Computational hydrology

Over the last few years, significant progress has been made in the development of convection-permitting climate models (CPCMs), especially in improving precipitation estimates in regions with complex terrain. Recently, the South American Affinity Group (SAAG) developed a novel high-resolution dataset – hereafter referred to as the WRF-SAAG dataset – by dynamically downscaling the ERA5 reanalysis using the Weather Research and Forecasting (WRF) model over South America for the period 2000–2021. Here, we evaluate WRF-SAAG outputs across continental Chile and compare them with two widely used gridded meteorological products, CR2MET and RF-MEP, which combine reanalysis data with in situ measurements. In particular, we analyze spatial differences among climatological estimates from the three products, and we assess the quality of daily precipitation and daily maximum and minimum temperature simulations using observations from meteorological stations over the period April/2001–March/2018. We also compare the potential of these climate datasets for hydrological modeling using a conceptual, bucket-style rainfall-runoff model, evaluating the results against streamflow observations, satellite remote sensing (SRS) products, and a snow water equivalent (SWE) reanalysis product. We found that the detection accuracy varied within each Chilean macrozone, with worse performance in the Far North (17.5–26° S) and Patagonia (43.7–56° S). The evaluation of daily precipitation against station observations using Tang's Kling–Gupta efficiency (KGE T ) reveals that all datasets performed better in reproducing precipitation in rainy regions (median KGE T ≥ 0.65 in the Southern macrozone), while in arid areas such as the Near North during summer, the median KGE T was negative. The CR2MET product consistently provided the best performance for extreme precipitation and temperature, partly because it includes information from the stations used for evaluation. Finally, the application of the hydrological model shows that although WRF-SAAG simulations achieved runoff estimations comparable to the best observation-based gridded products (i.e., CR2MET or RF-MEP), with the best results in the southern macrozone they provided relatively lower performance in evapotranspiration (ET) and soil moisture estimates, as well as SWE simulations of competitive quality. More broadly, the results…

Process-based land surface models, such as the Noah-Multiparameterization (Noah-MP) model, are widely used for large-scale hydrologic simulations because of their flexibility in selecting multiple parameterization schemes. However, limited guidance on choosing appropriate configurations constrains their reliability in representing runoff and snowmelt dynamics across diverse land-cover and snow conditions. This study evaluates the default and four alternative parameterization schemes in the Noah-MP land surface model, including Runoff and Groundwater (RUN), Surface Exchange Coefficient for Heat (SFC), Frozen Soil Permeability (INF), and Snow/Soil Temperature Time Scheme (STC), across 50 Hydro-Climate Data Networks (HCDNs) in the Missouri River Basin. Model performance was evaluated using USGS streamflow observations and snow water equivalent (SWE) estimates from the University of Arizona dataset for 2014–2023. Results showed that the alternative schemes improved runoff simulation compared to the default scheme through different approaches representing key hydrological and thermodynamic processes. Specifically, the RUN, SFC, INF, and STC experiments improved the Kling–Gupta Efficiency (KGE) by 0.19, 0.38, 0.44, and 0.14, respectively in representative subbasins, through enhanced groundwater dynamics, reduced evapotranspiration bias, improved rapid runoff response, and more accurate SWE evolution. SWE evaluation further indicates that the STC experiment most effectively reduced SWE bias in snow-dominated subbasins, though improvements in SWE did not universally translate into better runoff simulation. These results highlight the importance of basin-specific parameterization schemes within Noah-MP to improve hydrological prediction and water management across diverse hydroclimatic regions. The findings further indicate optimal parameterization schemes for different climates, land cover, and snow regimes.

Deterministic model predictions can struggle to adequately capture extreme events such as floods and droughts, which are of particular importance in hydrology. This limitation arises because deterministic models collapse the conditional discharge distribution to a single point estimate. Probabilistic modeling provides a way to address this issue by explicitly representing uncertainty and assigning non-zero probabilities to all possible outcomes, including extreme events, thereby capturing the entire conditional discharge distribution. Motivated by this perspective, we examine whether probabilistic Long Short-Term Memory (LSTM) models improve the representation of extreme events in rainfall–runoff simulations across Switzerland. The differing degree of flexibility in the models leads to different degrees and distinct patterns of miscalibration in the predictive distributions. For the extreme discharges (top 0.05 % of the marginal discharge distribution), the deterministic LSTM systematically underestimates 97 % of these events, whereas 83 % of these fall within the probabilistic model's 99th percentile prediction interval. Given the rarity of these events, representing them as outcomes near the upper tail of the predictive distribution is both expected and statistically consistent. Building on the additional information provided by probabilistic runoff predictions, we further show how they can be translated into actionable flood warnings using reinforcement learning. To this end, we introduce a Flood Risk Communication Agent (FRiCA) that operates on probabilistic runoff predictions and learns risk-aware decision rules for issuing warnings of varying intensity. Results indicate that the FRiCA outperforms simple fixed heuristics, such as issuing warnings based on the deterministic LSTM predictions or a fixed high quantile (e.g., the 99th percentile). The learnt behavior demonstrates the potential of reinforcement learning as a decision layer for improved flood risk communication, and motivates further exploration of better reward design and policy network optimization for context-dependent decision policies that adapt to varying hydrological and societal contexts.

We combine CONUS404 convection‐permitting climate simulations with the physics‐based WRF‐Hydro model to assess how climate warming may influence flood frequency and seasonality across the Midwestern United States. This framework captures interactions among precipitation, soil moisture, and snowpack that shape flood generation. Both simulated warming and historical records show a regional shift toward more extreme warm‐season, rainfall‐driven floods. This regional signal, however, does not translate uniformly to individual watersheds, where flood responses reflect not only warming but also the influence of subregional climate variability and sampling uncertainty. Local flood‐frequency changes thus remain highly uncertain despite the thermodynamic tendency toward stronger convective precipitation. Ensembles of convection‐permitting simulations are needed to constrain plausible local outcomes. These findings suggest that flood risk communication under climate warming should distinguish broad climate‐driven tendencies from uncertain local responses.

Physics‐informed neural networks (PINNs) provide a promising framework for river hydrodynamic simulation and parameter inversion under sparse observations. However, spatial heterogeneity in channel geometry and roughness causes strong variations in flood‐wave propagation and parameter sensitivity, challenging local conservation and joint state‐parameter estimation. This study proposes a Characteristic‐aligned Control‐volume Physics‐Informed Neural Network for Roughness Estimation (CC‐PINN‐RE). It constructs dynamic diamond‐shaped space‐time control volumes from local characteristic speeds of the Saint‐Venant equations, enforces mass and momentum conservation in integral form, and combines adaptive control‐volume management with local‐sensitivity‐weighted roughness regularization. The method is evaluated on three one‐dimensional channel cases of increasing complexity against strong‐form PINN‐RE and fixed‐control‐volume cvPINN‐RE. CC‐PINN‐RE achieves the lowest errors in eight of nine relative metrics for water depth, discharge, and roughness. In Case 3, where channel width, bed slope, and roughness vary spatially, it reduces the corresponding errors by 57.0%, 82.2%, and 57.3% relative to PINN‐RE and reduces the maximum absolute error in peak water depth by 69.7%. Ablation experiments with three random seeds support the contributions of the three proposed components. Observation‐deployment experiments show that increasing the number of monitoring stations is generally more effective than increasing temporal sampling frequency at individual stations. Under 0%–2% observational noise in Case 3, the model remains stable. These results indicate that CC‐PINN‐RE provides a physically consistent framework for integrating characteristic propagation, local conservation, and parameter inversion, offering a useful basis for joint hydrodynamic reconstruction and spatially varying roughness estimation in heterogeneous river channels.

The evaluation of hydrological models is essential for both model selection and reliability assessment. However, comparing predictions with observations alone is insufficient to fully characterize model behavior. This limitation is particularly relevant to deep learning (DL) models because of their complex internal structures. In risk‐averse operational settings, water managers require models whose reliability and failure modes have been carefully evaluated because unstable predictions can put critical infrastructure at risk. Motivated by the need to assess model reliability before operational deployment, we introduce and explore adversarial robustness analysis for hydrological modeling and evaluate whether small, targeted perturbations to meteorological forcings induce substantial changes in simulated discharge. We compare a physical‐conceptual model (HBV) with a deep learning‐based model (LSTM) across 1,347 German catchments using perturbations of varying magnitudes generated by the fast gradient sign method (FGSM). We find that FGSM perturbations generally reduce Kling–Gupta efficiency and increase mean squared error. However, widespread catastrophic failures under the specific FGSM perturbations are not observed, and LSTM generally exhibits less performance degradation than HBV. In the examined catchments, changes in the predicted hydrographs and time‐averaged internal model states often vary approximately linearly with perturbation magnitude, providing a compact summary of how errors grow under such perturbations. Similar patterns are observed along random perturbation directions in selected catchments, suggesting that, in those catchments, small input changes produce approximately proportional changes in model output. Overall, these findings characterize local predictive stability under controlled stress tests and motivate future studies of model robustness in real‐world operational settings.

