the key conceptual or scientific advances of this study compared to previous MODIS based snow phenology studies in other European or global mountain regions?

Author Response

Comment 1: The study applies a fixed NDSI ≥ 0.4 threshold to identify snow cover across all elevations and land cover types. However, snow detectability using NDSI can vary substantially with vegetation cover, terrain shading, and elevation. How sensitive are the derived snow phenology metrics (SCD, SOD, SED) to the chosen NDSI threshold, particularly in low elevation forested areas versus high-elevation alpine zones? Have the authors conducted any validation or uncertainty assessment (e.g., comparison with in situ observations or alternative snow products) to quantify potential biases?

Response 1: Thank you for raising this point. We agree that NDSI-based snow detectability can vary with vegetation cover, terrain shading, and elevation. Therefore, we quantified threshold sensitivity by repeating the full MOD10A1F processing workflow using scaled thresholds 30, 35, 40, and 45 (i.e., thresholding the product’s 0–100 NDSI_Snow_Cover layer) and computing SOD, SED, and SCD for each case. Differences were evaluated relative to the T40 baseline across five elevation bands (≤500, 500–1000, 1000–1500, 1500–2000, >2000 m) as an elevation-based proxy separating predominantly forested low elevations from open high-elevation terrain. Results are summarized in Fig. 17 (median offsets with IQR) and Supplementary Fig. S5 (percent of pixels within tolerance), with long-term robustness quantified in Supplementary Tables S2abc (trend and variability diagnostics). The sensitivity test shows that threshold effects are largest for day-level timing metrics (SOD/SED) and are most pronounced in transitional/intermittent snow conditions at mid elevations (e.g., 1000–1500 m, where T30–T40 yields median shifts up to ΔSED ≈ +12.2 days and ΔSCD ≈ +17.8 days; Fig. 17). In contrast, tolerance-based robustness is high at low and high elevations: at ≤500 m, 91.0–96.5% of pixels remain within |ΔSOD| ≤ 7 days, and >99% remain within |ΔSCD| ≤ 10 days across comparisons; at >2000 m, robustness is similarly high (|ΔSOD| ≤ 7 days: 97.5–100%; |ΔSCD| ≤ 10 days: 99.2–99.9%). Importantly, diagnostics of multi-decadal behavior remain highly consistent with the T40 baseline across all elevation bands: minimum Pearson correlations versus T40 remain high (~0.895–0.998) and maximum absolute Theil–Sen slope differences remain small (≤ ~0.571 days yr⁻¹) even under the strongest perturbation (Supplementary Tables S2abc). These results indicate that while day-level onset/melt timing can shift under larger threshold changes—especially around the snowline and in mixed/partly vegetated conditions—the regional spatial patterns, interannual variability, and long-term trend signals reported in the study are robust, supporting retention of NDSI ≥ 40 as the primary threshold. In addition to the threshold-sensitivity analysis, we added a quantitative accuracy assessment against in situ observations to contextualize potential biases in both snow classification and the derived phenology metrics. Specifically, we validated MODIS/Terra MOD10A1F CGF snow metrics using daily station snow observations available for Romania during July 2014–September 2025 from MeteoManz (land SYNOP/BUFR reports). We used 26 stations spanning 586–2504 m a.s.l. (Table S1), paired each station with the nearest MODIS 500 m pixel, and aggregated both datasets to hydrological years (1 Oct–30 Sep). Agreement for hydrological-year snow-cover duration (SCD) is strong overall (Pearson r = 0.95; MAE = 16.7 days; RMSE = 21.4 days; median bias = −8.0 days, IQR −19.8 to 4.0; Fig. 13), with performance highest at 500–1000 m (r = 0.91; MAE = 10.7 days) and remaining good at >2000 m (MAE = 11.7 days; Table 4). To evaluate daily snow-occurrence classification during the snow season (Oct–May), we computed the F1 score (harmonic mean of precision and recall) for each station-year and summarized results by elevation class; median F1 increases with elevation from ~0.73 (500–1000 m) to ~0.95 (>2000 m) (Fig. 14), indicating improved day-to-day agreement in the alpine zone. For phenology timing, station-based SOD and SED were derived using the same 5-day persistence rule as MODIS to ensure comparability, and timing differences were summarized as bias and MAE by elevation class (Table 4). Overall, this validation provides an empirical accuracy context showing that SCD is well captured at the station scale, while SOD/SED exhibit larger spread—particularly for SED at mid elevations—consistent with the greater intermittency and patchiness of snow cover during transition periods.

