acceptodds
Under review as a conference paper at ICLR 2027

SWECast: Enhancing Snow Water Equivalent Forecasting through Prior Modulation and Periodicity Modeling

Abstract

Snow water equivalent (SWE) directly measures the amount of water stored in snowpack. Accurate SWE forecasting is essential for water resource management and flood risk assessment in regions relying on seasonal snowpack. However, SWE evolution varies across locations and years and is sensitive to abrupt weather changes during the forecast horizon, which makes accurate SWE forecasting particularly challenging. We therefore propose SWECast, a novel model that improves SWE forecasting by combining periodicity modeling with prior modulation. We select informative meteorological inputs and construct an input configuration comprising complete historical snow seasons, current snow season history, and future meteorological inputs. We design a Periodicity Modeling Module (PMM) that analyzes SWE and meteorological inputs in the frequency domain, inferring the current SWE state and characterizing its response to weather. Based on this information, a gating mechanism dynamically weights each feature at each time step to incorporate meteorological information over the forecast horizon. In parallel, a Prior Conditioner (PC) derives priors from complete historical snow seasons in the time domain, capturing how SWE evolves at each location and thereby addressing the variability in SWE evolution across locations. Finally, the priors from the PC modulate the representations generated by the PMM. The evaluation results at stations excluded from training demonstrate that SWECast consistently outperforms all baselines. Over a forecast horizon of two weeks, SWECast reduces MAE by 13.1% and RMSE by 14.5% compared with the strongest baseline. By providing accurate SWE forecasting, SWECast can inform reservoir operations and flood preparedness in snow-dependent regions.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Snow Water Equivalent Forecasting

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_LnbdOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SWECast for 7–14 day snow water equivalent forecasting. It combines Soft-DTW-based historical-season retrieval, cross-channel attention, frequency-domain periodicity modeling, weather gating, and FiLM-style prior modulation. Experiments on unseen SNOTEL stations show consistent improvements over the evaluated baselines, including a 13.1% MAE reduction at the 14-day horizon.

Critical strengths

The model is well motivated and combines historical SWE patterns with future meteorological information in a structured way. The station-wise evaluation and ablation studies also provide reasonably strong empirical support for the proposed components.

Critical weaknesses

The main practical concern is that future temperature and precipitation are taken from ERA5-Land over the forecast horizon, which does not reflect uncertainty in operational weather forecasts. Since future meteorology directly affects the proposed weather-gating mechanism, robustness to forecast errors should be evaluated. The paper itself acknowledges this limitation. 38743_SWECast_Enhancing_Snow_W

The methodological novelty is also somewhat limited, as the approach mainly combines existing components such as Soft-DTW, TimesNet-style periodicity modeling, attention, and FiLM. Moreover, recent SWE-specific methods discussed in the related work are not included in the main comparison.

Questions for authors

Are the future Tm and TP inputs realized ERA5-Land values? If so, how does performance change under realistic meteorological forecast errors? Were the input-variable and history-length choices made exclusively using the validation set? Could the authors compare against recent SWE-specific models such as GEMS or ForeSWE, or explain why such comparisons are infeasible?

3 comments
AuthorsOpenReview ↗

Response to Reviewer Lnbd (1/3)

Dear Reviewer Lnbd,

Thank you for your careful review and constructive suggestions. We appreciate your recognition of SWECast’s motivation and the empirical support from the station-wise evaluation and ablation studies.

You raised concerns about meteorological forecast errors, methodological novelty, SWE-specific comparisons, and input and history-length selection. We have addressed them through experiments using archived weather forecasts, an expanded comparison with ForeSWE, and validation results documenting these choices. We have also clarified the design rationale and revised Figures 1 and 2.

W1 & Q1. Evaluation with operational weather forecasts

“…Are the future Tm and TP inputs realized ERA5-Land values? …how does performance change under realistic meteorological forecast errors?”

Response:

ERA5-Land provides meteorological data at 0.1° resolution throughout our 1993–2021 station record. The long record supports evaluation across diverse snow seasons, while the fine grid helps represent local meteorological conditions at the stations. These properties provide a suitable basis for comparing forecasting accuracy and generalization to unseen stations. We therefore used ERA5-Land reanalysis for all meteorological inputs in Table 1, including future temperature and precipitation.

Your suggestion allows us to test whether this accuracy advantage persists with imperfect weather forecasts. We therefore conducted additional experiments using archived operational forecasts from the Global Forecast System (GFS) as future meteorological inputs. GFS is run by NCEP within NOAA’s National Weather Service. The available archive begins in January 2015, with a 0.25° grid and forecasts up to 16 days ahead.

The experiments covered five snow seasons during 2016–2021. We used GFS future temperature and precipitation across training, validation, and testing and trained SWECast from scratch. We retained ERA5-Land for historical meteorological inputs. All models received identical future weather forecasts and used the same test stations, dates, and forecast windows. We have added the setup and results to Appendix C.4, Table 9, reproduced below.

Table 9. SWE forecasting with GFS future meteorological inputs and 60 days of current-season observations. MAE/RMSE: mean ± sample SD (mm) across three runs; NSE: mean. Bold: best; ☆: second-best. The same notation applies to Table 10.