Urban flood monitoring demands timely, reliable decisions under urgent and often unseen conditions. However, faults in sensor networks can undermine response, leading to human and economic losses. To address these issues, we propose a Nowcasting Physics‐Informed (NPI) model for fault detection in streamflow sensors, with a 100‐min memory window at 2‐min resolution. A Long Short‐Term Memory (LSTM) consumes fused features from measured signals and Storm Water Management Model simulator outputs to produce end‐of‐window fault probabilities. To ensure operational reliability, a lightweight per‐sensor eXtreme Gradient Boosting (XGBoost) stage calibrates the LSTM score using window‐local statistics of the measurement. In this study, we investigate whether and to what extent fusing a physical model with sensor data improves fault detection, and whether per‐sensor threshold calibration matters. Our results indicated yes. The Area under the precision–recall curve (AUC‐PR) was improved by 3.7%–12.5% over the physics‐informed model compared to baseline non‐physics‐informed learning. In addition, post‐hoc calibration improved AUC‐PR by an additional 0.2%–2.3%. Across five cross‐validation folds, F1‐score (F1) increased from 0.75 for the non‐physics LSTM baseline to 0.83 with calibration, and further to 0.88 for the physics informed model. Adding post‐hoc calibration to the physics informed model yielded the best result at 0.90. The NPI is causal and auditable, exporting calibrated probabilities and binary alerts for direct operator use without retraining the LSTM, and offers a practical template for warning systems and training strategies that remain robust to unseen conditions.

Accurate estimation of regional-scale terrestrial carbon budgets is of great importance but remains challenging. With particular advantages, the Long Short-Term Memory (LSTM) network method shows potential for improving regional carbon budget upscaling estimations. Here, using LSTM, we upscaled regional net ecosystem carbon exchange (NEE) with available flux tower measurements and satellite land surface observations in North America. With well-established ecosystem-specific LSTMs, we produced monthly NEE at a spatial resolution of 0.1 ° ×0.1 ° over 2001–2021 (labelled as MemoryFlux). Unlike existing upscaling estimates, our dataset properly identified the Midwest Corn Belt as a region of large seasonal carbon uptake during the peak growing season, a feature revealed by previous top-down studies and recognized as a model benchmark. Moreover, the seasonal variations in NEE estimated by MemoryFlux were strongly correlated with those from independent atmospheric inversions, including the ensemble mean of the Orbiting Carbon Observatory-2 Model Intercomparison Project (OCO-2 v10 MIP; r =0.96, p <0.001) and CarbonTracker2022 (CT2022) ( r =0.97, p <0.001). The mean annual NEE was estimated at −1.27 ± 0.12 Pg C yr −1 , which was closer in magnitude to inversions (−0.83 to −0.70 Pg C yr −1 ) than existing upscaling estimates (−3.30 to −1.68 Pg C yr −1 ). In addition, MemoryFlux captured spatial NEE anomaly patterns associated with six selected severe drought and flood events. We further found that explicitly incorporating historical predictor information improved the representation of NEE interannual variability and spatial anomalies associated with climate extremes. MemoryFlux provides an improved bottom-up estimation of North American NEE and shows greater consistency in magnitude with independent atmospheric inversion estimates than several existing EC-based upscaling products. The MemoryFlux dataset is available at https://doi.org/10.5281/zenodo.20482274 (Huang and He, 2026).

Ecohydrological models are increasingly applied in regional scenario‐based studies, yet their cross‐resolution application remains constrained by high computational costs of fine‐resolution simulations and by structural inconsistencies introduced during scale transfer. We propose a Structure–Function Coherent Coarsening (SFCC) framework that improves the preservation of hydrological structure and functional heterogeneity during model input coarsening. Using the VELMA model in 24 subbasins of the Salish Sea Basin, USA, we evaluate coarsening strategies for three input types: (a) DEMs coarsened with a Hydro‐Aware approach that better preserves drainage topology; (b) land cover and soil type data sets coarsened with function‐preserving methods ( Auto‐Weight and Auto‐Reassign ) that retain small but process‐dominant classes; and (c) initial conditions coarsened with Hydro‐Aware , Landcover‐Aware , and Soil‐Aware strategies to improve temporal stability. Results show that the Hydro‐Aware method better preserves watershed morphology and yields more consistent runoff and nitrate loss predictions across resolutions than mean‐based coarsening. For categorical inputs, the function‐preserving methods can reduce the dominant‐class bias of majority aggregation, particularly in basins where small, high‐impact patches drive nitrogen export. Long‐term simulations show that hydrological variables stabilize quickly, whereas biogeochemical variables stabilize more slowly. In both cases, differences caused by coarsening decrease over time and eventually level off. This suggests that preserving both structural consistency and functional heterogeneity can help maintain stable model behavior over time. By operating directly at the data‐input level, SFCC improves integration of multi‐source data sets while retaining high‐resolution information. It supports cross‐scale simulation, calibration, and model integration, and provides a scalable foundation for regional scenario analysis.

Recent advances have introduced diffusion models for probabilistic streamflow forecasting, demonstrating strong early flood‐warning skill. However, current implementations rely on Long Short‐Term Memory (LSTM) backbones and discrete‐time diffusion formulations. This leaves unresolved whether more advanced sequence‐modeling architectures and continuous‐time diffusion formulations can improve probabilistic streamflow forecasting. In this study, we developed HydroDiffusion, a diffusion‐based probabilistic forecasting framework integrating a decoder‐only state space model (SSM) backbone with a continuous‐time score‐based diffusion formulation. HydroDiffusion is evaluated across 531 watersheds in the contiguous United States (CONUS) in the CAMELS data set. We benchmark HydroDiffusion against two diffusion baselines with LSTM backbones, as well as the recently proposed Diffusion‐based Runoff Model (DRUM). Results show that HydroDiffusion achieves the strongest nowcast accuracy when driven by observed meteorological forcings. Moreover, HydroDiffusion delivers higher deterministic and probabilistic forecast skill than its LSTM‐based counterpart in operational forecasting. Attribution analysis further reveals that the SSM backbone excels in humid, densely vegetated, storage‐dominated watersheds with strong hydrologic memory, while its advantage becomes limited in arid regions and hydrologically complex systems such as karst and prairie pothole basins. These results establish HydroDiffusion as an effective generative modeling framework for medium‐range streamflow forecasting, providing both a new modeling benchmark and a foundation for future research on probabilistic hydrologic prediction at continental scales.

Estimation of surface soil moisture (SSM) is fundamental for hydrological monitoring and agricultural management, particularly in drought-prone regions such as central–northeastern Argentina. Data assimilation provides a robust approach for estimating SSM by integrating land surface models with remote sensing data. Satellite missions such as SMAP and SMOS deliver global SSM products; however, these observations often exhibit systematic biases arising from instrument noise, indirect measurements, and temporal and spatial heterogeneities. Since standard data assimilation techniques assume unbiased observations, bias correction is required to ensure optimal assimilation. This work evaluates off-line observation bias-correction methods within a land data assimilation framework based on the Noah-MP land surface model and an ensemble Kalman filter, focusing on the 2022 dry season in the endorheic Pampas region. During dry conditions, SSM probability density functions differ significantly among soil textures within the bin. To address sampling errors in cumulative distribution function (CDF) matching, we propose an approach in which climatological statistics are computed using homogeneous soil texture pixels, and a 45-day moving temporal window is applied to smooth CDF evolution. The proposed method is compared with standard CDF matching and normal deviate scaling. Results indicate that improved statistical sampling in CDF matching produces a measurable impact on SSM estimates, particularly during dry periods. It shows better agreement with independent in-situ soil measurements than standard CDF matching, provided that the soil texture classifications at the in-situ site and in the model grid cell are consistent.

Great Britain's variable maritime climate has until relatively recently limited the utility of seasonal hydrological forecasts. The latest generations of seasonal atmospheric forecasting systems have created new opportunities to improve flow forecasting across Great Britain, such as for the UK Hydrological Outlook. Here, we combine newly-developed high-resolution rainfall forecasts derived from historical weather analogues (HWA) conditioned on large-scale circulation patterns with a monthly-resolution national-scale hydrological model to produce seasonal and monthly forecasts. This forecasting scheme is evaluated using hindcasts from 1993–2017. We show that the skill of the rainfall forecasts is comparable to that achieved at coarser resolutions by other models. However, increasing the spatial resolution has degraded some of the skill found by previous authors producing HWA forecasts at lower resolutions. The river flow forecasts perform better than rainfall, with much of the improvement attributable to the skill provided by hydrological initial conditions. Both rainfall and river flow forecasts are most skilful in winter and perform adequately in spring and autumn. We show that in responsive catchments, primarily found in the northwest, rainfall forecasts contribute most of the skill in the river flow predictions but are much less important in regions with long hydrological memory. We thus indicate where flow forecasts would benefit most from improved rainfall forecasts or from better hydrological modelling. The introduction of these high-resolution rainfall forecasts now enables hydrological forecasting at unprecedented levels of detail across Great Britain and demonstrates a forecasting approach that may be similarly beneficial elsewhere in the world.

