the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
Better data or better architecture? Improving deep-learning-based prediction in ungauged basins
Abstract. Large-sample hydrology has recently been driven by two key developments. First, the introduction of hydrological benchmark datasets such as CAMELS-US and CARAVAN, and second, the emergence of deep‑learning modelling frameworks, particularly LSTM‑based regional models, which have demonstrated performance on par with, and in some cases exceeding, that of process-based models for streamflow prediction in gauged and ungauged settings. Building on these developments, we investigate whether (i) further enhanced LSTM architectures, (ii) new sets of static features, or (iii) a combination of both enable us to significantly improve Predictions in Ungauged Basins (PUB). In this study, we evaluate a state-of-the-art regional LSTM model (base LSTM) against embedded (EMB-LSTM) and cross‑attention enhanced (CA-LSTM) variants, in combination with a suite of newly applied static features, namely MODIS surface reflectance bands, ALPHAEARTH embeddings, DEM-, meteorology- and catchment coordinate-derived auxiliary aggregates, and conventional CAMELS attributes. We tested these model-and-data combinations in pseudo‑ungauged 5‑fold cross‑validation across the 531 CAMELS‑US catchments. Model performance was quantified by the Nash‑Sutcliffe Efficiency (NSE), while latent‑space complexity was assessed via the Shannon effective rank (erank). Results show that the quality of static features is more important than architectural improvements. ALPHAEARTH embeddings attained the highest median NSE, but only in combination with auxiliary static feature data (ALPHAEARTHplus). Architectural refinements yielded only modest improvements. Thereby the relatively simple EMB-LSTM, which allowed the LSTM layer to better ingest ALPHAEAERTHplus static features, outperformed the other architectures. With this combination, we achieved a median performance of NSE = 0.726, significantly improving the state-of-the-art PUB performance (NSE = 0.69) for the CAMELS-US dataset. Auxiliary analysis indicates that further improvement is possible when adding MODIS bands as additional dynamic features to the model. In conclusion, our study indicates that, broadly speaking, (a) better data is more important than better architecture, (b) better architecture is necessary only to accommodate better data, (c) the single layer LSTM remains the most suitable core model as of now, and (d) the Shannon effective rank complexity of the latent space is a useful diagnostic for linking improved PUB performance to improved quality of latent hydrological representation inside the model. Overall, this highlights the need for improved measurement‑derived descriptor datasets, especially for soil and geology.
Competing interests: At least one of the (co-)authors is a member of the editorial board of Hydrology and Earth System Sciences.
Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this paper. While Copernicus Publications makes every effort to include appropriate place names, the final responsibility lies with the authors. Views expressed in the text are those of the authors and do not necessarily reflect the views of the publisher.- Preprint
(1014 KB) - Metadata XML
- BibTeX
- EndNote
Status: open (until 18 Nov 2026)
-
CC1: 'Comment on egusphere-2026-1965', John Ding, 25 May 2026
reply
The comment was uploaded in the form of a supplement: https://egusphere.copernicus.org/preprints/2026/egusphere-2026-1965/egusphere-2026-1965-CC1-supplement.pdfReplyCitation: https://doi.org/
10.5194/egusphere-2026-1965-CC1 -
RC1: 'Comment on egusphere-2026-1965', Anonymous Referee #1, 20 Oct 2026
reply
General comments
I found the manuscript very well written and enjoyed reading it. The research questions are interesting and relevant to prediction in ungauged basins. In my view, the methodological contribution is clear, while the hydrological insights should be developed further before publication.
The comparison of conventional catchment attributes with alternative descriptors is interesting. However, the manuscript currently provides stronger evidence for differences in predictive performance than for the hydrological explanations proposed for those differences. In particular, improvements in NSE and changes in latent-space dimensionality do not, by themselves, establish that a model has learned a more realistic representation of catchment functioning. The discussion would also benefit from examining where improvements occur. Identifying the climatic and physiographic conditions under which the new descriptors are useful would provide more hydrological insight than reporting improvements primarily through a median score across all catchments.
Finally, the broader question of how to select and aggregate hydrologically meaningful catchment descriptors is not entirely new. The manuscript would benefit from a stronger connection to previous work on catchment characterization (e.g. https://doi.org/10.1016/j.earscirev.2024.104739).
Specific comments
- Section 2.3.1 mainly explains the choice of source datasets, rather than the choice of the specific descriptors derived from them. This is especially relevant for the topographic descriptors. Why were descriptors of slope, drainage organization, or landscape connectivity not considered?
- Line 509: How do the authors explain the finding that higher-elevation catchments appear simpler for the LSTM to represent? To me, this finding is not immediately intuitive.
- Lines 511-520: The explanations proposed for arid and densely vegetated catchments are plausible but would benefit from further support. Low effective rank in arid catchments is interpreted as insufficient representation, whereas higher rank in vegetated catchments is interpreted as greater physical complexity. What independent evidence supports these different interpretations?
- Please reconcile the best NSE reported in the abstract (0.726) with Table 3 and the conclusions (0.721). Please also check the statements concerning statistical significance (p=0.061 and p=0.072 do not meet the 5% significance threshold).
