DeepBase: A Deep Learning-based Daily Baseflow Dataset across the United States

Authors: Parnian Ghaneei, Hamid Moradkhani Year: 2025 Tags: baseflow-estimation, lstm, ungauged-basins, conus-hydrology, transfer-learning, data-descriptor

TL;DR

DeepBase generates daily baseflow estimates for 1,661 CONUS basins from 1981–2022 using a pipeline of UMAP dimensionality reduction, Growing Neural Gas clustering, and transfer-learning-finetuned LSTM models. It addresses the shortage of high-quality, spatially broad baseflow data by explicitly supporting ungauged basin prediction without using any streamflow or baseflow attributes as inputs.

First pass — the five C's

Category. Data descriptor / research prototype — primary output is a publicly released dataset; methodology is novel but the paper's purpose is dataset generation and validation.

Context. Hydrology / machine learning for hydrological estimation. Builds on: Kratzert et al. 2019 (large-sample LSTM hydrology); Xie et al. 2022 (LSTM-based monthly gridded baseflow for CONUS); Beck et al. 2013 (global BFI mapping via ANN); NLDAS-2 Noah/VIC as the comparison baseline.

Correctness. Load-bearing assumptions: (1) digital-filter-derived baseflow from USGS streamflow is a valid training target; (2) 12 separation methods evaluated against strict baseflow points provide a defensible "observed" baseflow; (3) UMAP-GNG clusters on static hydroclimatic attributes adequately characterize ungauged basin similarity; (4) excluding all flow-related attributes truly prevents information leakage in the ungauged scenario. Assumptions (1) and (2) are partially contested in the literature but not challenged within the paper.

Contributions. - Daily baseflow dataset for 1,661 CONUS basins (1981–2022), freely available on Figshare. - UMAP-GNG clustering pipeline applied to 21 static hydroclimatic/physical attributes to identify 14 homogeneous basin groups for guiding transfer learning. - Transfer learning framework (general LSTM → cluster-finetuned LSTM) enabling baseflow estimation in fully ungauged basins without any streamflow input. - Systematic comparison of LSTM-generated baseflow against NLDAS-2 Noah and VIC products across all 14 clusters.

Clarity. Writing is generally understandable but repetitive across the Background and Methods sections; the description of hyperparameter tuning relies heavily on "trial-and-error" without reporting the search space outcomes.

Second pass — content

Main thrust: A UMAP-GNG clustering + fine-tuned LSTM pipeline trained on 530 gauged basins produces daily baseflow estimates for 1,661 CONUS basins that outperform NLDAS-2 Noah and VIC on both NSE and LTRMSE, and the dataset is released for 1981–2022.

Supporting evidence: - Gauged scenario (test period 2007–2014): LSTM median NSE > 0.7 in 12 of 14 clusters; lowest cluster (Cluster 14, southern Texas) achieves median NSE = 0.56. - Ungauged scenario (8-fold cross-validation, 1981–2014): 12 of 14 clusters achieve median NSE > 0.6; Clusters 5 and 14 (high-aridity) are the poorest performers. - LSTM LTRMSE is lower than both Noah and VIC in all 14 clusters in both gauged and ungauged scenarios (exact values not tabulated in the text excerpt). - Clustering quality: Silhouette Coefficient = 0.56, Davies-Bouldin Index = 0.65 for the 14-cluster solution from 530 training basins. - Mann-Kendall trend analysis (1981–2022) shows consistent positive baseflow trends in eastern clusters (e.g., 12, 13) and negative trends in western clusters (e.g., 3, 5, 14), particularly in summer.

Figures & tables: - Fig. 2 (cluster spatial maps for 530 and 1,661 basins with FS/AI overlays): key for understanding cluster geography; axes/legends present but no uncertainty bounds — qualitative interpretation only. - Fig. 3 (spatial NSE and LTRMSE maps for LSTM, Noah, VIC in gauged scenario): carries the main validation argument; color maps used but no confidence intervals or statistical significance reported. - Fig. 4 (NSE and LTRMSE spatial maps for ungauged scenario): same limitations as Fig. 3. - Fig. 5 (multi-year monthly average baseflow maps, 1981–2022): illustrates spatiotemporal patterns; no uncertainty shown. - Fig. 6 (Mann-Kendall Z-value maps by season): significance levels stated (p < 0.1, 0.05, 0.01) — the only figure with explicit statistical thresholds. - Table 1 (static attributes with units and sources): well-structured. - Table 2 (12 baseflow separation methods): complete reference list. - No box plots, violin plots, or quantitative summary tables for NSE/LTRMSE distributions by cluster are present in the excerpt.