Large-scale hydrological models like CTRIP and MGB are essential for simulating river dynamics and supporting large-scale climate studies. Their accuracy can be significantly improved through satellite data assimilation. This study leverages the stand-alone value of 20 years of ESA Climate Change Initiative (CCI) high-resolution discharge and water surface elevation (WSE) products (2000–2020) for improving large-scale hydrological simulations through data assimilation. To evaluate the added-value of these produtcs across contrasting modelling and hydrological contexts, we assimilate altimetry-derived discharge, multispectral-imagery-derived discharge, and WSE anomalies into two existing ensemble Kalman Filter frameworks: HyDAS in CTRIP, a global-scale physically based and uncalibrated river-routing model, and HYFAA in MGB, a calibrated semi-distributed regional hydrological model. The experiments are conducted over the Niger and Congo basins, which differ in hydrological variability, river-network structure, wetland influence, and CCI product availability. Across the experiments, discharge assimilation generally outperformed WSE anomaly assimilation because discharge is directly represented in the routing models, providing a more direct and physically consistent correction, while WSE requires consistency between observed and simulated rating curves to be converted into effective discharge corrections. In the Niger basin, where the seasonal signal and station coverage better constrain the main river dynamics, assimilating altimetry-derived discharge led to the strongest improvement in MGB, increasing the median Nash-Sutcliffe Efficiency (NSE) to 0.83 and the correlation coefficient to 0.94. WSE anomaly assimilation was beneficial in specific cases, particularly when the baseline simulation was poor and when observed and simulated rating curves were well aligned. Temporal data density in discharge assimilation emerged as a key driver of performance gains. Assimilating high-frequency discharge data from multispectral imagery significantly reduced bias, from 1.2 to near 1 in MGB, and from 2.3 to 1.78 in CTRIP (median values), supporting hydrological assessments related to long-term variability. Furthermore, the higher temporal resolution allowed for better capture of flow variability, with Kling-Gupta Efficiency γ approaching 1.0 in MGB, which is relevant for…

Accurate streamflow prediction is critical for effective water resource management, and recent advances in deep learning and multi‐basin data sets have significantly improved model skill. Global deep‐learning models, trained on thousands of catchments have demonstrated strong generalization across diverse climates and continents. However, these models rely solely on globally available meteorological inputs, potentially overlooking the value of nationally developed data sets with enhanced regional accuracy. In this study, we assess the benefit of using Canada's high‐resolution national reanalysis data set (CaSR) in long short‐term memory (LSTM) models for streamflow prediction. We compare model performance using different meteorological forcings, global (ERA5‐Land), continental (Daymet), national (CaSR), and a combined “Complete” model, across over 1,000 Canadian stream gauges. Model performance is evaluated using the Nash–Sutcliffe Efficiency (NSE), Kling‐Gupta Efficiency (KGE) and related metrics. We also benchmark our best‐performing model against Google's global LSTM streamflow model at 412 Canadian gauges. Results show that national‐scale deep‐learning models using CaSR outperforms deep‐learning models using only global or continental inputs. The Complete model achieved the highest performance, with a median KGE 0.886−exceeding the 0.839 achieved by Google's global model within Canada. These findings suggest that nationally trained models using a wide array of inputs, including national meteorological data sets, can outperform state‐of‐the‐art global deep‐learning models that do not use such data when applied within their region of origin. For countries with dense observational networks and distinct hydrological regimes, investing in national data sets and models can enhance deep‐learning model prediction accuracy, leading to improved resilience to climate extremes and local water management decisions.

Streamflow reanalysis is vital for water resources management and climate impact assessment; however, the extent to which it is affected by precipitation forcing errors remains poorly understood. Focusing on the Global Flood Awareness System driven by the European Centre for Medium-Range Weather Forecasts Reanalysis v5 (GloFAS-ERA5), this paper details how streamflow reanalysis errors respond to precipitation errors. Specifically, the root mean square errors (RMSEs) are calculated by hydrological year for reanalysis products across 671 catchments in the Catchment Attributes and Meteorology for Large-sample Studies (CAMELS) dataset; and by combining catchment-specific linear regression with global panel regression, the effects of precipitation errors on streamflow errors are quantified. The results demonstrate an improved performance from GloFAS-ERA5 v2.1 to v4.0, with the median RMSE decreasing from 2.16 to 1.81 mm. For GloFAS-ERA5 v4.0, the panel regression estimates an average increase of 0.51 mm in streamflow RMSE for each 1 mm increase in precipitation RMSE across the 671 catchments. In the meantime, the corresponding catchment-specific increase of streamflow RMSE reaches up to 2.5 mm in humid catchments but remains below 0.7 mm in arid catchments. These divergent responses suggest that streamflow errors may be more directly associated with precipitation errors under the saturation-excess conditions while soil moisture deficits can dampen their effects. Furthermore, incorporating interaction terms into panel regression increases the coefficient of determination ( R 2 ) from 0.16 to 0.36, indicating these responses are modulated by catchment heterogeneity. This modulation is further presented by targeted case studies, indicating that temperature controls the storage and release of snow water, thereby dampening and delaying the responses of streamflow errors to precipitation errors in snow-dominated catchments. These findings provide a valuable diagnostic method and practical guidance for applications of global streamflow reanalysis to complex, heterogeneous catchments.

Seasonal snow is a critical component of the water cycle in the Southern Andes of Chile and Argentina. Quantitative assessments of snow accumulation remain constrained by the scarcity, heterogeneity, and inconsistency of in situ observations, resulting in substantial uncertainties in mountain hydrological modeling. To address this gap, we compiled and quality-controlled snow depth observations to produce a consistent daily dataset from 81 stations between 21 and 54° S covering 2010–2024. Our quality-control procedure involved data compilation and preprocessing, followed by harmonization of snow depth observations, primarily by correcting the ground reference level relative to the soil surface during snow-free periods and removing anomalous spikes and observations outside physically plausible ranges. This process improved data reliability, increasing the Physical Consistency Index (PCI) from 87 % to 95 % at selected stations in the Maipo River Basin, while reducing the median number of days with data across stations by 23 % (from 1392 to 1074 d). The snow depth data availability increased from one station in 2010 to 57 stations in 2024, largely driven by the expansion of the monitoring network operated by the General Directorate of Water (DGA), Chile. However, this expansion remains spatially uneven across the Andean zones. The Mediterranean Andes have the largest number of stations (39), as well as the largest number of stations with highly complete records with 17 stations reaching 80 %–100 % data coverage, compared with only nine stations each in the Arid and Wet Andes, highlighting persistent spatial and temporal gaps. Using this newly quality-controlled dataset, we find that snow depth generally increases from the Arid to the Wet Andes in association with increasing precipitation, whereas its relationship with elevation is not consistently positive. The snow depth-elevation relationship is nonlinear in the Arid and Mediterranean Andes, with maximum observed snow depths at 4300 m a.s.l. in the Elqui River Basin and 3300 m a.s.l. in the Maipo River Basin. In contrast, the available observations indicate a positive snow depth-elevation relationship in the Wet Andes (Maule–Itata River Basin). However, these relationships should be interpreted in the context of the specific spatial and temporal conditions represented by the available observations. This…

Peatlands play a crucial role in regional water balance and carbon dynamics but are often degraded due to drainage and agricultural use. In Germany, many drained peatlands have shifted from carbon sinks to CO 2 sources. Rewetting these ecosystems is therefore essential to restore their ecological functions and mitigate greenhouse gas emissions. However, effective rewetting requires a detailed understanding of peatland hydrology and its response to climatic and management conditions. To address this need, this study employs a fully coupled surface–subsurface hydrological model (HydroGeoSphere) to analyze the complex hydrological functioning of a typical degraded fen peatland site (11.6 ha) in Brandenburg, Germany. The model-based quantification of hydrological fluxes is basis for assessing peatland vulnerability to climate variability and land use while providing a hydrological baseline for future evaluation of rewetting strategies. The studied peatland is connected to a regional aquifer and intensively drained by a system of ditches. Simulations used daily meteorological inputs and detailed field measurements from 2015 to 2023. Evapotranspiration (ET) was parameterized using field-measured vegetation dynamics (seasonal leaf area index and management schedules), while measured ditch water levels served as hydraulic boundary conditions. The site was spatially divided into different management units with distinct vegetation parameters. The peat profile was represented by two layers (a 0.3 m highly degraded surface peat overlying a 0.7 m less degraded layer) overlying sand (aquifer) and till (aquifer base). The model was evaluated from different angles against eddy covariance ET and groundwater table dynamics during a calibration period (2016–2020) and a validation period (2021–2023) using a multi-metric approach. The model successfully reproduced seasonal water-table fluctuations and ditch–peatland interactions, including ET-driven hydraulic gradient dynamics between summer and winter. Simulated ET closely matched eddy covariance measurements, with RMSE values of 64 mm yr −1 , 10.2 mm month −1 , and 1.01 mm d −1 , and showed only minor biases during dry conditions, while over the year seasonal dynamics of ET were also well captured by the model. The model reproduced groundwater variations with sufficient accuracy, achieving KGE values of 0.80–0.85, NSE of…

Watershed streamflow is often the focus of hydrological model calibration and evaluation, despite other potential objectives, including water quality management, flood protection, or agricultural management. When hydrological models are calibrated on streamflow, intermediate processes such as those affecting soil-moisture are not necessarily well represented. This research evaluated whether calibration using downscaled and bias-corrected satellite soil-moisture improves prediction of field-scale soil-moisture relative to conventional streamflow-based calibration. In this work, downscaled satellite soil-moisture and streamflow data are used to calibrate a soil and water assessment tool – variable source area model initialized using a terrain informed process to create hydrologic response units. In-situ soil-moisture measurements at 25 locations across a 4.2-ha mixed-grass pasture located in southwestern Virginia were used to estimate field-scale average soil-moisture variability for model evaluation. Leveraging downscaled satellite soil-moisture data substantially improved estimation of temporal soil-moisture variability without affecting the model streamflow performance. The multi-objective calibration using streamflow and satellite soil-moisture improved soil-moisture performance while maintaining streamflow performance comparable to streamflow-only calibration. These results demonstrate the potential for satellite soil-moisture–informed calibration to improve internal hydrologic state estimation in small, saturation excess watersheds. Furthermore, these results highlight the importance of coupling statistical performance gains with evaluation of hydrologic realism when extending such approaches to broader modeling applications.

