{"id":"643c43bc-d7a9-46b2-9d47-f0a8dd4868c0","arxiv_id":"2504.20129","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"An LSTM trained on ERA5 forcings and SNOTEL measurements estimates snow presence and snow water equivalent with above 93 percent accuracy and roughly 0.9 correlation, outperforming ERA5 at SNOTEL sites.","lead":"A NASA-affiliated team trains an LSTM neural network to estimate daily snow water equivalent across the continental US from weather and terrain data, splitting the task into snow-presence classification and snow-depth regression. The model reports about 93 to 95 percent classification accuracy and correlations near 0.9 on held-out SNOTEL sites, beating ERA5 reanalysis at the same locations.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Weakest point is that the headline generalization claim rests on the primary model using ERA5 snow-state inputs, making the bias-correction reading viable; the appendix no-snow-input result, which would settle the distinction, is only partially reported.","rationale":"The paper is a plausible, well-designed application of an established LSTM architecture, and the reader's conditional verdict is appropriate. The strongest claim, generalizing to previously unseen data and outperforming ERA5 while learning physical associations, depends on whether the model's success relies on direct snow-state inputs. The reader's weakest_assumption matched mine closely: both center on the use of ERA5 snow variables vs. genuinely independent forcings, and both flag the spatiotemporal generalization caveat and numeric inconsistencies. My contribution narrows the concern to a precise, falsifiable requirement: the appendix's no-snow-input experiment is the decisive one, yet it is reported only partially (classification table plus a qualitative figure, with no regression metrics). That missing evidence is load-bearing because without it the strong independent-estimation reading cannot be checked. The numeric discrepancies (Table 5 ERA5 recall mean 0.265 vs min 0.790; Discussion numbers not matching Table 3/Fig. 4) are secondary but strengthen the need for a conditional verdict rather than an unconditional one. The appropriate verdict remains CONDITIONAL: accept once the no-snow-input regression metrics are reported, the tables are reconciled, and either code or detailed reproducibility notes are released.","tokens_in":941,"tokens_out":952,"duration_ms":21354,"concrete_test":"Add the missing LSTM-est regression statistics for the Sec. 4.1 test protocol (MAE, MdAE, bias, Pearson-r, and a scatter plot with the same y-axis limits as Fig. 4), and re-run the original model with only the non-snow ERA5 variables to produce a side-by-side numbers table. If LSTM-est regression metrics are within, say, 10-15% of the LSTM-with-snow-inputs metrics and the Pearson-r remains near 0.85, the learning-physical-relationships reading holds. If the gap is large, the headline claim should be revised to describe a snow-aware bias-correction model.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that a physically driven LSTM generalizes to previously unseen data and outperforms ERA5 because it learns physical and morphological relationships (Abstract; Sec. 6). But the primary experiments include ERA5 sd, csfr, es, and smlt as inputs (Table 1; Sec. 3.1). With sd being the ERA5 SWE-equivalent and the other three being snow-process fluxes, a model that refines these inputs against SNOTEL targets can score well by learning a station-level bias correction rather than by learning snow physics from weather and static attributes. The paper itself concedes this interpretation in Appendix-A: without ERA5 snow variables the model would only be a bias corrector, and the appendix is the only experiment that tests whether non-snow forcings are sufficient. That appendix is underreported: Table 7 gives classification metrics but the regression comparison is only a qualitative figure (Fig. 11), with no table of MAE, MdAE, bias, or Pearson-r for LSTM-est, so the claim that no significant performance degradation occurs is not backed by reported numbers. Additionally, Table 5 has an apparent typo: the ERA5 recall column lists 0.265 as the mean, a value below the stated minimum of 0.790; the discussion also quotes test-set values (e.g., accuracy 94.49%, Pearson-r 0.868) that do not appear verbatim in Table 3 or Fig. 4, making the public-facing numbers hard to verify from the manuscript alone. The spatiotemporal generalization result is materially weaker (mean Pearson-r 0.669 vs 0.866 temporal), and the manuscript itself notes that no entirely novel biogeographic regions were tested; therefore the claim of spatial generalization to previously unseen regions is stronger than the evidence supports.