Model7d MAE ↓7d RMSE ↓7d NSE ↑14d MAE ↓14d RMSE ↓14d NSE ↑
Random Forest12.51±0.31☆18.73±0.210.971020.34±0.5931.83±1.420.9210
XGBoost12.75±0.4618.39±0.19☆0.974420.03±0.72☆30.10±1.66☆0.9333☆
DLinear57.38±2.4572.14±2.720.551563.00±3.4579.44±3.570.4817
PatchTST29.41±0.8639.59±0.490.876137.44±2.3451.99±3.100.7883
iTransformer20.93±0.9128.50±1.120.937127.81±1.2640.41±0.870.8729
TimesNet19.17±0.9426.80±2.040.941326.95±1.2139.20±1.750.8780
SWECast11.63±0.3217.23±0.480.9733☆18.62±0.5528.02±1.300.9378

The results show that SWECast achieves the lowest MAE and RMSE at both horizons with GFS inputs. Forecasting accuracy decreases relative to the ERA5-Land setting, consistent with the additional challenge of uncertain future meteorological inputs.

SWECast also outperforms most baselines using ERA5-Land inputs across all three metrics at both horizons, even with imperfect GFS forecasts as future inputs. Its RMSE is numerically lower than that of every ERA5-Land baseline at both horizons. We have compiled this descriptive comparison in Appendix C.4, Table 10, reproduced below. These findings further support SWECast’s potential for practical SWE forecasting.

Table 10. SWECast using GFS forecasts compared with baselines using ERA5-Land reanalysis.

ModelFuture meteorology7d MAE ↓7d RMSE ↓7d NSE ↑14d MAE ↓14d RMSE ↓14d NSE ↑
Random ForestERA5-Land11.88±0.1619.69±0.510.96818.87±0.2532.57±0.320.922
XGBoostERA5-Land10.97±0.27☆17.68±0.710.975☆17.06±0.40☆28.31±0.530.942☆
DLinearERA5-Land43.64±2.1357.89±0.960.73954.82±1.8174.59±0.710.601
PatchTSTERA5-Land26.02±2.7935.45±3.280.89634.30±2.2751.00±2.520.809
iTransformerERA5-Land14.03±0.4721.64±0.780.96120.19±0.4232.37±0.350.917
TimesNetERA5-Land13.70±0.4721.04±1.260.96319.77±0.6130.97±0.450.923
ForeSWEERA5-Land17.82±2.0326.65±2.110.94124.86±3.3039.06±4.160.874
SNOW-17ERA5-Land75.38±1.26105.32±2.450.37776.38±2.44109.38±3.100.308
SNOW-17†ERA5-Land58.74±1.9781.75±4.130.63260.40±1.4484.20±2.090.568
SWECastERA5-Land10.10±0.2716.10±0.760.97814.82±0.3724.20±1.040.954
SWECastGFS11.63±0.3217.23±0.48☆0.973318.62±0.5528.02±1.30☆0.9378

† SNOW-17 calibrated using all available historical records at each station.

AuthorsOpenReview ↗

Response to Reviewer Lnbd (2/3)

W2. Methodological contributions of SWECast

“…methodological novelty is also somewhat limited, as the approach mainly combines existing components…”

Response:

SWECast addresses two practical challenges in SWE forecasting: generalization to stations excluded from training and abrupt weather changes during the forecast horizon. The central contribution is a forecasting architecture that uses historical patterns at the forecast station to modulate how a shared model processes current observations and future meteorological information.

1. Historical priors for generalization to unseen stations.

We observed recurring patterns in SWE time series across complete snow seasons at the same station. This suggests that past snow seasons can serve as a reference for forecasting SWE in the current season. We therefore construct a historical prior from the forecast station’s own records to guide the shared model’s predictions at stations excluded from training.

A useful prior requires a historical reference that matches the current conditions despite differences in seasonal timing. For example, a warm spring may cause snowmelt to begin earlier and progress faster than in a cooler year. SWECast therefore searches the preceding snow season for segments with similar SWE and meteorological patterns. Soft-DTW accommodates local differences in their timing and pace during matching. We then extract the matched segment and subsequent historical observations to construct the prior.

The resulting prior guides the forecast through FiLM modulation. Feature-wise scaling and shifting allow the prior to recalibrate the representations learned from current observations and future weather inputs. SWECast generates these modulation parameters for each prediction, enabling the shared network to adapt its forecasts to local SWE patterns.

2. Future meteorological information for responding to weather changes.

Historical SWE patterns alone may not anticipate rapid accumulation or melt caused by weather changes during the forecast period. We therefore include future temperature and precipitation to account for these changes. Cross-Channel Attention models the relation between SWE and meteorological variables. Within each Adaptive Periodicity Block, the weather gate adjusts feature channels according to meteorological conditions. These operations allow the forecast representations to respond to predicted weather changes.

We also conducted an additional comparison under three settings: ERA5-Land future inputs, archived GFS future inputs, and no future weather inputs. All three settings used identical test stations, dates, and forecast windows. The results show that future meteorological inputs improve forecasting performance for most models, even with imperfect weather forecasts. SWECast achieves the best MAE, RMSE, and NSE in each setting in this comparison, supporting the effectiveness of its architecture across different meteorological input conditions. We have provided the complete comparison in Appendix C.5, Table 11, reproduced below.

Table 11. Effects of removing future weather inputs. ↓: lower is better; ↑: higher is better. Bold: improvement over the same model’s w/o setting. w/o: no future weather inputs.