Achieving high prediction accuracy while maintaining low‐complexity and interpretability has been a critical challenge in streamflow forecasting. To address this issue, this study proposes an improved reservoir computing (RC) framework that simultaneously supports structural inference and dynamic forecasting of streamflow‐meteorology networks. The framework operates in two stages: first, a cross‐validation‐based interpretable dynamic weight inference algorithm is designed to uncover the relational structures within the streamflow–meteorology network; subsequently, the inferred structural information is integrated into an architecture embedded with catchment similarity to predict future streamflow states. This process guides the internal dynamics of the framework toward representations that align with the data‐driven dependency relationships between streamflow and meteorological variables, thereby enhancing the interpretability of the predictions. We partition a subset of catchments from CAMELS data set into eight groups and conduct a series of experiments to evaluate the proposed framework. Results demonstrate that the proposed method achieves an average improvement of 38.87% in Nash–Sutcliffe Efficiency (NSE) and 38.21% in peak flow prediction accuracy compared to benchmark models and effectively mitigates the temporal lag and instability commonly observed in deep learning approaches, exhibiting superior adaptability and robustness in handling spatial heterogeneity under diverse hydrological conditions. With fewer than 10% of the trainable parameters of benchmark models and a runtime 2 to 3 times faster, the framework enables rapid inference, adaptation, and modeling in cross‐catchment scenarios within complex hydrological dynamical systems, thus holding considerable promise for practical applications.

Exploring the future evolution of runoff within reservoir drainage areas and its dominant drivers provides insights for future reservoir management. We combine a georeferenced reservoir data set with historical runoff simulations and future projections from an ensemble of Global Hydrological Models (GHMs) to investigate the long‐term runoff evolution across 913 drainage areas of large reservoirs in China. To reduce uncertainties in the GHM ensemble, we adopt an ensemble best approach, which outperforms the ensemble mean and median. Using this approach, we incorporate observational constraints to project three aspects of runoff‐related variations: trends in inter‐annual magnitude, inter‐annual variability, and intra‐annual variability. A total of 337–438 reservoirs are projected to exhibit decreasing inter‐annual magnitude, weakening inter‐annual variability, and enhancing intra‐annual variability during the historical period, while increasing magnitude and enhancing variability under future scenarios. These three aspects of runoff‐related variations do not occur synchronously within the same reservoir groups, highlighting the need to consider each reservoir individually when developing effective management strategies. Attribution analyses indicate that anthropogenic climate change predominates over natural climate variability and human water and land management, serving as the dominant driver of runoff‐related variations. Although the ensemble best approach performs well at large spatial scales, uncertainties remain at the individual reservoir level, as relative errors exceed ±50% for 90–259 reservoirs in drylands and cold regions. To promote the reliability of future projections, we advocate continued efforts to improve GHM mathematical structures and parameterization schemes, alongside the integration of additional in situ observations for model calibration.

Accurate estimation of snowpack dynamics is essential for predicting snowmelt‐driven hydrological processes in mountainous regions. Although data assimilation techniques can integrate remote sensing observations into snow models to reduce uncertainties, their operational implementation remains constrained by sparse and irregular coverage of satellite data, complexity of mountainous terrain, and the computational demands of high‐dimensional assimilation. In this study, an Observation‐Masked Localized Ensemble Kalman Filter (OML‐EnKF) framework was developed to restrict analysis updates to observed regions while preserving localized spatial covariance among nearby observations. This framework assimilated daily Copernicus snow cover fraction (SCF) observations into the Wflow_sbm distributed hydrological model across the Rhône River basin, France, for 2017–2019. We compared its performance against an open‐loop run, a direct insertion scheme, and a one‐dimensional EnKF. Results demonstrate that incorporating a snow module in the Wflow_sbm model improved seasonal discharge patterns. The OML‐EnKF produced spatially coherent snow water equivalent patterns consistent with topography and improved snow depth estimates in mid‐elevation regions. Both EnKF schemes refined SCF and improved discharge predictions across multiple gauged stations. Furthermore, temporal shifts in snowpack accumulation and ablation modulated runoff timing, with increased snowmelt contributions during winter (October–March) and diminished discharge in late spring (April–June). Sensitivity analyses showed that a moderate localization radius (0.025–0.1, approximately 2.8–11.1 km) effectively captured spatial correlations within the OML‐EnKF, whereas excessively restrictive radius values degraded performance. These findings show that the OML‐EnKF enhances snowpack characterization and discharge prediction in the Rhône River basin, offering a promising tool for water resource management in snow‐dominated basins.

When transported in rivers by short‐term episodic events such as floods and debris flows, large wood interacts with the flow, sediment, and river boundaries, leading to physical degradation such as weathering, abrasion, and breakage. This degradation, manifested as mass loss and structural alteration, weakens wood stability, increases mobility, and diminishes its geomorphic and ecological functions. However, this process remains largely overlooked when assessing wood transport and accumulation. Here, we introduce laboratory tumbling experiments to investigate the physical degradation of fresh large wood, focusing on mechanical bark removal during fluvial transport. Fifteen tests were conducted under varying flow velocities and water volumes, with three White Willow samples in each test. Wood debarking causes changes in the volume of wood samples during the experiment. Therefore, Structure‐from‐Motion photogrammetry was employed to quantify changes in wood volume and thereby evaluate the debarking process. Results revealed that debarking involves two coupled processes: (a) expansion due to mechanical cell‐wall breakdown, increased porosity, and hydration, and (b) abrasion caused by friction and repeated contact. Three debarking modes were identified under varying hydraulic conditions: abrasion‐dominated, expansion–abrasion, and expansion‐dominated. These modes correspond to distinct transport regimes: congested (sliding–rolling), semi‐congested (sliding–rolling–floating), and uncongested (floating). The maximum expanded volume increased with water volume, while abraded volume first increased then decreased due to reduced wood‐chamber interactions. Under different water‐volume conditions, increasing flow velocity decreased the maximum expanded volume and increased the abraded volume, with minor variations in the mean values, but relatively large standard deviations. A conceptual model integrating both expansion and abrasion processes is further proposed to describe physical wood debarking during fluvial transport. This study provides first experimental evidence and conceptual framework for understanding large wood degradation in rivers.

Stationary precipitation measurements are frequently affected by undercatch errors, which are particularly pronounced in cold and alpine regions with strong winds. Since gridded precipitation products used in land surface modeling are often derived from spatial interpolation of meteorological station data, these measurement errors propagate directly into gridded datasets. In this study, we train monthly Generalized Additive Models (GAMs) using undercatch corrected station observations, with geographical exposure and terrain elevation as predictors, achieving R 2 values above 0.76 in Leave-One-Out Cross-Validation. We apply these models to generate monthly undercatch correction factors for Austria and – combined with an exposed terrain penalty – use them to adjust existing station-based gridded precipitation products. We validate the undercatch correction using the conceptual rainfall-runoff model COSERO across Austria and in two high-alpine reservoir catchments: Kölnbrein and Schlegeis. Our results demonstrate that retrospectively corrected precipitation reduces runoff simulation biases across Austria, especially in catchments above 1500 m elevation, and closes the water balance in both alpine study regions where uncorrected data showed runoff deficits exceeding 20 %. Biases in snow depth simulations – assessed using the physically-based snowpack model Alpine3D and validated against stereo-satellite observations – decrease from a median difference of −0.87 to +0.15 m. Additionally, undercatch-corrected precipitation enables more realistic simulations of snow covered area during the melting season and long-term glacier volume changes. The proposed method shows promising results in both alpine case study catchments and across Austria, highlighting the importance of accounting for undercatch errors in high-alpine terrain and indicating the need for further research into their magnitude at high elevations.

The availability of large-sample hydrometeorological datasets, now widespread across many regions worldwide, has changed hydrological catchment modelling. Assessing model performance is an essential component of any modelling exercise, and an important question is how to interpret performance measure values. Performances of uncalibrated bucket-type models vary significantly across regions and can reach NSE values of 0.8 or higher, particularly in humid or snow-dominated catchments. This implies that using a fixed value for a performance measure to judge model performance, as sometimes suggested in the literature, is inappropriate. Instead, one should consider that, given local hydroclimatic conditions and the quality of the available data, the performance we should expect from any model in a particular catchment can vary widely. At the same time, a perfect fit (NSE value of 1) is usually impossible to achieve due to errors and uncertainties in the model and data. Therefore, it is helpful to compare model performances to lower and upper benchmarks. The purpose of this study was two-fold. First, we examined how to compute lower bounds, including determining appropriate ensemble sizes, assessing the effects of parameter ranges, deciding whether to use random or regional parameter sets, and evaluating how best to aggregate the ensemble of simulations. We also examined the relationships between lower and upper benchmarks and catchment characteristics. Secondly, we utilised these findings to compute both lower and upper benchmarks for many of the existing large sample datasets. By providing these values to the modelling community, we aim to facilitate the broader use of lower and upper benchmarks in large-sample hydrological modelling studies. We argue that these values are valuable as they provide a basis for evaluating model performance across the various large-sample datasets. This will allow assessment of model performance, considering what one could and should expect for a particular catchment.