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript presents an LSTM-based model that estimates daily snow water equivalent (SWE) at SNOTEL stations by first classifying snow presence and then regressing SWE on snow-present days. The model uses ERA5 meteorological time series and static physiographic attributes as inputs, with a 365-day lookback and a static-repetition scheme. The authors compare the LSTM against ERA5's snow-depth variable on a temporal test set, across 11 temporal cross-validation folds, and across a 33-fold spatiotemporal cross-validation, reporting that the LSTM consistently achieves higher classification accuracy (about 93-95%) and higher SWE correlation (around 0.9) than ERA5. An appendix removes the ERA5 snow-state variables to test whether the model can learn snowpack dynamics from non-snow forcings alone.","tokens_in":16982,"tokens_out":4768,"duration_ms":46645,"significance":"If the results are confirmed, the model would be a fast, point-scale SWE estimator that could serve as an LSM surrogate and provide useful snow products at SNOTEL locations. The experimental design has real strengths: the classification/regression decomposition is sensible for sparse SWE data, temporal buffers are used to reduce leakage, evaluation is against external in situ SNOTEL data, and the appendix ablation addresses an important confound. However, the headline comparison is confounded because the primary model ingests ERA5 snow-depth and snow-flux variables while being benchmarked against ERA5's snow-depth product, and the only experiment that controls for this is underreported. The significance of the paper is therefore conditional on completing that ablation and reconciling several numerical inconsistencies.","major_comments":[{"comment":"The only experiment that distinguishes \"learning snow physics from weather and static inputs\" from \"bias-correcting ERA5 snow fields\" is reported only partially. Table 7 gives classification metrics for LSTM-est, but the regression comparison is confined to Fig. 11, with no tabulated MAE, MdAE, bias, or Pearson-r for the snow-excluded model. The statement that \"there appears to be no significant performance degradation\" is therefore not supported by reported numbers. Please add a table of regression metrics for LSTM-est on the 2021 test set and, ideally, for the temporal and spatiotemporal cross-validation folds.","section":"Appendix-A"},{"comment":"The temporal cross-validation protocol as written is internally inconsistent. For the example fold with test year 2015, the training set is described as [01/01/2010–12/31/2013, 01/01/2016–12/31/2021], yet the text says the year after the test year is masked out; 2016 is the year immediately after 2015. Please clarify whether the after-year mask is applied and specify the training intervals for boundary folds such as test year 2011.","section":"Sec. 3.3.2"},{"comment":"Table 5 reports the mean ERA5 recall across the 33 spatiotemporal splits as 0.265, which is below the reported minimum of 0.790 for the same column. This is arithmetically impossible and indicates an error in the aggregated row; please recompute the ensemble statistics.","section":"Table 5"},{"comment":"Several test-set numbers quoted in the Discussion do not appear verbatim in the tables. The LSTM test accuracy is quoted as 94.49% whereas Table 3 gives 0.949; ERA5 accuracy is quoted as 89.10% whereas Table 3 gives 0.894; and the regression values (LSTM MAE 0.101, MdAE 0.06, Pearson-r 0.868; ERA5 MAE 0.204, MdAE 0.114, Pearson-r 0.391) are not tabulated anywhere. Please reconcile the text with the tables or add a regression-results table so that the public numbers can be verified.","section":"Sec. 5"},{"comment":"The conclusion states that the model \"generalizes effectively to previously unseen data\" without the caveat that the spatiotemporal test (Sec. 3.3.3) only holds out stations within the same western-US SNOTEL network. Since Sec. 5 explicitly acknowledges that the design does not stress-test novel biogeographic regions, the abstract and conclusions should qualify the spatial-generalization claim accordingly.","section":"Sec. 6"}],"minor_comments":[{"comment":"The qualitative snow-map comparison in Fig. 10 would be more informative with a quantitative map-error metric or at least an explicit color bar; as presented it is illustrative only.","section":"Sec. 4.4B"},{"comment":"The regression metrics are computed on snow-present days only, but this is not stated in the metrics section; please make the evaluation domain explicit there.","section":"Sec. 3.3.4"},{"comment":"The phrase \"physiological factors\" in the introduction appears to mean \"physiographic\"; please correct the terminology.","section":"Sec. 1"},{"comment":"No code or data availability statement is provided; adding one would materially improve reproducibility, especially for the cross-validation splits.