ModelMAE (mm) ↓RMSE (mm) ↓NSE ↑
ERA5-LandGFSw/oERA5-LandGFSw/oERA5-LandGFSw/o
Random Forest18.5820.0624.9432.7732.8041.230.9240.9170.885
XGBoost17.1419.5424.7928.8629.8640.630.9430.9330.890
DLinear53.4562.9949.1374.7881.0476.220.6030.4340.586
PatchTST35.8737.1033.7253.7051.4150.430.7940.7790.818
iTransformer19.7126.6727.0131.9740.6143.480.9200.8620.862
TimesNet19.0926.3530.0430.5438.4745.180.9290.8770.854
ForeSWE25.4829.8628.4039.9442.6544.980.8770.8450.856
SWECast14.4619.1523.0623.8329.2338.280.9580.9340.904

3. Experimental support for the design choices.

The component ablations support the roles of historical matching, cross-channel attention, periodicity modeling, and prior modulation in the full model (Table 2). For example, MAE increases from 14.46 to 21.93 mm when we remove the Prior Modulation Layer. This result supports the use of historical priors to recalibrate forecast representations. SWECast also achieves the best results against general forecasting models and ForeSWE in Table 1, reproduced in our response to W3 & Q3.

We have revised Figure 1 to clarify how the forecasting challenges motivate our design. We have simplified Figure 2 to make the inputs, outputs, and connections between modules clearer.

AuthorsOpenReview ↗

Response to Reviewer Lnbd (3/3)

W3 & Q3. Comparison with data-driven SWE forecasting models

“…compare against recent SWE-specific models such as GEMS or ForeSWE, or explain why such comparisons are infeasible?”

Response:

Our original comparison included a broad set of machine-learning and time-series baselines, together with SNOW-17. The consistent gains across these approaches provide strong evidence of SWECast’s forecasting accuracy and its ability to generalize to stations excluded from training.

Your suggestion prompted us to examine GEMS and ForeSWE as additional SWE-specific data-driven forecasting approaches. We found that GEMS used SNOTEL station observations for pretraining (Umirbekov et al., 2024). We therefore did not include its pretrained model to avoid potential data leakage.

We have conducted additional experiments with ForeSWE (Thapa et al., 2026) and updated Table 1 in the main text with the results. We have documented its implementation alongside those of the other baselines in Appendix B.2. We also present the updated table below for your convenience.

Table 1. SWE forecasting with 60 days of current-season observations. Three runs; MAE/RMSE: mean ± SD (mm); NSE: mean. ↓: lower is better; ↑: higher is better. Bold: best; ☆: second-best.

Model7d MAE ↓7d RMSE ↓7d NSE ↑14d MAE ↓14d RMSE ↓14d NSE ↑
Random Forest11.88±0.1619.69±0.510.96818.87±0.2532.57±0.320.922
XGBoost10.97±0.27☆17.68±0.71☆0.975☆17.06±0.40☆28.31±0.53☆0.942☆
DLinear43.64±2.1357.89±0.960.73954.82±1.8174.59±0.710.601
PatchTST26.02±2.7935.45±3.280.89634.30±2.2751.00±2.520.809
iTransformer14.03±0.4721.64±0.780.96120.19±0.4232.37±0.350.917
TimesNet13.70±0.4721.04±1.260.96319.77±0.6130.97±0.450.923
ForeSWE17.82±2.0326.65±2.110.94124.86±3.3039.06±4.160.874
SNOW-1775.38±1.26105.32±2.450.37776.38±2.44109.38±3.100.308
SNOW-17†58.74±1.9781.75±4.130.63260.40±1.4484.20±2.090.568
SWECast (Ours)10.10±0.2716.10±0.760.97814.82±0.3724.20±1.040.954

† SNOW-17 calibrated using all available historical records at each station.

SWECast achieves the best results across all three metrics at both forecast horizons in the expanded comparison. Relative to ForeSWE, it reduces MAE by 43.3% at 7 days and 40.4% at 14 days. This additional comparison extends the evidence for SWECast’s accuracy advantage to a recent model developed specifically for SWE forecasting.

Reference: Umirbekov, A., Essery, R., and Müller, D. (2024). GEMS v1.0: Generalizable Empirical Model of Snow Accumulation and Melt, based on daily snow mass changes in response to climate and topographic drivers. Geoscientific Model Development, 17, 911–929. DOI: 10.5194/gmd-17-911-2024.

Q2. Selection of meteorological inputs and history lengths

“…Were the input-variable and history-length choices made exclusively using the validation set?”

Response:

We selected the meteorological inputs, current-season history length, and number of complete historical snow seasons exclusively based on validation MAE. The meteorological inputs were chosen from physically motivated candidate combinations. We then evaluated the configurations on the test set, as reported in Table 3 and Figure 4. We have clarified this procedure in Section 5.5 and added the corresponding validation results to Appendix C.2, Tables 6–8, reproduced below. These results support our choices of temperature and total precipitation, 60 steps of current-season history, and one complete historical snow season.

Table 6. Validation MAE for meteorological input selection using 60 days of current-season observations and a 14-day forecast horizon. The candidate configurations correspond to those in Table 3. MAE is in mm; ↓ indicates lower is better. Best results are in bold. Tm: temperature; TP: total precipitation.

InputValidation MAE ↓ (mm)
SWE only28.50
SWE + Tm27.33
SWE + TP21.24
SWE + Tm + TP (Base)16.54
Base + Snow16.75
Base + Surface16.77
Base + Atmospheric18.18
Base + Vegetation16.94

Table 7. Validation MAE for current-season history lengths over 14 prediction steps. MAE in mm; ↓: lower is better; bold: best.

Current-season history length (steps)Validation MAE ↓ (mm)
1020.60
1518.58
3017.48
4518.74
6017.12
7521.26
9023.13
10524.09
12027.75

Table 8. Validation MAE for the number of complete historical snow seasons over 14 prediction steps. MAE in mm; ↓: lower is better; bold: best.