Although geothermal energy is a low‐carbon resource, some powerplants can coproduce and release , which is then subject to emissions monitoring and sometimes financial penalty. Accurate attribution of these emissions is difficult because exploitation of the reservoir alters hydrological flows that in turn modify surface fluxes. Hence, emissions metered through the powerplant do not represent the true anthropogenic impact. This study develops a lumped‐parameter model of a geothermal reservoir that tracks pressure evolution, dissolved and emissions through the powerplant and the Earth's surface. Closed‐form analytical solutions are obtained for reservoir degassing, outflow reversal, delayed reinjection returns and mineralization reactions, and these are calibrated against emissions data from the Rotokawa and Ngā Tamariki geothermal fields in New Zealand. The interplay of characteristic timescales for pressure equilibration, loss, reservoir residence, breakthrough delay and mineralization are shown to control whether a field's emissions decline, stabilize or rebound over decades. Calibrated models indicate that plant‐metered emissions at Rotokawa overcount the true anthropogenic impact by 17–45 kt , because production suppresses surface emissions that are not credited to the operator. storage in geothermal reservoirs is shown to be fundamentally different from conventional carbon capture and storage, as it is a flow‐through system in which dissolved only resides temporarily. Long‐term sequestration requires either mineralization or reinjection at sufficient distance to delay returns. Field‐specific models and repeat flux surveys are needed to constrain these effects and support emissions accounting based on true anthropogenic impact.

Precipitation estimates are subject to substantial errors, but their propagation into streamflow simulations remains an open question. This study uses a combinatorial approach to investigate this problem. First, the Raven hydrological modeling framework is used to identify behavioral model configurations by accounting for both parameter and structural uncertainties. Second, unlike conventional approaches using synthetic perturbations, we use empirically derived precipitation differences from 405 observed–estimated satellite and reanalysis-based pairs and use them as estimates of precipitation error. Third, we characterize the relationship between precipitation input errors and resulting streamflow errors ( P – Q error relationship) using two methods: the random forest model and precipitation elasticity analysis. Finally, streamflow uncertainty is partitioned using Shannon entropy and Sobol indices. The methodology is applied to 16 Canadian Prairie catchments; the model is calibrated over 2008–23 to obtain behavioral parameter sets, which are then used to simulate streamflow under 405 precipitation estimates for 2018–23. Results show that precipitation-induced streamflow errors depend strongly on model parameterization, particularly as input errors increase. Behavioral parameter sets do not exhibit uniform error propagation despite similar calibration performance, highlighting the critical role of model configuration in shaping the P – Q error relationship. This relationship is captured through precipitation elasticity, with higher elasticity leading to greater error amplification. Furthermore, Shannon entropy decomposition reveals substantial precipitation–parameter interactions that are not fully captured by the Sobol method. Overall, our study shows that precipitation uncertainty strongly influences streamflow uncertainty in the Canadian Prairies, while commonly used performance metrics [Kling–Gupta efficiency (KGE), Nash–Sutcliffe efficiency (NSE), etc.] do not necessarily indicate model robustness to input errors.

High-resolution, reliable surface meteorological estimates are essential for hydrologic modeling, climate research, and disaster management. Previous phases of the North American Land Data Assimilation System (NLDAS-1 and NLDAS-2), developed jointly by NOAA, NASA, and academic partners, established widely used observation-driven land hydrology modeling environments over the contiguous United States, supporting applications such as drought monitoring, water resource management, agricultural planning, and land–atmosphere research. However, increasing scientific and operational demands require finer spatial resolution, broader geographic coverage, and near-real-time availability. Building on this foundation, we present the North American Land Data Assimilation System phase 3 (NLDAS-3), which expands coverage to all of North and Central America and increases spatial resolution to 1 km. We developed a 1-km, hourly gridded surface meteorological dataset available in both retrospective and near-real-time modes. NLDAS-3 integrates satellite-based precipitation products and reanalysis fields from NASA and partner agencies. Meteorological variables are downscaled using elevation-based corrections to better represent fine-scale spatial variability across complex terrain. Evaluation against more than 3000 ground stations and 10 widely used reference datasets demonstrates that NLDAS-3 consistently outperforms existing products. Relative to NLDAS-2, NLDAS-3 achieves substantial average reductions in root-mean-square error (RMSE), including 18% for precipitation, 10% for air temperature, 42% for surface pressure, and 22% for longwave radiation. Extended triple collocation analysis further corroborates these improvements. Overall, NLDAS-3 represents a significant advancement in the quality, consistency, availability, and spatial resolution of surface meteorological data, providing a valuable resource for research and operational applications across North and Central America.

Soil-Vegetation-Atmosphere Transfer (SVAT) models are essential tools for simulating and underploting the dynamic interactions governing water balance components within forest ecosystems. These models are widely employed to predict hydrological responses to environmental change, including the impacts of shifting meteorological conditions on forested landscapes. Despite their usefulness, the reliability of SVAT models is frequently compromised by uncertainties arising from incomplete or imprecise input data. These limitations often result in model assumptions that may lead to over- or underestimation of critical water balance components such as groundwater recharge. In order to improve the accuracy of SVAT models, observed soil moisture data are integrated to enhance parameterization processes by aligning simulated outputs with measured values. However, uncertainties remain regarding the selection of representative soil moisture profiles for calibration and the extent of measurements necessary to robustly characterize a forest plot. To address these challenges, the present study explores the spatial variability of soil moisture across two forested plots with contrasting soil and vegetation conditions by the deployment of an extensive network of soil moisture probes in 11 profiles per plot. The influence of soil moisture variability on the adjustment of model input parameters during the calibration process and its subsequent impact on the computation of groundwater recharge is evaluated. The findings reveal that soil moisture variability at the plot characterized by a heterogeneous soil was greater, both horizontally and in depth, throughout the study period. These patterns of variability are also mirrored in the different parameter sets obtained from the calibration of the LWF Brook90 model, based on the recorded soil moisture time series in each of the 11 profiles per plot. The most significant variation is observed in the infiltration and hydraulic soil parameters, whereby this is more pronounced at the plot with heterogeneous soil structure. When examining the groundwater recharge rates calculated using the 30 best-performing parameter sets for each of the 11 profiles, both plots exhibited comparable temporal patterns but substantial differences in total groundwater recharge volumes across profiles. The results also suggest that model-inherent…

Despite the widespread use of physically based groundwater models, their deployment at the catchment scale remains challenging and time-consuming. HydroModPy was developed to address this gap by enabling automated and streamlined multi-site development of hydrogeological models at the catchment scale. This open-source Python toolbox facilitates the construction, execution, calibration, and analysis of unconfined shallow groundwater models. The current version integrates well-established geospatial tools, such as Whitebox Tools, and hydrogeological libraries with FloPy-driven MODFLOW-NWT simulations, along with optional particle-tracking and solute transport modules (MODPATH and MT3DMS), to provide a fully scriptable, end-to-end workflow. Automation is achieved through dedicated functions and classes capable of performing watershed delineation from digital elevation models, preparing spatial and temporal recharge forcings, generating computational meshes and vertical discretization schemes, assigning model parameters, and running simulations in steady-state or transient modes. The overall framework supports systematic and reproducible calibration routines that leverage subsurface data such as groundwater head measurements, as well as surface observations – including stream network maps and stream intermittency patterns – to constrain model estimates of aquifer hydraulic properties. Model outputs and provenance metadata are exported in standard geospatial formats to ensure interoperability and alignment with FAIR data principles. Built-in visualization tools and integration with Jupyter Notebooks support interactive exploration, teaching applications, and fully reproducible analyses. In this technical note, we present the HydroModPy architecture and its core functionalities, demonstrate model deployment across various hydrogeological contexts, and discuss ongoing and planned developments for future versions of this collaborative tool. The codebase is modular and extensible, making it suitable for adoption by a broad user community. Planned enhancements include tighter coupling with land-surface or ecohydrological models adding new numerical solvers, integrating advanced calibration and uncertainty quantification algorithms, and improving user interfaces to facilitate application in various environmental settings. HydroModPy contributes to improving the…

Data‐driven rainfall–runoff models have advanced rapidly, yet the majority of large‐scale applications still rely on lumped inputs that smooth out spatial variability in precipitation, temperature, and landscape properties. This simplification can introduce substantial biases in flood peaks, hydrograph timing, and water balance estimates. Here, we develop a deep learning framework that ingests spatially distributed (gridded) meteorological forcings and is evaluated across 921 catchments in North America using the CAMELS‐SPAT data set. We compare two strategies for fusing spatial and temporal information, namely a spatial‐to‐temporal (S2T) architecture and a temporal‐to‐spatial (T2S) architecture, and assess the added value of gridded catchment attributes and finer spatial resolutions. The T2S approach consistently achieves the best performance, increasing median Nash–Sutcliffe Efficiency (NSE) from 0.60 (lumped baseline) to 0.65 and Kling–Gupta Efficiency (KGE) from 0.63 to 0.69, with NSE improvements observed in 75.4% of catchments. Performance gains are consistent across catchment sizes and aridity regimes, with enhanced skill in low‐flow, high‐flow, and peak timing representations. Integrated Gradients (IG) analysis shows that the model learns physically coherent spatial behavior, highlighting grid cells near the outlet and identifying dominant meteorological drivers of peak runoff events. These results demonstrate the practical value of spatially distributed forcings for data‐driven hydrologic modeling and outline a pathway toward interpretable, large‐scale streamflow prediction.