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper's central claim—that the LSTM learns physical relationships rather than simply bias-correcting ERA5 snow fields—rests on an appendix that is reported incompletely. I would not reject the manuscript if the authors supply the missing regression metrics for the snow-excluded model, correct the Table 5 inconsistency, and reconcile the Discussion numbers with the tables. The research question is appropriate for the journal, and the experimental scaffolding is otherwise careful."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is a competent, incremental extension of the LSTM-SWE idea from Pflug et al., and the cross-validation design is genuinely careful. But the abstract and conclusions sell it as a 'physically driven' estimator that generalizes to unseen regions, when the default model gets a big assist from ERA5 snow-state variables as inputs. You can read the main result as a station-level bias correction of ERA5, not independent physics. That reading is exactly why Appendix A exists, and the paper is honest about it—but the appendix is where it falls short.\n\nWhat's new and good: breaking the problem into snow-presence classification and SWE regression, dropping the MODIS fSCA priors, and adding a large static feature set. The temporal and spatiotemporal CV designs, with buffer years and spatial splits, are better than most of what I see in this literature. The appendix ablation, even partially reported, suggests non-snow forcings do carry enough signal—that would be the genuinely interesting result if fully supported.\n\nSoft spots, in rough order of severity:\n\n1. The no-snow-input experiment is underreported. Table 7 gives classification only; the regression comparison is a qualitative figure with no MAE/MdAE/bias/r values. The text claims 'nearly identical performance,' but there are no numbers to back it. If the authors make this experiment primary and report the full metric table, the bias-correction concern largely dissolves.\n\n2. Table 5 has an obvious typo: the ERA5 recall mean is listed as 0.265, below the stated minimum of 0.790. That's the kind of error that makes a reader doubt the rest of the tables.\n\n3. The Discussion quotes test-set numbers (accuracy 94.49%, Pearson-r 0.868) that I could not verify in Table 3 or Fig. 4. Not necessarily wrong, but hard to check.\n\n4. The 'generalize to previously unseen regions' claim is stronger than the evidence. The spatiotemporal CV gets mean r ≈ 0.67, well below the temporal 0.87, and it only moves stations between splits within the same western US mountain network. The paper itself admits no truly novel bioregions were tested.\n\nWho should read it: anyone building ML surrogates for snow or land-surface models, and anyone thinking about what 'physically driven' means when reanalysis products are in the input set. It deserves a serious referee. I'd send it out, but I'd insist the appendix result be elevated to a main result with full metrics, and the table/discussion numbers reconciled.","headline":"A solid but oversold SWE surrogate: the main experiments are confounded by ERA5 snow inputs, and the appendix that could settle the 'physically driven' claim is underreported.","tokens_in":17567,"tokens_out":2912,"would_cite":false,"duration_ms":28458,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"An LSTM snow model beats ERA5 at estimating snow water equivalent.","keywords":["snow water equivalent","long short-term memory","ERA5 reanalysis","SNOTEL snow pillows","snow-presence classification","spatiotemporal cross-validation","land surface model surrogate","zero-inflated regression"],"falsifier":"Hold out one complete mountain range or biogeographic region absent from training (for example, the Alaska SNOTEL stations or a Sierra-only split) and compare the model's classification accuracy, MAE, and correlation with ERA5's; if the gains over ERA5 shrink to near zero in that region, the generalization claim fails. A complementary check is to compare predicted SWE at stations whose elevation or aspect lie far outside the training distribution.","tokens_in":16456,"feed_emoji":"❄️","tokens_out":6952,"duration_ms":71184,"temperature":0.7,"pith_summary":"Seasonal snow is a key land-surface variable, but process-based reanalysis is expensive and coarse while in situ snow pillows are sparse. The paper proposes a Long Short-Term