Complete historical snow seasonsValidation MAE ↓ (mm)
0 (w/o historical priors)28.38
116.96
217.03
317.18
417.22
517.48
Reviewer_xjzGOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SWECast for snow water equivalent (SWE) forecasts, which combines periodicity modeling with prior modulation. The inputs comprise complete historical snow seasons, current snow season history and future meteorological inputs. Over a forecast horizon of 14 days, SWECast reduces MAE and RMSE compared with the baseline models.

Critical strengths
  1. A useful and well-defined task. The paper studies short-range SWE forecasts at stations held out from model training. This is a meaningful test of whether the trained model can be used at other monitored stations.
  2. Consistent gains across the reported metrics. SWECast leads Table 1 at both forecast horizons. The paper also reports results over three random seeds and includes input and component studies. These results make the performance gain worth examining further.
Critical weaknesses

Limited evidence for real-world forecasting. The model uses future temperature and precipitation from ERA5-Land, a reanalysis product. These values are not available when a real forecast is made. The main results therefore do not show how well the method works with weather forecasts that contain errors. The authors should test it with archived weather forecasts. It is also unclear how much of the forecast skill comes from future weather inputs and how much of the skill comes from the model design. Table 3 includes a SWE-only setting, but it removes both past and future weather inputs. A separate test that removes only future weather, for SWECast and the baselines, would isolate this effect. The comparison also lacks a simple persistence baseline.

Questions for authors
  1. Please define SWE with a simple equation. Stating its units would also help readers understand what the model predicts.
  2. Can you repeat the main comparison using archived weather forecasts for future temperature and precipitation? If these forecasts are unavailable, please test several realistic levels of weather forecast error. Use the same weather inputs for every method.
  3. How do simple baselines perform on the same station split? Please include a persistence baseline.
  4. Please add a map showing the stations and mark which ones are used for training, validation, and testing. This would make the geographic coverage and station split easier to assess.
  5. Some lines and arrows in Figures 1 and 2 are meaningless. Please simplify the diagrams and make the inputs, outputs, and connections between modules clearer.
7 comments
AuthorsOpenReview ↗

Response to Reviewer xjzG (1/3)

Dear Reviewer xjzG,

Thank you for your careful review and constructive suggestions. We appreciate your recognition of our evaluation at stations held out from training and the consistent gains across the reported metrics.

You raised concerns about real-world forecasting, future weather inputs, simple baselines, and presentation. We have addressed them through experiments using archived weather forecasts, a comparison that removes only future weather inputs, and a persistence baseline. We have also clarified the SWE definition and units, added station maps, and revised Figures 1 and 2.

W1 & Q2. Evaluation with operational weather forecasts

“…repeat the main comparison using archived weather forecasts for future temperature and precipitation? …Use the same weather inputs for every method.”

Response:

ERA5-Land provides meteorological data at 0.1° resolution throughout our 1993–2021 station record. The long record supports evaluation across diverse snow seasons, while the fine grid helps represent local meteorological conditions at the stations. These properties provide a suitable basis for comparing forecasting accuracy and generalization to unseen stations. We therefore used ERA5-Land reanalysis for all meteorological inputs in Table 1, including future temperature and precipitation.

Your suggestion allows us to test whether this accuracy advantage persists with imperfect weather forecasts. We therefore conducted additional experiments using archived operational forecasts from the Global Forecast System (GFS) as future meteorological inputs. GFS is run by NCEP within NOAA’s National Weather Service. The available archive begins in January 2015, with a 0.25° grid and forecasts up to 16 days ahead.

The experiments covered five snow seasons during 2016–2021. We used GFS future temperature and precipitation across training, validation, and testing and trained SWECast from scratch. Historical meteorological inputs remained from ERA5-Land. All models received identical future weather forecasts and used the same test stations, dates, and forecast windows. We have added the setup and results to Appendix C.4, Table 9, reproduced below.

Table 9. SWE forecasting with GFS future meteorological inputs and 60 days of current-season observations. MAE/RMSE: mean ± sample SD (mm) across three runs; NSE: mean. Bold: best; ☆: second-best. The same notation applies to Table 10.

Model7d MAE ↓7d RMSE ↓7d NSE ↑14d MAE ↓14d RMSE ↓14d NSE ↑
Random Forest12.51±0.31☆18.73±0.210.971020.34±0.5931.83±1.420.9210
XGBoost12.75±0.4618.39±0.19☆0.974420.03±0.72☆30.10±1.66☆0.9333☆
DLinear57.38±2.4572.14±2.720.551563.00±3.4579.44±3.570.4817
PatchTST29.41±0.8639.59±0.490.876137.44±2.3451.99±3.100.7883
iTransformer20.93±0.9128.50±1.120.937127.81±1.2640.41±0.870.8729
TimesNet19.17±0.9426.80±2.040.941326.95±1.2139.20±1.750.8780
SWECast11.63±0.3217.23±0.480.9733☆18.62±0.5528.02±1.300.9378

The results show that SWECast achieves the lowest MAE and RMSE at both horizons with GFS inputs. Forecasting accuracy decreases relative to the ERA5-Land setting, consistent with the additional challenge of uncertain future meteorological inputs.

SWECast achieves better results than most baselines using ERA5-Land inputs across all three metrics at both horizons, even with imperfect GFS forecasts as future inputs. Its RMSE is numerically lower than that of every ERA5-Land baseline at both horizons. We have compiled this descriptive comparison in Appendix C.4, Table 10, reproduced below. These findings further support SWECast’s potential for practical SWE forecasting.