Vapor pressure deficit (VPD), a key indicator of atmospheric dryness, has markedly increased across most global land areas due to progressive warming. Although site‐scale studies show that increasing VPD may enhance plant transpiration (Tran) and influence water and energy fluxes between land and atmosphere, this relationship has not been thoroughly investigated on the global scale. Here, we apply a binning approach to quantify the contribution of VPD to global Tran by ruling out the effect of soil moisture. To this aim, we integrate satellite‐based observations, reanalysis data sets, and simulations from sixteen Coupled Model Intercomparison Project Phase 6 (CMIP6) models into a coherent framework. VPD shows a broadly positive influence on Tran over most parts of the global land consistently across all data sets. Exceptions occur in tropical rainforests where reanalysis data and CMIP6 model outputs suggest a negative VPD‐Tran interplay. Attribution analysis suggests that increasing leaf area index reduces the positive effect of VPD on Tran in reanalysis data sets and CMIP6 models compared to satellite‐based observations. This discrepancy may reflect differences in the representation of the effective canopy‐level response to high atmospheric water demand, potentially involving stomatal regulation, plant hydraulics, and vegetation structure. Such differences may contribute to an underestimated transpiration response to VPD. These findings highlight the need to better represent physiological processes, hydraulic processes, and vegetation structural characteristics in process‐based models to capture vegetation responses to increasing water stress.

Accurately simulating overland flow in vegetated landscapes remains a challenge in hydrological modeling due to the complex interactions between vegetation, surface roughness, and soil infiltration. This study evaluates multiple methods for estimating Manning's roughness coefficient and explores the influence of vegetation on infiltration parameters, namely saturated hydraulic conductivity ( K sat ) and wetting front suction (Psi), using the OpenLISEM model. Based on 132 artificial rainfall experiments across 22 sites in southwest Germany, the model was calibrated and validated against observed runoff data, incorporating both depth-independent and depth-dependent roughness formulations. Incorporating water depth-dependent roughness into the model can improve its performance in simulating overland flow. Beyond roughness effects, vegetation was shown to significantly alter soil hydraulic properties, particularly saturated hydraulic conductivity ( K sat ). Paired site comparisons revealed that increased vegetation cover corresponded with higher infiltration capacities, emphasizing vegetation's role not only in surface resistance but also in enhancing subsurface water fluxes. The findings demonstrate that models must account for both surface and subsurface impacts of vegetation to improve runoff predictions.

We evaluate the performance of the Super-Fast INundation of CoastS (SFINCS) hydrodynamic model for simulating riverine floods, combined with a fully automated open-source data preprocessing pipeline. To do this, we assessed the simulated extent of 499 historic flood events against the satellite derived flood extents using the Critical Success Index (CSI) as a performance metric. We utilised simulated discharges from the Global Flood Awareness System (GloFAS) hydrological model and found that SFINCS performance improved with upstream basin size, with a global mean CSI of 0.42 for basins with large upstream area (>1000 km 2 ) and a CSI of 0.29 for basins with small upstream area (<50 km 2 ). Our results illustrate the importance of accurate discharge data input to flood hazard simulations. When the (globally simulated) GloFAS data is replaced with observed discharge data for ten events in the US, the CSI improved from 0.39 to 0.67. These results suggest that global hydrological model performance limits the accuracy of the flood hazard simulations. Our findings also showed a significant improvement in the CSI (from 0.37 to 0.57) when changing to a higher-resolution elevation input by contrasting a ∼1 m digital elevation model (DEM; 3DEP) with our default ∼30 m global DEM (FABDEM) in six US events. Sensitivity analysis of bathymetric calculations revealed a systematic underestimation of the default 2-year return period estimated by GloFAS discharge, likely driven by underrepresentation of annual block maxima, which resulted in underestimated channel dimensions. All of these factors resulted in a loss of detail, which impacted model performance, especially in smaller headwater rivers. We recommend to improve the estimation of bathymetry, for instance by employing the “gradually varying solver” method or using data from the SWOT mission. Furthermore, incorporating additional validation data which ideally includes flood depth measurements can largely enhance our understanding of the model performance.

The simulation of streamflow is a complex task due to its intricate formation process. Existing single models struggle to accurately capture the stochastic, non-stationary, and nonlinear dynamics of basin streamflow in changing environments. This study combined process-driven hydrological mechanism models with data-driven deep learning models, considering various factors like hydrology, meteorology, environment, and the interconnected effects of upstream and downstream rivers, to develop an interpretable hybrid streamflow simulation model. The study collected multiple external variables to better understand the hydrologic system complexity and used the Maximum Information Coefficient (MIC) to analyze their relationship with streamflow. Subsequently, the Xinanjiang (XAJ) model with physical mechanisms was employed, alongside the TCN-GRU model integrating Temporal Convolutional Network (TCN) and Gated Recurrent Unit (GRU), for separate streamflow simulations. Furthermore, a robust integration method was adopted, realizing nonlinear ensemble through Random Forest (RF), thus establishing the hybrid XAJ-TCN-GRU model. This model exhibits promising results in simulating streamflow in four different basins in China, achieving high Nash-Sutcliffe Efficiency (NSE) values with 0.991, 0.971, 0.984, and 0.986 for the Wuding River, Chu River, Jianxi River, and Qingyi River respectively. In terms of streamflow simulation, flood simulating, and interval simulation, this model outperforms other benchmark models. Additionally, the study quantified the contributions of each hydro-meteorological variable to the long-term streamflow trend using mean absolute SHAP values (SHAPABS), Feature Importance (FI), and Permutation Feature Importance (PFI), thereby enhancing the model's external interpretability. The results of this study are of significant importance for optimizing water resource management and mitigating flood disasters.

Wildfires fundamentally alter snow accumulation and melt dynamics by modifying forest canopy structure, snow albedo, and snowpack physical properties. Despite the growing body of work on fire impacts on snow, few studies have used physically based models to simulate post‐fire changes, and none have examined the global parameter sensitivities that govern these dynamics. In this study, we apply the Structure for Unifying Multiple Modeling Alternatives in a high elevation Colorado watershed to identify and evaluate the model parameters that influence post‐fire snowpack evolution. We conduct a global sensitivity analysis, then sample sensitive parameters to identify behavioral parameter sets with respect to airborne LiDAR‐based snow depth in burned and unburned forests. Among behavioral parameter sets, we identify 4 parameters that differ significantly in burned versus unburned forests. These include the albedo decay rate (45% faster in burn than unburned forest), maximum canopy interception for rain (60% lower), minimum temperature for snow unloading (3.9°C warmer), and snow surface roughness (902% higher). Model simulations using behavioral parameter sets for burned and unburned conditions indicate 17% lower peak absolute cold content on average and 17% lower peak snow‐water equivalent (SWE) in burned conditions on average, with differences in depth, SWE, and cold content increasing throughout the snow season. These findings highlight the combined influence of canopy loss and accelerated albedo decay on snow energy balance and reveal additional post‐fire snow processes that are not typically represented in existing modeling frameworks. This study therefore provides a process‐level foundation for improving the simulated representation of post‐fire snow dynamics.

Hydrologic models applied in snow‐dominated basins are typically evaluated by comparing simulated outputs, such as snow water equivalent and streamflow, with observations. However, this approach may overlook the energy balance aspect of snow modeling. Snowmelt initiation and rate are primarily determined by the net energy balance, which varies across snow models due to diverse model inputs, parameterizations, and layering structures. Here, we discuss how decisions regarding model parameters must be made in conjunction with energy budget considerations and model layer structure to model snowmelt effectively, and highlight the importance of incorporating multiple validation data sources, including snow surface temperature. We use the two‐layer Distributed Hydrology Soil Vegetation Model in California's upper Tuolumne River watershed as a case study and explore the influence of model top layer thickness (TLT) on simulating snowmelt, focusing on its significant effects on modeled snow surface temperature ( T ss ), energy propagation through the snowpack, and the timing of snowmelt. Additionally, we examine the interplay between model forcing and parameters, specifically incoming longwave radiation and albedo, to reveal how variations in these factors interact with TLT and shape snowmelt simulations. Our results reveal that TLT has a significant impact on diurnal variations in T ss , snowpack energy distribution, and the overall timing of snowmelt. Lower albedo enhances melting rates more strongly in the thinner upper layers, and biases in incoming longwave radiation can either amplify or mitigate these effects. Snow surface temperature was a key variable to investigate, and we recommend considering it in snowmelt investigations.