Comment 2: The authors conclude that snow cover at elevations ≥2000 m is relatively stable and resilient to recent climate change, based largely on non-significant Mann-Kendall trends. How do the authors distinguish between true climatic stability and potential limitations related to sample size, reduced spatial coverage, or limited statistical power at high elevations?

Response 2: We agree that non-significant Mann–Kendall results alone should not be interpreted as proof of climatic stability. We therefore revised the wording to avoid implying “resilience” and instead report that high-elevation terrain (≥2000 m) shows no robust evidence of change and weaker trend magnitudes relative to lower elevations. To explicitly address concerns about spatial coverage and statistical power, we quantified the extent of the ≥2000 m class on the MODIS grid (Table S3a): it includes 1,943 pixels (~395.2 km²), representing 0.589% of the study domain (330,150 pixels; ~67,154.7 km²). We further tested trends using an area-aggregated annual (hydrological-year) time series over ≥2000 m, which reduces pixel-level noise and improves power (Table S3b). This aggregated analysis yields near-zero effect sizes with confidence intervals spanning zero (SCD Sen slope −0.303 days yr⁻¹, 95% CI [−1.313, 0.835], MK p=0.624; SOD +0.012 days yr⁻¹, 95% CI [−0.234, 0.560], MK p=0.815; SED +0.079 days yr⁻¹, 95% CI [−0.803, 1.258], MK p=0.834). These results support our interpretation that changes at ≥2000 m are small in magnitude over 2000–2025, while also acknowledging the limited areal extent of this elevation band.

Comment 3: While the manuscript attributes the observed shortening of snow cover mainly to earlier snowmelt, the analysis does not explicitly link snow changes to temperature or precipitation variability.

Response 3: We agree and have added an explicit climate–snow linkage analysis to directly relate snow-phenology variability and trends to temperature and precipitation variability. This new analysis is presented in Section 4.3 and Figs. 15–16, using seasonal air temperature and precipitation from 43 homogenized meteorological stations (Dumitrescu et al., 2025). We aggregated station data into metric-relevant seasonal windows (Nov–May for SCD, Sep–Nov for SOD, Mar–May for SED) and quantified associations using Spearman rank correlations and trend statistics. In addition, we provide spatial context for regional climate change using E-OBS Sen-slope maps for cold-season temperature and precipitation trends (Figs. S3–S4). Overall, the added results show that air temperature exhibits coherent warming and stronger covariation with snow metrics than precipitation, supporting temperature as the dominant climatic covariate of the observed snow-season shortening, whereas precipitation trends/relationships are weaker and less consistent (Figs. 15–16; Figs. S1–S2).

Comment 4: The methodological framework (MODIS data, Mann-Kendall test, Sen’s slope) is widely used in snow-cover trend analyses. Beyond providing regional coverage for the Romanian Carpathians, what are the key conceptual or scientific advances of this study compared to previous MODIS based snow phenology studies in other European or global mountain regions?

Response 4: We agree that the core processing and trend-detection methods are established. We therefore clarified the conceptual contribution of the study and added a short statement at the beginning of Section 4.1 emphasizing what is new beyond applying standard tools in a new area. In particular, this work delivers the first Carpathians-wide, internally consistent, 25-year (2000–2025), daily 500 m MODIS/Terra CGF baseline of snow-cover duration and timing (SCD/SOD/SED) for a climatically distinct and historically under-instrumented mountain system. The advance is not only expanded spatial coverage, but a snow-regime framework that translates pixel-level information into structured, comparable diagnostics along environmental gradients: we systematically stratify snow metrics by elevation bands, aspect, and mountain units, which resolves within-belt heterogeneity and identifies where sensitivity to change is greatest—patterns that cannot be robustly inferred from sparse station networks or coarse regional summaries. In addition, we complement pixel-wise phenology with regional-scale diagnostics of snow extent and elevation control (snow-covered area and snowline elevation), providing a unified basis for interpreting snow persistence and timing in relation to periglacial and high-mountain sensitivity. Together, these elements provide a reproducible regional benchmark that supports cross-region comparison with other MODIS-based mountain studies and provides an explicit baseline for future monitoring and model-based investigations in Eastern European mountain cryosphere research.