Table 10. SWECast using GFS forecasts compared with baselines using ERA5-Land reanalysis.

ModelFuture meteorology7d MAE ↓7d RMSE ↓7d NSE ↑14d MAE ↓14d RMSE ↓14d NSE ↑
Random ForestERA5-Land11.88±0.1619.69±0.510.96818.87±0.2532.57±0.320.922
XGBoostERA5-Land10.97±0.27☆17.68±0.710.975☆17.06±0.40☆28.31±0.530.942☆
DLinearERA5-Land43.64±2.1357.89±0.960.73954.82±1.8174.59±0.710.601
PatchTSTERA5-Land26.02±2.7935.45±3.280.89634.30±2.2751.00±2.520.809
iTransformerERA5-Land14.03±0.4721.64±0.780.96120.19±0.4232.37±0.350.917
TimesNetERA5-Land13.70±0.4721.04±1.260.96319.77±0.6130.97±0.450.923
ForeSWEERA5-Land17.82±2.0326.65±2.110.94124.86±3.3039.06±4.160.874
SNOW-17ERA5-Land75.38±1.26105.32±2.450.37776.38±2.44109.38±3.100.308
SNOW-17†ERA5-Land58.74±1.9781.75±4.130.63260.40±1.4484.20±2.090.568
SWECastERA5-Land10.10±0.2716.10±0.760.97814.82±0.3724.20±1.040.954
SWECastGFS11.63±0.3217.23±0.48☆0.973318.62±0.5528.02±1.30☆0.9378

† SNOW-17 calibrated using all available historical records at each station.

AuthorsOpenReview ↗

Response to Reviewer xjzG (2/3)

W2. Contribution of future meteorological inputs

“…how much of the forecast skill comes from future weather inputs and how much of the skill comes from the model design.”

Response:

Our original input study examined the contribution of meteorological information to SWE forecasting and the choice of input variables. We used the SWE-only setting as a reference and compared configurations that added temperature, precipitation, or both. This design allowed us to assess the contribution of each variable and whether their combination improved forecasting accuracy, as reported in Table 3.

Your suggestion has helped us extend this analysis by distinguishing the contribution of future weather from that of historical meteorological observations. We have added a separate test that removes only future weather inputs for SWECast and the baselines. We compared this setting with those using ERA5-Land and archived GFS forecasts as future inputs. All three settings used identical test stations, dates, and forecast windows. We have added the complete comparison to Appendix C.5, Table 11, reproduced below.

Table 11. Effects of removing future weather inputs. ↓: lower is better; ↑: higher is better. Bold: improvement over the same model’s w/o setting. w/o: no future weather inputs.

ModelMAE (mm) ↓RMSE (mm) ↓NSE ↑
ERA5-LandGFSw/oERA5-LandGFSw/oERA5-LandGFSw/o
Random Forest18.5820.0624.9432.7732.8041.230.9240.9170.885
XGBoost17.1419.5424.7928.8629.8640.630.9430.9330.890
DLinear53.4562.9949.1374.7881.0476.220.6030.4340.586
PatchTST35.8737.1033.7253.7051.4150.430.7940.7790.818
iTransformer19.7126.6727.0131.9740.6143.480.9200.8620.862
TimesNet19.0926.3530.0430.5438.4745.180.9290.8770.854
ForeSWE25.4829.8628.4039.9442.6544.980.8770.8450.856
SWECast14.4619.1523.0623.8329.2338.280.9580.9340.904

The results show:

  1. Most models benefit from future weather inputs, even when the weather forecasts contain errors. Most models improve across all three metrics with ERA5-Land future inputs. Even with imperfect archived weather forecasts, most models achieve higher forecasting accuracy than in the w/o setting. These results support our design of incorporating future meteorological information into the inputs.

  2. Effective use of future weather information depends on how models connect meteorological variables to SWE predictions. The original architectures of DLinear and PatchTST process variables independently, without explicit mechanisms for modeling interactions between them. This limits their ability to use future meteorological information for SWE forecasting. SWECast establishes these interactions through Cross-Channel Attention and weather gating, enabling meteorological conditions to inform SWE predictions. Its leading performance across the three input settings supports this design.

W3 & Q3. Comparison with a persistence baseline

“…Please include a persistence baseline.”

Response:

Our original comparisons with machine-learning models, time-series models, and SNOW-17 provided strong evidence of SWECast’s forecasting performance. Your suggestion has led us to add persistence, which predicts the last observed SWE value at every forecast step. We used the same station splits and forecast windows. We have added the complete comparison to Appendix C.6, Table 12, reproduced below.

Table 12. Results from Table 1 with persistence added. Three runs; MAE/RMSE: mean ± SD (mm); NSE: mean. Bold: best; ☆: second-best. † SNOW-17 calibrated on all historical records at each station.

Model7d MAE ↓7d RMSE ↓7d NSE ↑14d MAE ↓14d RMSE ↓14d NSE ↑
Persistence18.94±0.3032.93±0.560.91734.52±0.2757.12±0.740.754
Random Forest11.88±0.1619.69±0.510.96818.87±0.2532.57±0.320.922
XGBoost10.97±0.27☆17.68±0.71☆0.975☆17.06±0.40☆28.31±0.53☆0.942☆
DLinear43.64±2.1357.89±0.960.73954.82±1.8174.59±0.710.601
PatchTST26.02±2.7935.45±3.280.89634.30±2.2751.00±2.520.809
iTransformer14.03±0.4721.64±0.780.96120.19±0.4232.37±0.350.917
TimesNet13.70±0.4721.04±1.260.96319.77±0.6130.97±0.450.923
ForeSWE17.82±2.0326.65±2.110.94124.86±3.3039.06±4.160.874
SNOW-1775.38±1.26105.32±2.450.37776.38±2.44109.38±3.100.308
SNOW-17†58.74±1.9781.75±4.130.63260.40±1.4484.20±2.090.568
SWECast (Ours)10.10±0.2716.10±0.760.97814.82±0.3724.20±1.040.954

SWECast achieves the best results across all three metrics at both horizons. These results support its ability to forecast SWE evolution at stations excluded from training.