Snow plays a critical role in cold-region environments, from climate regulation to water storage and from spring freshet to ecosystem services. However, quantification of snow accumulation and redistribution across spatio-temporal scales is challenging, owing to non-uniform snowfall, topographic attributes, wind redistribution, and interception and sublimation by vegetation. Manual snow surveys are often conducted to improve understanding of snow processes across scales and to support societal needs (e.g., seasonal water supply projections, flood forecasting, reservoir filling, etc.). Nevertheless, manual snow surveys are time-consuming and expensive, which limits their temporal frequency and are rarely conducted in remote locations. This manuscript presents snow data from manual snow surveys conducted from 2009 to 2024 across four small watersheds in high-latitude boreal plains of northern Alberta in Canada. The dataset includes various snow properties, including depth, density and snow water equivalent across different watersheds and land cover types (i.e., Open, Flat Low Lying, Coniferous and Mixed Deciduous). The dataset is expected to be useful in understanding snow processes in high-latitude boreal plains, to help calibrate snow models and to validate airborne lidar or satellite-derived snow depth measurements.

Hydrological impact assessments under climate change commonly rely on conventional modeling chains where climate projections are bias-corrected before being used in hydrological simulations. While this improves agreement with historical observations, it can introduce methodological uncertainties, reduce the diversity of climate ensembles, and smooth out extreme events. Asynchronous methods have been proposed as an alternative, allowing hydrological models to be calibrated directly with raw climate model outputs. However, fully asynchronous methods often fail to capture the timing of key hydrological processes, especially in snow-affected regions. This study introduces and evaluates a semi-asynchronous calibration on monthly observed discharge distributions to address these limitations. Using the physically based WaSiM model, we compare the semi-asynchronous, fully asynchronous, and conventional methods across ten snow-influenced catchments in southern Quebec, Canada, under historical and future climate conditions. The results show that while the fully asynchronous and semi-asynchronous methods perform well in preserving streamflow distributions and high-flow extremes, only the semi-asynchronous method succeeds in restoring the seasonal timing of key processes such as snowmelt and low flows. The semi-asynchronous method notably reduces intermodel variability in streamflow and snow water equivalent compared to the fully asynchronous approach. It also exhibits seasonal dynamics that closely align with observations and the conventional method, despite relying on uncorrected climate inputs. In contrast, the fully asynchronous method shows signs of desynchronization, with unrealistic snowmelt timing and elevated variability across projections. The conventional method, while more stable in the historical period, exhibits an increase in intermodel variability under future conditions, likely due to divergent magnitudes of projected change across climate models. Compared to the conventional method, which benefits from stable and consistent simulations but tends to dampen extremes through bias correction, the semi-asynchronous approach offers a compelling alternative. It strikes a different balance between realism, ensemble diversity, and the ability to represent extreme flood events, making it particularly valuable for future-oriented climate impact assessments.…

Accurate simulation of snowmelt runoff (SMR) is critical for water resource management. However, despite the abundance of global hydrological models, little is known about their SMR performance. This study presents a comprehensive evaluation of SMR across 15 state-of-the-art large-scale models and runoff products by focusing on their biases in first-order indices, i.e., the total volume ( Q sum ), peak flow ( Q max ), and centroid timing (CTQ) of runoff in the snowmelt period. Then by introducing 1455 snow-dominated basins with diverse topography and vegetation complexities, we further proposed a novel model robustness metric to test how different models perform under increasing basin complexity, thereby allowing for a quantification on how they adapt to complex environmental conditions. Our results reveal that (1) most models exhibit underestimated Q sum and Q max and predict CTQ too early. These biases are particularly pronounced in regions such as the western United States, northern Europe, and northeastern China. (2) Model biases systematically increase with basin complexity, with CTQ exhibiting strong sensitivity to mean elevation and topographic variability, while Q sum and Q max being shaped more by mean elevation and the diversity of vegetation types in the basin. (3) The robustness assessment further shows that observation-constrained runoff products exhibit the most outstanding performance (i.e., low biases and strong adaptability to stern conditions), followed by the hydrological and land surface models. Notably, while global hydrological models generally exhibit stronger robustness in simulating Q sum and Q max , land surface models show a clear advantage in simulating CTQ, highlighting their structural strength in capturing melt timing rather than runoff magnitude. This study provides a large-sample benchmark for SMR evaluation and complements existing model assessment approaches by examining model performance across basin complexity gradients, offering useful insights for future model development and uncertainty reduction.

Streamflow prediction is essential for water resources management and flood mitigation but remains challenging in large basins due to spatial heterogeneity. Machine learning (ML) models have been widely applied to streamflow prediction, typically relying on explicit routing modules, whereas the potential of pure ML models remains underexplored. This study presents a purely ML‐based model (Convolutional Entity‐Aware Long Short‐Term Memory) for fully distributed streamflow prediction. Spatial cross‐validation across 50 gauges in the Amazon Basin yields a median Nash‐Sutcliffe Efficiency of 0.73 and an R 2 of 0.84, comparable to or exceeding benchmark models with routing modules. The model reduces annual peak flow bias (APFB = 0.12) and accurately reproduced spatial autocorrelation patterns (Pearson's r > 0.6, Moran's I > 0.8, and Geary's C < 0.2), indicating spatially consistent grid‐scale discharge predictions. These findings demonstrate the potential of a pure data‐driven ML model for fully distributed streamflow prediction in ungauged basins while ensuring spatial consistency.

Large‐sample deep learning runoff models rely primarily on aggregated catchment attributes, which may obscure informative spatial patterns. Using 531 watersheds in the contiguous United States, we develop a hierarchical deep learning framework that represents static watershed properties from watershed‐averaged raster aggregation to convolutional neural network encodings at fixed‐grid and adaptive multi‐patch scales, thereby providing a controlled test of whether increasingly explicit spatial encoding improves temporal generalization and spatial transferability to ungauged watersheds. Watershed‐averaged rasters perform comparably to curated attribute tables and show a slight advantage under the prediction‐in‐ungauged‐basins benchmark. Incorporating intermediate spatial structure via fixed‐grid encoding provides the clearest temporal‐split benefit relative to watershed‐averaged aggregation, with 26.2% of watersheds showing uncertainty‐supported Nash–Sutcliffe efficiency improvement and 6.0% showing uncertainty‐supported degradation, although this advantage weakens under spatial holdout. Under spatial holdout, diagnostic analyses indicate that fixed‐grid encoding tends to outperform watershed‐averaged raster aggregation in higher‐elevation, steeper watersheds. The adaptive multi‐patch level further reveals that finer spatial representation does not necessarily provide consistent gains, highlighting a trade‐off between spatial detail and computational demand. Overall, this framework provides practical guidance for selecting raster‐based spatial representations in large‐sample deep learning runoff prediction.

Hydrologic models are often calibrated using streamflow alone, but increasing availability of in situ and satellite-based observations provide numerous opportunities to constrain model outputs and improve process representation. However, as new observation data emerges, it is often unclear whether calibration with additional data would inform or misinform streamflow prediction. Here, we carry out a multi-observational sensitivity and uncertainty analysis using the U.S. Geological Survey's National Hydrologic Model (NHM) in four headwater catchments in the Upper Colorado River Basin. We use seven different observational data products that pertain to discharge, snow water equivalent, snow-covered area, soil moisture, and evapotranspiration. Informative model parameters are identified using the Morris screening method across all data sets, followed by parameter estimation and streamflow performance assessment using a Latin Hypercube Sample Monte-Carlo filtering approach. Results show that an increased number of informative parameters are determined through the screening process with the use of observation data representing terms beyond streamflow, and that forcing corrections and rain-snow partitioning parameters are particularly impactful to the model fit to observations. Multi-objective Monte Carlo filtering reduces the number of behavioral parameter sets, and estimated parameter values can depend strongly on the observation data criteria. Evapotranspiration is informative for streamflow prediction across all catchments included in this study, but snow and soil moisture datasets are only informative in some. These results provide new insight into the variable value of alternative observation data for streamflow prediction and highlight challenges related to model/observation scale mismatches, compensating errors, and misinformative data.

Differentiable hydrologic models that use machine learning to infer parameters for process-based models show promise for both prediction and inference. However, these models are often developed with time-varying parameters, despite evidence that such flexibility can undermine physical consistency and yield only marginal predictive improvements over simpler static approaches. In this study, we revisit the comparison between static and dynamic differentiable models across 531 CAMELS-US basins, evaluating key architectural choices: (1) neural network type (multi-layer perceptron (MLP) vs. long short-term memory network (LSTM)); (2) process model configuration (single- versus ensemble-parameter estimation); and (3) comprehensive versus alternative input feature sets. Using the Hydrologiska Byråns Vattenbalansavdelning (HBV) conceptual model, we find that although ensemble parameterizations improve performance relative to single-parameter configurations, they also alter the conclusions about the relative value of network architecture: LSTMs outperform MLPs in single-parameter configurations, but static, MLP-based ensembles achieve performance comparable to dynamic, LSTM-based ensembles despite their simpler structure. Additionally, we find that LSTM-estimated parameters rarely exhibit meaningful temporal variability despite their time-varying inputs, and when they do, this temporal variability may reflect hydrologic model equifinality rather than process dynamics. We further show that models using only latitude and longitude as static inputs achieve spatial generalization comparable to models using comprehensive feature sets describing climate, topography, geology, soils, and land cover. Similarly, temporal generalization is retained even when comprehensive features are replaced with physically meaningless values. These results raise the possibility that static inputs may function less as direct representations of physical basin processes and more as spatial proxies for generalization in space or as site identifiers when generalizing in time. Overall, our results support reduced complexity in differentiable hydrologic modeling to provide greater transparency while retaining predictive performance.