 

Reviewer 3 Report

Comments and Suggestions for Authors

This study presents a comprehensive analysis of snow cover dynamics in the Romanian Carpathians over a 25-year period using MODIS CGF snow products. The manuscript is well-structured, methodologically sound, and addresses an important research gap in a region where long-term, spatially continuous snow observations are limited. The findings are relevant for understanding cryospheric changes in southeastern European mountains. However, several aspects require clarification, deeper analysis, and improved presentation before the manuscript can be considered for publication.

1. While the study is clearly valuable for the Romanian Carpathians, the Introduction could better articulate how this work advances beyond existing regional studies (e.g., Birsan & Dumitrescu, 2013; Amihaesei et al., 2024). Specifically, clarify how the use of pixel-level MODIS CGF data over 25 years provides new insights compared to previous station-based or coarser-scale analyses.
2. The choice of NDSI ≥ 40 as the snow-presence threshold is standard. A short sensitivity analysis or reference to validation studies specific to similar mountain environments would strengthen this methodological choice.
3. The use of a 5-day persistence filter for SOD/SED is appropriate, but the manuscript should explicitly state how temporary snow events (e.g., late spring snowfall after melt onset) are handled. A brief discussion on the potential impact of this filter on trend detection, especially at high elevations, would be helpful.
4. The study documents clear spatial and temporal patterns but does not link them explicitly to climatic variables (air temperature, precipitation, atmospheric circulation). While this may be beyond the scope, a brief correlation analysis or reference to known warming trends in the region would help contextualize the observed snow declines.
5. The implications are briefly mentioned but could be more concrete. For example, how might the observed ~11-day/decade SCD decline affect spring runoff timing or peak flow in Carpathian rivers? Similarly, can the reported changes be linked to observed shifts in vegetation phenology or alpine species distribution in the region?
6. Provide the full name and abbreviation when an abbreviation first appears in the text. In figures and tables, abbreviations should be given in their full names. Moreover, abbreviations in all figure captions must be consistent; for example, SCD is snow cover duration, not snow cover day.
7. The main focus of the article is on the spatiotemporal characteristics of snow cover. It is recommended to add a section analyzing the spatiotemporal distribution characteristics of snow cover area.
8. Include climatic factors influencing snow cover changes in the abstract.
9. Provide p-values in the trend analysis.

Author Response

Comment 1: While the study is clearly valuable for the Romanian Carpathians, the Introduction could better articulate how this work advances beyond existing regional studies (e.g., Birsan & Dumitrescu, 2013; Amihaesei et al., 2024). Specifically, clarify how the use of pixel-level MODIS CGF data over 25 years provides new insights compared to previous station-based or coarser-scale analyses.

Response 1: We agree and have revised the Introduction to more clearly state how this work advances beyond prior snow studies in the Romanian Carpathians, which are largely based on sparse station records or aggregated regional summaries (e.g., Bîrsan & Dumitrescu, 2013; Amihăesei et al., 2024). Specifically, we added a dedicated paragraph (Introduction, rows 83–94) explaining that pixel-level MODIS Terra CGF observations provide spatially continuous snow-phenology fields over the full domain, enabling analyses that are not feasible with station data alone. The added text highlights that the MODIS CGF time series allow (i) mapping sharp spatial contrasts driven by elevation, aspect, and terrain shading, (ii) consistent stratification of SCD/SOD/SED across topographic classes, and (iii) identification of within–elevation-band heterogeneity and potential elevational breakpoints in snow persistence. This clarification emphasizes that the novelty of our study lies in delivering a 25-year, near-daily, domain-wide and spatially explicit characterization of snow phenology and related diagnostics, rather than extrapolating from a limited number of point observations.