AuthorsOpenReview ↗

Response to Reviewer xjzG (3/3)

Q1. Definition and units of SWE

“…define SWE with a simple equation. Stating its units would also help…”

Response:

The introduction of our paper already described the physical meaning of SWE: “Snow water equivalent (SWE) is a measure of the water stored in snowpack and indicates subsequent snowmelt runoff.” Nevertheless, we have followed your suggestion by adding a simple equation and explicitly stating its units in Section 3.1:

SWE is the equivalent depth of water obtained by completely melting the snowpack over a unit area:

SWE=ρsρwhs,\mathrm{SWE} = \frac{\rho_s}{\rho_w} h_s,

where hsh_s is snow depth, ρs\rho_s is snow density, and ρw\rho_w is water density. SWE and hsh_s are expressed in millimeters (mm), with both densities expressed in the same units. One millimeter of SWE corresponds to one kilogram of water per square meter.

Q4. Geographic coverage and station splits

“…add a map showing the stations and mark which ones are used for training, validation, and testing.”

Response:

SWECast forecasts SWE at individual stations without modeling spatial relations between them. But as you pointed out, a map helps clarify the geographic coverage and station splits. We have therefore added a map to Appendix B.5, Figure 5 showing the training, validation, and test stations for the three splits generated with different random seeds in our repeated experiments.

Q5. Clarity of Figures 1 and 2

“…make the inputs, outputs, and connections between modules clearer.”

Response:

We have simplified Figure 1 to connect the two forecasting challenges with their corresponding solution ideas: historical priors for generalization to unseen stations and future meteorological information for anticipating weather changes. The updated architecture overview in Figure 2 clarifies the inputs, outputs, and connections between modules. The Prior Conditioner’s output now connects explicitly to the Prior Modulation Layer. We have also clarified the residual connection, removed redundant arrows, and labeled the final SWE forecast.

Reviewer_s77LOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a snow water equivalent (SWE) forecasting framework designed to address heterogeneous SWE evolution across locations and abrupt changes in meteorological conditions. The method retrieves similar segments from complete past snow seasons to construct historical priors and combines this cross-year information with adaptive weighting of current SWE and meteorological inputs in its forecasting architecture. Experiments are conducted to evaluated the proposed method using SNOTEL observations matched with ERA5-Land meteorological data and compared with a range of forecasting baselines.

Critical strengths
  1. The proposed method and model architecture are well aligned with the identified SWE forecasting challenges, combining historical-prior conditioning to account for heterogeneous SWE evolution with dynamic integration of SWE and meteorological information to respond to changing weather conditions.

  2. The proposed method achieves consistently competitive or superior performance across the reported experiments compared with a reasonably broad set of machine-learning, time-series, and snow-model baselines.

Critical weaknesses
  1. The evaluation uses ERA5-Land meteorological variables over the forecast horizon rather than clearly using weather forecasts available at prediction time. Since adapting to abrupt weather changes is a central motivation of the method, it remains unclear how the reported performance would translate to realistic forecasting when future meteorological inputs are uncertain.

  2. Although the paper compares against a broad set of generic machine-learning and time-series baselines and SNOW-17, it does not compare with recent SWE-specific data-driven forecasting approaches. This makes it difficult to assess the proposed method's improvement over the domain-specific state-of-the-art.

Questions for authors
  1. Could you clarify whether the meteorological inputs over the forecast horizon are weather forecasts available at prediction time or future ERA5-Land values? If the latter, how should the practical forecasting performance be interpreted under realistic meteorological forecast uncertainty?

  2. Could you discuss how the proposed method compares with recent SWE-specific data-driven forecasting approaches, beyond the generic time-series models and SNOW-17 considered in the experiments?

  3. Could you clarify whether the selection of temperature and total precipitation as meteorological inputs was performed using only the training/validation data, without using the test set?

3 comments
AuthorsOpenReview ↗

Response to Reviewer s77L (1/2)

Dear Reviewer s77L,

Thank you for your careful review and constructive suggestions. We appreciate your recognition of SWECast’s design and its consistently competitive or superior performance.

We have addressed your concerns about uncertain future meteorology, SWE-specific comparisons, and input selection through experiments using archived operational weather forecasts, an expanded comparison with ForeSWE, and validation results documenting input selection. Our responses to each point are presented in detail.

W1 & Q1. Evaluation with operational weather forecasts

“...how should the practical forecasting performance be interpreted under realistic meteorological forecast uncertainty?”

Response:

ERA5-Land provides 0.1° meteorological data throughout our 1993–2021 station record. Its long record and fine grid support evaluation across diverse snow seasons and local conditions, providing a suitable basis for comparing forecasting accuracy and generalization to unseen stations. We therefore used ERA5-Land reanalysis for all meteorological inputs in Table 1, including future temperature and precipitation.

Your suggestion provides a valuable test of whether this accuracy advantage persists with imperfect weather forecasts. We therefore conducted additional experiments using archived forecasts from the Global Forecast System (GFS), an operational weather forecasting system run by NCEP within NOAA’s National Weather Service. The archive begins in January 2015 and provides forecasts on a 0.25° grid up to 16 days ahead.