Follow-up references: - Kratzert et al. 2019 (HESS) — foundational multi-basin LSTM hydrology paper this framework directly extends. - Xie et al. 2022 (WRR) — closest prior work (monthly gridded LSTM baseflow for CONUS); direct methodological predecessor. - Beck et al. 2013 (WRR) — global BFI mapping via ANN; provides the benchmark framing for spatial baseflow products. - Xie et al. 2020 (J. Hydrology) — evaluation of baseflow separation methods for CONUS; provides the separation-method toolkit used here as training targets.

Third pass — critique

Implicit assumptions: - Digital-filter and graphical baseflow separation applied to USGS streamflow is treated as ground truth for training; these methods carry their own structural uncertainties and parameter sensitivities, which are not propagated into model uncertainty estimates. If separation methods systematically over- or under-estimate baseflow in specific climate regimes, the LSTM inherits that bias. - Strict baseflow points used to select the "best" separation method per basin assume the four Brutsaert rules correctly isolate true baseflow; violation of these rules (e.g., in flashy or regulated streams) would corrupt the training target. - Basin area < 2,000 km² cutoff is applied without justification of whether this threshold meaningfully separates basins where the LSTM physics approximation holds. - UMAP neighbor count (10) and GNG hyperparameters are chosen by trial-and-error without a held-out cluster quality test; the 14-cluster solution is not compared against alternative cluster counts quantitatively. - Transfer of cluster labels from 530 training basins to 1,131 ungauged basins (by forward-running UMAP-GNG) assumes the dimensionality-reduction manifold generalizes — not validated independently.

Missing context or citations: - No comparison against process-based models beyond NLDAS-2 (e.g., MOSAIC, SAC-SMA also in NLDAS-2, or mHM, HBV) despite these being available benchmarks. - No comparison against the Xie et al. 2022 LSTM monthly product at matched temporal aggregation, which is the most directly comparable DL baseline. - Uncertainty quantification of the generated baseflow is absent; no ensemble, Monte Carlo, or Bayesian interval is reported, which limits use for drought/flood risk applications. - Basins excluded due to gaps > 3 years (>2/3 of the 1,661 basins) receive no cross-check of generated baseflow quality; the paper does not discuss how data gaps in the training pool affect spatial extrapolation. - The paper does not engage with regulated vs. unregulated streamflow: GAGES-II "least disturbed" criterion is mentioned but not verified to eliminate all anthropogenic baseflow modification.

Possible experimental / analytical issues: - Training target is itself a model output (digital filter applied to streamflow), not a direct measurement; the paper acknowledges field-based validation is lacking but does not quantify the resulting uncertainty floor. - The gauged scenario uses a single temporal split (train 1981–2001, validation 2001–2007, test 2007–2014); this does not account for inter-decadal non-stationarity and is not replicated across different period splits. - Ungauged validation uses 8-fold cross-validation but the fold composition (random or spatially stratified) is not described; spatially clustered folds could inflate apparent generalization. - Exact NSE and LTRMSE values for each cluster are not reported in tables; only qualitative thresholds ("above 0.7," "above 0.6") are stated, preventing full reproduction or meta-analysis. - The LSTM is trained using NSE* (basin-averaged NSE) as the loss function, but the chosen hidden state, dropout, and learning rate hyperparameter combination actually selected is not reported. - No ablation is provided comparing (a) general LSTM only vs. (b) general + finetuned LSTM, so the contribution of the finetuning step to performance gains is unquantified. - Comparison with NLDAS-2 is unfair in one direction: Noah and VIC are gridded products evaluated at basin-average via zonal statistics, whereas the LSTM is trained basin-by-basin — spatial resolution mismatch is acknowledged only briefly.

Ideas for future work: - Introduce conformal prediction or Bayesian LSTM to attach calibrated uncertainty intervals to each daily baseflow estimate, enabling probabilistic drought/flood risk assessments. - Conduct an ablation study isolating the contribution of UMAP-GNG clustering and finetuning vs. a single general LSTM applied to all 1,661 basins, with reported confidence intervals on NSE differences. - Extend the framework to regulated basins by incorporating reservoir operation data, allowing coverage beyond GAGES-II "least disturbed" basins. - Validate generated baseflow against available field-based measurements (e.g., tracer-based baseflow estimates, groundwater-level recession data) in a subset of basins to constrain the error floor introduced by using digital-filter baseflow as training labels.