Memory (LSTM) network that, given a year of daily ERA5 meteorological fields plus static soil, topographic, irrigation, and vegetation attributes, first classifies whether snow is present and then regresses snow water equivalent where snow exists. Trained on quality-controlled SNOTEL pillow measurements at western United States stations, the model reports snow-presence accuracy above 93 percent and SWE correlation near 0.9, with larger gains over ERA5 in deep-snow situations. It generalizes to held-out years and to held-out station groups in spatiotemporal cross-validation. An ablation in an appendix indicates the model does not depend on ERA5's snow-related variables for this performance.","feed_headline":"An LSTM snow model beats ERA5 at estimating snow water equivalent","feed_subtitle":"Trained on a year of weather plus terrain, it classifies snow days at ~94% accuracy and tracks SWE at ~0.9 correlation.","key_machinery":"The carrying mechanism is a two-headed LSTM with a one-year daily lookback: 256 LSTM cells receive the dynamic meteorological series and the static attributes through a static-repetition scheme at each time step, and the shared hidden state feeds a sigmoid snow-presence head and a two-layer regression head. The regression head is trained with a zero-masked root-mean-square error so that snow-free days contribute nothing to the loss, while the classification head uses cross-entropy, and the full loss weights classification at 0.1. The model multiplies its regression output by the rounded classification output, so predicted SWE is exactly zero on classified snow-free days. This architecture's role is to disentangle the binary presence of snow from its depth, letting the model handle a highly sparse, zero-heavy SWE distribution without a satellite snow prior.","core_discovery":"The paper's central claim is that a single LSTM, with a 365-day lookback and static physiographic inputs repeated at every time step, can learn the local physics of snow accumulation and ablation well enough to beat the ERA5 reanalysis at point scale. In both the temporal holdout test and the spatiotemporal cross-validation, the model's snow-presence classification accuracy, precision, and recall are consistently above those of ERA5's snow-depth variable, and its regression errors are roughly half as large. The paper explains this with the two-task structure: a separate snow-presence classifier and a zero-masked regression loss let the model treat snow-free days correctly instead of including them as noisy near-zero targets. The ablation shows that removing ERA5 snow depth, snowfall, snow evaporation, and snow melt from the inputs leaves performance essentially unchanged, suggesting the model is learning the relation between meteorological and physiographic drivers and snow rather than merely bias-correcting ERA5's snow field.","pith_inferences":["If the mapping transfers to regions outside the western United States, the same architecture could serve as an inexpensive, globally deployable SWE surrogate for water-resource and land-surface modeling; the paper itself tests only within the western US mountain network.","The zero-masked regression loss and static-repetition design could be adapted to estimate other zero-inflated hydrologic states, such as soil moisture or snow depth, where static controls dominate.","A direct testable extension would be to train the model on the western US and evaluate it on independently measured snow courses or on Alaska SNOTEL stations; a large performance drop there would bound how much of the reported generalization is physical transfer versus overfitting to station geography.","A natural next test is to quantify performance during extreme snow events, since the paper's qualitative maps suggest the model captures such events more sharply than ERA5."],"forward_implications":["A trained model can produce daily SWE estimates at Continental US stations at a small fraction of reanalysis computational cost, with point accuracy above ERA5.","Because the model uses no snow priors or satellite snow-cover inputs, it can be used for forecasting and scenario runs where future snow cover itself is unknown.","The two-task decomposition of presence-then-depth is a transferable design for other sparse, zero-inflated geophysical variables.","The spatiotemporal cross-validation results suggest the learned mapping captures at least part of the terrain and climate signal, though performance drops slightly when both space and time are held out.","The snow-excluded ablation implies the model is estimating snow from physical drivers rather than copying ERA5 snow outputs."],"supporting_citations":[{"why":"Supplies the daily