The additional experiments covered five snow seasons during 2016–2021. We used GFS future temperature and precipitation across the training, validation, and test sets and trained SWECast from scratch. We used the same future weather forecasts, test stations, dates, and forecast windows for all models in this comparison. We have added the results and experimental setup to Appendix C.4, Table 9, with the results reproduced below.

Table 9. GFS future meteorological inputs. MAE/RMSE: mean ± SD (mm) across three runs; NSE: mean. Bold: best; ☆: second-best. The same notation applies to Table 10.

Model7-day MAE ↓7-day RMSE ↓7-day NSE ↑14-day MAE ↓14-day RMSE ↓14-day NSE ↑
Random Forest12.51±0.31☆18.73±0.210.971020.34±0.5931.83±1.420.9210
XGBoost12.75±0.4618.39±0.19☆0.974420.03±0.72☆30.10±1.66☆0.9333☆
DLinear57.38±2.4572.14±2.720.551563.00±3.4579.44±3.570.4817
PatchTST29.41±0.8639.59±0.490.876137.44±2.3451.99±3.100.7883
iTransformer20.93±0.9128.50±1.120.937127.81±1.2640.41±0.870.8729
TimesNet19.17±0.9426.80±2.040.941326.95±1.2139.20±1.750.8780
SWECast11.63±0.3217.23±0.480.9733☆18.62±0.5528.02±1.300.9378

The results show that SWECast achieves the lowest MAE and RMSE at both forecast horizons. Relative to its ERA5-Land results in Table 1, its MAE increases from 10.10 to 11.63 mm at 7 days and from 14.82 to 18.62 mm at 14 days. This change is consistent with the additional challenge of uncertain future meteorological inputs.

SWECast achieves better results than most baselines using ERA5-Land inputs across all three metrics at both horizons, even with imperfect GFS forecasts as future inputs. Its RMSE is numerically lower than that of every ERA5-Land baseline at both horizons. We have compiled these results in Appendix C.4, Table 10 to make the comparison easier to follow. These findings further support SWECast’s potential for practical SWE forecasting.

Table 10. SWECast using GFS forecasts compared with baselines using ERA5-Land reanalysis. SWECast with ERA5-Land inputs is also shown for reference.

ModelFuture meteorology7-day MAE ↓7-day RMSE ↓7-day NSE ↑14-day MAE ↓14-day RMSE ↓14-day NSE ↑
Random ForestERA5-Land11.88±0.1619.69±0.510.96818.87±0.2532.57±0.320.922
XGBoostERA5-Land10.97±0.27☆17.68±0.710.975☆17.06±0.40☆28.31±0.530.942☆
DLinearERA5-Land43.64±2.1357.89±0.960.73954.82±1.8174.59±0.710.601
PatchTSTERA5-Land26.02±2.7935.45±3.280.89634.30±2.2751.00±2.520.809
iTransformerERA5-Land14.03±0.4721.64±0.780.96120.19±0.4232.37±0.350.917
TimesNetERA5-Land13.70±0.4721.04±1.260.96319.77±0.6130.97±0.450.923
ForeSWEERA5-Land17.82±2.0326.65±2.110.94124.86±3.3039.06±4.160.874
SNOW-17ERA5-Land75.38±1.26105.32±2.450.37776.38±2.44109.38±3.100.308
SNOW-17†ERA5-Land58.74±1.9781.75±4.130.63260.40±1.4484.20±2.090.568
SWECastERA5-Land10.10±0.2716.10±0.760.97814.82±0.3724.20±1.040.954
SWECastGFS11.63±0.3217.23±0.48☆0.973318.62±0.5528.02±1.30☆0.9378

†SNOW-17 calibration uses all available historical records at each station.

AuthorsOpenReview ↗

Response to Reviewer s77L (2/2)

W2 & Q2. Comparison with data-driven SWE forecasting models

“…how the proposed method compares with recent SWE-specific data-driven forecasting approaches…”

Response:

Our original comparison included representative methods from machine learning, time series forecasting, and process-based snow modeling. This coverage provided a broad benchmark for assessing SWECast’s forecasting accuracy and generalization to unseen stations. The original results established its consistent advantage across these complementary approaches.

In response to your suggestion, we have conducted additional experiments with the SWE forecasting model ForeSWE (Thapa et al., 2026) and updated Table 1 in the main text with the results. We have documented its implementation alongside those of the other baselines in Appendix B.2. We present the updated table below for your convenience.

Table 1. SWE forecasting performance using the preceding 60 days of observations from the current snow season. Results are averaged over three independent runs with different random seeds. MAE and RMSE are reported as mean ± standard deviation across runs (mm), and NSE as the mean. ↓ indicates lower is better; ↑ indicates higher is better. Best results are shown in bold, and second-best results are marked with ☆.

Model7d MAE ↓7d RMSE ↓7d NSE ↑14d MAE ↓14d RMSE ↓14d NSE ↑
Random Forest11.88±0.1619.69±0.510.96818.87±0.2532.57±0.320.922
XGBoost10.97±0.27☆17.68±0.71☆0.975☆17.06±0.40☆28.31±0.53☆0.942☆
DLinear43.64±2.1357.89±0.960.73954.82±1.8174.59±0.710.601
PatchTST26.02±2.7935.45±3.280.89634.30±2.2751.00±2.520.809
iTransformer14.03±0.4721.64±0.780.96120.19±0.4232.37±0.350.917
TimesNet13.70±0.4721.04±1.260.96319.77±0.6130.97±0.450.923
ForeSWE17.82±2.0326.65±2.110.94124.86±3.3039.06±4.160.874
SNOW-1775.38±1.26105.32±2.450.37776.38±2.44109.38±3.100.308
SNOW-17†58.74±1.9781.75±4.130.63260.40±1.4484.20±2.090.568
SWECast (Ours)10.10±0.2716.10±0.760.97814.82±0.3724.20±1.040.954

† SNOW-17 calibrated using all available historical records at each station.