Comment 2: The choice of NDSI ≥ 40 as the snow-presence threshold is standard. A short sensitivity analysis or reference to validation studies specific to similar mountain environments would strengthen this methodological choice.

Response 2: Thank you for this suggestion. We strengthened the justification for using NDSI ≥ 40 in two complementary ways. First, we added additional citations documenting the standard MODIS snow-mapping threshold and its general suitability for 500 m snow products. Second, we introduced a dedicated threshold-sensitivity analysis. Specifically, we repeated the complete MOD10A1F processing workflow using scaled thresholds 30, 35, 40, and 45 (applied to the product’s 0–100 CGF_NDSI_Snow_Cover layer) and compared the resulting SCD, SOD, and SED against the T40 baseline across five elevation bands (≤500, 500–1000, 1000–1500, 1500–2000, >2000 m). The sensitivity results are presented in Fig. 17 (median offsets with interquartile range) and Supplemental Fig. S5 (percentage of pixels within tolerance), and the robustness of multi-decadal behavior (trend and variability diagnostics relative to T40) is summarized in Suppplemental Tables S2abc. Overall, the analysis shows that threshold effects are concentrated mainly in day-level SOD/SED timing under intermittent/transitional snow conditions, whereas SCD and the long-term trend/variability signals remain highly consistent with the NDSI ≥ 40 baseline across elevation classes. This supports retaining NDSI ≥ 40 as a consistent regional threshold for long-term comparisons.

Comment 3: The use of a 5-day persistence filter for SOD/SED is appropriate, but the manuscript should explicitly state how temporary snow events (e.g., late spring snowfall after melt onset) are handled. A brief discussion on the potential impact of this filter on trend detection, especially at high elevations, would be helpful.

Response 3: We agree and have clarified this explicitly in the revised manuscript (Chapter 4.5). We now state that temporary snow events are handled through a 5-day persistence criterion: brief autumn snowfalls, short melt interruptions, and late-season return snow are not treated as season onset or season end unless they persist for ≥5 consecutive days, so SOD/SED represent the seasonal snowpack rather than episodic events. We also added a short note on sensitivity: conceptually, a shorter window would tend to yield earlier SOD and later SED (greater sensitivity to transient events and higher interannual scatter), whereas a longer window would increasingly emphasize only the most persistent phases, potentially delaying SOD and advancing SED in marginal, patchy conditions at low–mid elevations. At high elevations, where snow cover is generally more continuous, we note that trend detection is expected to be less sensitive to the exact window length than near the snowline.

Comment 4: The study documents clear spatial and temporal patterns but does not link them explicitly to climatic variables (air temperature, precipitation, atmospheric circulation). While this may be beyond the scope, a brief correlation analysis or reference to known warming trends in the region would help contextualize the observed snow declines.

Response 4: We agree and have added an explicit climate-context analysis in the revised manuscript. Specifically, we introduced a new Section 4.3 (Climate–Snow Interactions) and expanded the Methods to describe how we used seasonal air temperature and precipitation from 43 homogenized Romanian meteorological stations (Dumitrescu et al., 2025). We aggregated climate variables into metric-relevant windows (Nov–May for SCD, Sep–Nov for SOD, Mar–May for SED) and quantified associations using Spearman rank correlations, considering both (i) regional (station-median) year-to-year series and (ii) station–year pairs (climate values paired with snow metrics sampled at station locations). The main results are presented in Figs. 15–16 and in Supplemental Figs. S1–S2: temperature shows the most consistent relationships with snow variability (including a significant Nov–May warming of +0.64 °C decade⁻¹, p = 0.018, alongside declining SCD in the same panel), whereas precipitation exhibits weaker and generally non-significant trends/associations across the corresponding windows. In addition, we provide spatial context for regional cold-season climate change using E-OBS trend maps (included in Supplemental Figs. S3-S4).

 

Comment 5: The implications are briefly mentioned but could be more concrete. For example, how might the observed ~11-day/decade SCD decline affect spring runoff timing or peak flow in Carpathian rivers? Similarly, can the reported changes be linked to observed shifts in vegetation phenology or alpine species distribution in the region?