ERA5 time-series forcings (temperature, precipitation, radiation, and related variables) that drive the model.","marker":"[11]"},{"why":"Provides the in situ SNOTEL snow-pillow SWE observations used as training and test targets.","marker":"[25]"},{"why":"The recent lightweight LSTM for SWE whose approach this paper extends by omitting satellite snow priors.","marker":"[19]"},{"why":"The western United States snow reanalysis used as the SWE target in related work, motivating the need for cheaper ML surrogates.","marker":"[14]"},{"why":"Supplies the three-stage quality-control procedure applied to raw SNOTEL data before use.","marker":"[30]"},{"why":"Supplies the temperature and precipitation bias correction that produces the bias-corrected, quality-controlled SNOTEL dataset.","marker":"[31]"},{"why":"Supplies the static-repetition scheme for feeding static physiographic variables into each LSTM time step.","marker":"[33, 34]"},{"why":"Supplies soil texture fractions used as static inputs describing surface properties.","marker":"[26]"}],"fun_headline_variants":["LSTM snow model beats ERA5 with 93% accuracy","Two-task LSTM estimates SWE beating ERA5","Snow water equivalent LSTM beats ERA5","LSTM learns snow physics, beats ERA5 SWE","Snow presence and depth predicted by LSTM, tops ERA5"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim rests on the assumption that daily ERA5 grid-cell weather and coarsely downsampled static attributes contain enough local information to predict snow at the exact point of a SNOTEL tower; if sub-grid terrain, aspect, canopy, and wind redistribution matter more than these inputs capture, the learned mapping will not transfer to genuinely new locations.","fun_headline_variants_meta":{"raw":{"variants":["LSTM snow model beats ERA5 with 93% accuracy","Two-task LSTM estimates SWE beating ERA5","Snow water equivalent LSTM beats ERA5","LSTM learns snow physics, beats ERA5 SWE","Snow presence and depth predicted by LSTM, tops ERA5"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000651,"raw_usage":{"total_tokens":3021,"prompt_tokens":1018,"completion_tokens":2003,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":634,"completion_tokens_details":{"reasoning_tokens":1924}},"tokens_in":634,"tokens_out":2003,"duration_ms":14636,"temperature":1.0,"reasoning_tokens":1924,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:37:59.089341+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Hold out one complete mountain range or biogeographic region absent from training (for example, the Alaska SNOTEL stations or a Sierra-only split) and compare the model's classification accuracy, MAE, and correlation with ERA5's; if the gains over ERA5 shrink to near zero in that region, the generalization claim fails. A complementary check is to compare predicted SWE at stations whose elevation or aspect lie far outside the training distribution.","supporting_citations":[{"cited_title":"The ERA5 global reanalysis","cited_arxiv_id":null,"evidence_quote":"Supplies the daily ERA5 time-series forcings (temperature, precipitation, radiation, and related variables) that drive the model."},{"cited_title":"Characteristics of the western United States snowpack from snowpack telemetry (SNOTEL) data","cited_arxiv_id":null,"evidence_quote":"Provides the in situ SNOTEL snow-pillow SWE observations used as training and test targets."},{"cited_title":"Lightweight and Regionally Transferrable Snow Water Equivalent Estimation Using a Long Short-Term Memory Network","cited_arxiv_id":null,"evidence_quote":"The recent lightweight LSTM for SWE whose approach this paper extends by omitting satellite snow priors."},{"cited_title":"A western United States snow reanalysis dataset over the Landsat era from water years 1985 to 2021","cited_arxiv_id":null,"evidence_quote":"The western United States snow reanalysis used as the SWE target in related work, motivating the need for cheaper ML surrogates."},{"cited_title":"Next-generation intensity- duration-frequency curves for hydrologic design in snow-dominated environments","cited_arxiv_id":null,"evidence_quote":"Supplies the three-stage quality-control procedure applied to raw SNOTEL data before use."},{"cited_title":"Regional snow parameters estimation for large-domain hydrological applications in the western United States","cited_arxiv_id":null,"evidence_quote":"Supplies the temperature and precipitation bias correction that produces the bias-corrected, quality-controlled SNOTEL dataset."},{"cited_title":"The global land data assimilation system","cited_arxiv_id":null,"evidence_quote":"Supplies soil texture fractions used as static inputs describing surface properties."}],"review_version":1}