SWECast achieves the best results across all three metrics at both forecast horizons in the expanded comparison. Relative to ForeSWE, it reduces MAE by 43.3% at 7 days and 40.4% at 14 days. This additional evidence extends the demonstrated accuracy advantage to a recent model developed specifically for SWE forecasting.

Q3. Selection of meteorological inputs

“…using only the training/validation data, without using the test set?”

Response:

We selected the meteorological inputs based on validation MAE among physically motivated candidate combinations, and then evaluated their effectiveness on the test set, as reported in Table 3. We have clarified this procedure in Section 5.5 of the revised manuscript and added the results obtained on the validation set to Appendix C.2, Table 6, reproduced below. The combination of temperature and total precipitation achieves the lowest validation MAE, supporting our choice of meteorological inputs.

Table 6. Validation MAE for meteorological input selection using 60 days of current-season observations and a 14-day forecast horizon. The candidate configurations correspond to those in Table 3 of the main paper. MAE is in mm; the best result is in bold. ↓ indicates that lower values are better. Tm: temperature; TP: total precipitation.

InputValidation MAE ↓ (mm)
SWE only28.50
SWE + Tm27.33
SWE + TP21.24
SWE + Tm + TP (Base)16.54
Base + Snow16.75
Base + Surface16.77
Base + Atmospheric18.18
Base + Vegetation16.94
AuthorsOpenReview ↗

Looking Forward to Your Feedback

Dear Reviewer s77L,

We believe that we have addressed all your concerns in our rebuttal. As the discussion period is coming to an end, we would appreciate it if you could review our responses. We would be happy to discuss any remaining questions. Thank you for your time and consideration.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Brief summary of the discussion

Dear Reviewers, Area Chair, Senior Area Chair, and Program Chairs,

Thank you for the time and care you have devoted to reviewing our submission.

Our work began with the observation that patterns of snow accumulation and melt can recur at the same station across snow seasons, even when their timing and pace differ. SWECast retrieves a matching period from the station’s preceding snow season and uses the subsequent observations to guide the forecast. A weather gate adjusts the model’s response to forecast temperature and precipitation. This enables a shared model to forecast snow water equivalent (SWE) at monitored stations excluded from training. SWECast retained its forecasting advantage at these stations even when using imperfect operational weather forecasts.

As the rebuttal period draws to a close, we would like to offer a brief summary of the discussion and the revisions we have made in response to the reviewers’ feedback.

1. Forecasting with uncertain future weather. All three reviewers asked whether SWECast’s accuracy advantage would hold under realistic weather forecast uncertainty. To address this concern, we evaluated SWECast and the baselines using archived GFS forecasts of temperature and total precipitation as future inputs. SWECast achieved the best overall forecasting performance among the models evaluated with GFS inputs. Even with these imperfect weather forecasts, it outperformed most baselines using ERA5-Land inputs across all reported metrics at both horizons. These findings show that SWECast can produce accurate SWE forecasts using weather information available at forecast time, supporting its potential for operational use.

2. Additional baselines. Reviewers s77L and Lnbd asked for a comparison with a recent model developed specifically for SWE forecasting. Reviewer xjzG requested a persistence baseline and later suggested including a recent pretrained forecasting model. We have expanded our comparison to include ForeSWE, persistence, and Chronos-2. SWECast achieved the best overall forecasting performance in the expanded comparison, providing stronger evidence for its accuracy across a broader range of forecasting approaches.

3. Model design. Reviewer Lnbd asked what SWECast contributes beyond its individual components. We clarified how SWECast retrieves a matching segment from the forecast station’s preceding snow season and uses the resulting historical prior to modulate the shared model’s forecast. The component ablation showed that removing this modulation increased forecasting error. Reviewer xjzG also asked how much future weather contributes to forecast accuracy. We added an experiment that removed only future weather inputs while retaining historical observations. Most models benefited from future weather information, and SWECast performed best under each input setting.

4. Input Selection and Configuration. Reviewer s77L asked whether the test set influenced the selection of meteorological inputs, and Reviewer Lnbd raised the same question about history lengths. We clarified that we selected the meteorological inputs, current-season history length, and number of complete historical snow seasons exclusively based on validation MAE. The revised manuscript provides the corresponding validation results in the appendix and reports performance on the test set in the main text.

5. Discussion outcomes. We also addressed Reviewer xjzG’s requests to define SWE and its units, map the station splits, and clarify the model diagrams. After our initial response, xjzG wrote, “It addressed most of my concerns.” Alongside reviewer xjzG suggested including a recent pretrained model and added that they were curious why Random Forest and XGBoost performed so well on this task. Our additional experiments showed that adding temperature and total precipitation to recent SWE observations improved forecasting accuracy for both Random Forest and XGBoost. We also examined periods of rapid SWE change and found that SWECast produced more accurate forecasts than either tree-based model. We have not received further feedback on this latest response. Reviewers s77L and Lnbd have not posted further written comments, although one of them has raised their rating.

Thank you for your time and consideration.

Best regards,

The authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.