Response 5: We agree and have revised Section 4.1 to make the environmental implications more explicit while keeping them appropriately scoped. We added two concise linkages based on published regional evidence. First, we relate the observed ~11 days decade⁻¹ reduction in snow-cover duration to hydrological expectations for snow-influenced catchments: earlier melt-out and reduced seasonal snow storage tend to shift the nival contribution toward earlier spring and increase the relative importance of winter/early-spring runoff, while emphasizing that peak-flow changes are not uniform and remain catchment- and event-dependent, particularly under rain-on-snow conditions [59–63]. Second, we strengthened the ecological context by linking earlier snowmelt and reduced spring snow persistence to earlier growing-season onset/green-up and potential sensitivity of alpine snowbed habitats and the forest–alpine ecotone/treeline, citing available Carpathian evidence and explicitly noting that species-range shifts are not quantified in the present study [55–58].

Comment 6: Provide the full name and abbreviation when an abbreviation first appears in the text. In figures and tables, abbreviations should be given in their full names. Moreover, abbreviations in all figure captions must be consistent; for example, SCD is snow cover duration, not snow cover day.

Response 6: Thank you for this correction. We revised the manuscript to ensure that all abbreviations are defined at first use with their full term followed by the abbreviation in parentheses (e.g., snow-cover duration (SCD), snow onset date (SOD), snow end date (SED), snow-covered area (SCA), snowline elevation (SLE)). We also standardized abbreviation usage across the entire manuscript, including all figure captions and table headings, and ensured that captions provide the full names where appropriate. In particular, we corrected inconsistent caption wording so that SCD is consistently reported as “snow-cover duration” (not “snow cover day”), and we harmonized terminology in figures/tables to match the definitions used in the text.

Comment 7: The main focus of the article is on the spatiotemporal characteristics of snow cover. It is recommended to add a section analyzing the spatiotemporal distribution characteristics of snow cover area.

Response 7: We agree and have added a dedicated snow-covered area (SCA) analysis to better characterize the spatiotemporal distribution of snow extent across the Romanian Carpathians (added at the end of the Results section). Specifically, we compute mean SCA (%) for three seasonal windows—winter (Dec–Feb), spring (Mar–Apr), and the full snow season (Nov–May)—and quantify interannual variability and linear trends separately for the Western, Southern, and Eastern Carpathians (Fig. 11). To relate areal snow extent to elevational controls, we also introduced a snowline elevation (SLE) diagnostic, summarized as monthly boxplots (Nov–May) for the same regions (Fig. 12), capturing the seasonal lowering and spring rise of the snowline that modulates SCA. Together, these additions explicitly quantify both the temporal evolution and regional contrasts in snow-cover area, complementing the pixel-based phenology metrics (SCD/SOD/SED).

Comment 8: Include climatic factors influencing snow cover changes in the abstract.

Response 8: We agree that climate-snow interaction information should be added to the Abstract, so we have revised it to explicitly incorporate the climatic context of the observed snow-cover changes. In the revised version, we now added a statement regarding climate-control analysis by mentioning that interannual variability in snow phenology is primarily associated with air-temperature variability, while precipitation relationships are weaker. At the same time, we updated the Abstract to reflect the main additions introduced during revision, including the new snow-covered area (SCA) and snowline elevation (SLE) diagnostics and the quantitative uncertainty assessment based on station comparisons (reporting correlation, bias, and MAE). These changes strengthen the Abstract by linking the reported snow-phenology trends to their dominant climatic covariate and by concisely highlighting the new analyses performed.

Comment 9: Provide p-values in the trend analysis.

Response 9: We have addressed this request by explicitly reporting p-values in the trend analysis. In the original trend figures, we now annotate each panel with the Theil–Sen slope together with the corresponding p-value for the trend significance test, so statistical significance can be assessed directly from the figure (not only inferred from confidence intervals or visual impression). In addition, we applied the same p-value reporting format to all newly added trend figures introduced in the revision (both in the main text and the Supplementary Material), ensuring a consistent presentation of trend magnitude + significance across the entire manuscript.

 

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The authors have thoroughly revised the manuscript and addressed most of my previous concerns.