Citation: https://doi.org/10.5194/egusphere-2026-1965-RC1 -
RC2: 'Comment on egusphere-2026-1965', Anonymous Referee #2, 02 Nov 2026
reply
Review Comments on the Manuscript
Manuscript title : Better data or better architecture? Improving deep-learning-based prediction in ungauged basins
General Comment
The manuscript addresses an important problem in hydrological modelling by investigating the relative contributions of environmental input data and deep-learning architecture to streamflow prediction in ungauged basins. The comparison of conventional catchment descriptors, satellite-derived embeddings, and alternative LSTM architectures is potentially valuable. However, several methodological details require clarification, and some conclusions appear stronger than the evidence presented. The following concerns should be addressed to improve transparency, reproducibility, and scientific robustness.
Major Comment 1: Unclear Treatment and Normalization of Observed Discharge
The manuscript identifies observed daily streamflow from CAMELS-US as the target variable but does not clearly specify whether model training was conducted using volumetric discharge (m³/s) or catchment-area-normalized runoff depth (mm/day).
This distinction is particularly important because the study considers 531 catchments with potentially substantial differences in drainage area. Volumetric discharge depends strongly on catchment size, and using raw discharge without appropriate treatment of catchment area may influence regional model training and generalization. The units of the observed streamflow are used during training.
Without this information, it is difficult to fully assess the hydrological consistency and reproducibility of the reported predictions.
Major Comment 2: Insufficient Evaluation of Catchment Area Effects:
Catchment area is a fundamental hydrological characteristic controlling total runoff volume. Its importance becomes particularly relevant when transferring rainfall–runoff relationships among watersheds of substantially different sizes. Author need to discuss about impact of catchment area on results, How it was used in training (Units). Because catchment area impacts on discharge, it cannot be ignored.Comment 4: Revision of the Manuscript Title
The current title, “Better data or better architecture? Improving deep-learning-based prediction in ungauged basins,” is relatively generic and does not adequately communicate the specific methodological scope and scientific contribution of the study. It is suggested to revise the title, such as Assessing Satellite-Derived Catchment Attributes and LSTM Architectures for Ungauged Streamflow Prediction: Evidence from CAMELS-US
Comment 5: Clarification of Training and Testing Periods
Section 3.2 reports the training period as 1 October 1999 to 30 September 2008 and the testing period from 1 October 1989 to 30 October 1999. As written, these periods overlap during October 1999. The authors should verify whether the testing end date is a typographical error and clarify the precise temporal partitioning
Overall Recommendation
The study addresses a relevant research question and provides potentially useful evidence regarding the role of satellite-derived environmental descriptors in deep-learning-based streamflow prediction. Nevertheless, clarification of discharge units and normalization, further analysis of catchment-area effects, correction of numerical inconsistencies, and more cautious interpretation of statistical significance are necessary.
Citation: https://doi.org/10.5194/egusphere-2026-1965-RC2 -
RC3: 'Comment on egusphere-2026-1965', Anonymous Referee #3, 09 Nov 2026
reply
General comments
The authors investigate whether better data or a better architecture matters more for improving LSTM predictions in ungauged basins. They test several combinations of model architectures and static predictors. They show that the model’s predictive skill is more sensitive to the choice of static attributes than to the choice of LSTM structure, among the models tested. This work is valuable in guiding where the large-sample hydrology community should focus its future efforts.
However, I have some concerns about the framing. First, the title, “Better data or better architecture?”, frames a much larger problem than the one the paper addresses. The investigation is limited to static data, which is fine, but the title should be adjusted to reflect this scope.
Second, the paper discusses model architecture in general, but only a limited number of architectures were investigated, and results are reported for only three. The conclusions are nevertheless generalised from these. Such a broad conclusion would require a more systematic selection of models. Otherwise, the title and conclusions should be narrowed to make these limitations clear.
Specific comments
- L57–68. The argument assumes the model represents a complete input–output relation. For an LSTM, this means the look-back window must contain all relevant history. This may not hold for catchments with slow stores (groundwater, multi-year storage), whose state is only partly captured by a 365-day window.
- L57–68. Even in the gauged case (L65–67), random identifiers work only when each catchment has a long record. When records are short, or for rare events such as extreme floods that are poorly sampled in any single catchment, meaningful attributes help the model learn from similar catchments; random identifiers cannot.
- L86–87. “Deeper architectures” could be read as more LSTM layers, whereas the variants tested add a layer before a single LSTM layer. Consider “more complex architectures” instead, to avoid confusion.
- L157–160, L180–182, L187–190. The preliminary experiments mentioned here (raw MODIS bands vs derived indices, 500 m vs 90 m resolution, and quantiles vs PCA or autoencoder compression) would be interesting to readers. The authors may wish to include them in an appendix or supplement.
- L251–252. Section 3.2 refers to Appendix C, which is not in the manuscript. Please add it or remove the reference.
- L253–256. Please state how the catchments were assigned to the five folds (random or spatially grouped).
- Section 4.1.1 and Table 2. It would be interesting to see CDF plots of the NSE across catchments, in addition to the medians.
- L368–369. The sentence says AlphaEarth “significantly outperforms” the CAMELS model, “however without significance” (p = 0.364). Please correct.
- L376. AlphaEarthplus vs MODISplus gives KS = 0.094, p = 0.018, which is significant at the 95% level. The text states the opposite. Please check the numbers or the statement, as it bears on the paper’s main conclusion.
- L448–452. The text says the “None” model has the lowest erank, but Figure 2 shows it has the highest. Please correct.
Citation: https://doi.org/10.5194/egusphere-2026-1965-RC3
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 989 | 510 | 67 | 1,566 | 53 | 57 |
- HTML: 989
- PDF: 510
- XML: 67
- Total: 1,566
- BibTeX: 53
- EndNote: 57
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1