The use of deep learning models in hydrology is becoming an ever more prevalent application in operational flood forecasting. Such operational systems face performance degradation when transitioning from high quality reanalysis to meteorological forecast data with lower accuracy. This study investigates training strategies and Long Short-Term Memory network architectures to mitigate meteorological forecast-induced bias in maximum daily discharge predictions using the Extended LamaH- CE dataset and a subset of 451 basins. We systematically evaluated cross-domain generalization, transfer learning approaches, Encoder–Decoder LSTMs, Sequential Forecast LSTMs, and the role of input embeddings and integrating past discharge observations. The results show that domain shifts between reanalysis and forecast data lead to substantial skill loss, with median Nash–Sutcliffe Efficiency decreasing from 0.58 to 0.33. Among the tested strategies, the Sequential Forecast LSTM demonstrated the most stable improvements, achieving a median NSE of 0.63. Integrating recent discharge observations further enhanced performance, raising median NSE to 0.71 and surpassing even the reanalysis-driven baseline. In contrast, integrating archived forecasts or using more complex input embeddings did not yield consistent benefits and in some cases degraded model stability. Basin-level analysis reveals that forecast skill improvements compared to our baseline are not uniformly distributed across catchment types: the largest gains are concentrated in arid and precipitation-limited catchments, while alpine and snow-dominated catchments, despite experiencing the largest meteorological domain shift, show smaller improvements. This is likely because the LSTM cell state retains strong seasonal signals and thereby compensates for forecast input bias through its long-term memory mechanism, particularly in alpine and snow-dominated catchments where strong seasonal cycles dominate the hydrological response. These findings highlight the value of training strategies that allow models to directly learn bias correction during forecast transitions, emphasize the operational potential of combining sequential processing with near real-time discharge observations and identify physiographic catchment characteristics as key modulators of forecast skill improvement across diverse hydroclimatic settings.

Deep learning hydrological models rely on static catchment attributes to characterize basin heterogeneity, but these traditional descriptors suffer from regional inconsistency and limited temporal representativeness. This study evaluates geospatial foundation model embeddings (from AlphaEarth and StefaLand) as transferable catchment descriptors for rainfall‐runoff modeling. Temporal validation across 531 CAMELS basins showed that deep learning models trained on AlphaEarth embeddings alone achieved comparable performance to those trained on traditional attributes (median Nash‐Sutcliffe Efficiency (NSE) 0.779 vs. 0.781). In a global prediction in ungauged basins (PUB) scenario with 3,434 basins, combining embeddings with traditional attributes improved median NSE from 0.683 to 0.720. A dual‐path gated model revealed structured spatiotemporal reliance patterns: models preferentially utilized embeddings during snowmelt transitions and in humid basins, with limited reliance in arid catchments. Independent validation with StefaLand embeddings from a land‐surface foundation model yielded consistent temporal validation results, supporting the broader use of learned embeddings as transferable catchment descriptors.

This study investigates the application of Physics‐Informed Neural Operators (PINOs) for solving the two‐dimensional shallow water equations (2D SWE) in the context of flood modeling. Unlike Physics‐Informed Neural Networks (PINNs), which require retraining for each new initial or boundary condition (BC), PINOs learn the solution operator itself, enabling fast inference across a range of conditions without retraining. PINNs have moreover been shown to struggle with complex bed topography due to the difficulty of satisfying the well‐balanced property (Tian et al., 2025, https://doi.org/10.1029/2025wr040052 ; Dazzi, 2024, https://doi.org/10.1029/2023wr036589 ), a limitation that PINOs can partially mitigate by incorporating data alongside the physics loss. We evaluate the method on experiments of increasing physical complexity: a radial dam break, constant BCs with and without friction, time‐dependent BCs, and a real‐world test case. The results show that PINOs capture key flood dynamics, particularly water depth, while reducing inference time by up to two orders of magnitude compared to numerical solvers. Relative test errors for water depth ranged from 0.3% for the radial dam break to 10.9% for cases with bottom topography, 7.3% with friction, and 9.0% under time‐dependent BCs. For the real‐world case, an additional data loss term yielded a water depth error of 25.8%. Larger errors were consistently observed for velocity components. The findings establish PINOs as a promising complement to traditional numerical solvers, offering a balance between computational efficiency and solution accuracy. Future work will focus on improving accuracy and extending the framework toward real‐world flood forecasting applications.

Precipitation is important for hydrological monitoring and modeling. The accuracy of mean areal precipitation (MAP) estimation relies largely on the configuration of the precipitation network. This study proposes a spatial pattern–based framework for optimizing rain gauge networks that uses the climatological precipitation field from high-quality reanalysis data as an external benchmark to evaluate MAP estimation. In contrast to entropy- and variance-based methods, the framework requires no long in situ records, explicitly matches both the magnitude and the full spatial distribution of the precipitation field, and can assess as-yet-uninstrumented locations. Using the Qingyi River basin as a case study, we employ the China Meteorological Administration Multisource Precipitation Analysis System (CMPAS) fusion data to characterize precipitation spatial patterns. The framework introduces two metrics, the mean areal precipitation bias (MB) and the Kullback–Leibler (KL) divergence, to quantify differences between gauge-derived and CMPAS-derived precipitation fields. Evaluation of the current network reveals significant MAP estimation discrepancies in subbasins with high precipitation variability. Through a significance index analysis of candidate gauges and a hierarchical optimization, strategic gauge placement reduced MAP bias from up to 16% to below 1% in the optimized subbasins. An independent hydrological validation with the Xin’anjiang model showed that the optimized network improved streamflow simulation, with the basin-average Nash–Sutcliffe efficiency increasing from 0.63 to 0.76, confirming that better MAP estimation propagates to improved hydrological prediction. The framework is computationally efficient and applicable to both network expansion and rationalization.

The changing Arctic climate alters the dynamics of melting and freezing in the ground. A methodology has been developed to assess small magnitude, ice-induced earthquakes, called frost quakes, by estimating thermal stresses in the soil in Oulu, Finland. Information on temporally and spatially varying soil properties, such as soil temperature and soil ice content, is required to calculate thermal stress. Developing this methodology further on a larger scale, over the whole country, is challenging due to a lack of in-situ measurements of those parameters with high spatial and temporal coverage. However, they can be simulated using land surface models, one of which is HydroBlocks. Previously, HydroBlocks has been applied in the contiguous United States. The goal of this paper was to configure the model in subarctic and arctic Finland. However, another challenge with large scale LSMs is that they often use coarse global soil datasets, and their simplified soil presentations can introduce errors in the modelling results. For that reason, using local soil data helps to achieve a more accurate representation of soil and improves the model performance. Hence, in this paper, the HydroBlocks model was run first by using coarser SoilGrids soil data and second, by using local soil data provided by the Geological Survey of Finland (GTK). HydroBlocks' ability to produce accurate snow accumulation and melt approximations, as well as estimate soil temperature and soil water content at different depths in Finland, has not been evaluated before. The snow model (snow water equivalent) and the modeled soil temperatures and soil water contents were compared with observational data to evaluate the model performance. For six observational snow water equivalent stations, with GTK's soil data the average RMSE and KGE were 29 mm and 0.34, respectively. The worst KGE was −0.30, and the best was 0.90. For the three observational soil stations, the soil temperature had an average RMSE and KGE of 1.3 °C and 0.7, respectively. The worst KGE was 0.22, and the best was 0.86. For the soil water content, using local soil data provided by the GTK significantly improved the modelling. The average RMSE and KGE were first, 0.15 vol/vol and −3.5, and after calibration, they were reduced to 0.03 vol/vol and increased to 0.09, respectively. For the calibrated model, the worst KGE was −0.8, and the…

Socio‐hydrology models explore interactions between human and hydrological systems, enabling the analysis of dynamic feedbacks between society and flooding. Simplified conceptual models of socio‐hydrology often treat the social or human domain as single homogenous entities, failing to account for disparities among the communities that make up a city, and thus underestimating the true impacts of flooding on vulnerable populations. We propose a disaggregated socio‐hydrology model that incorporates a Resilience domain to capture differences among communities within a single city. Using empirical socioeconomic data from United States metropolitan areas, we inform the model's representation of inequality and its influence on community‐level adaptive capacities. With the further addition of a disaster relief component, our approach captures community‐specific recovery trajectories, linking economic inequalities to urban flood resilience. Results show that a community's adaptive capacity is shaped by a city's overall growth rate and the degree of economic disparities between communities. We find that improving a city's economic growth rate can unintentionally widen disparities in community adaptive capacity, consistent with the Kuznets curve hypothesis. However, simulations show that equitable resource allocation reduces recovery gaps without hindering overall city growth. Our results emphasize the necessity of incorporating socioeconomic heterogeneity into socio‐hydrological models to inform equitable disaster response policies. By grounding resilience in empirical economic data, this study advances socio‐hydrology modeling to provide more actionable insights for disaster response and community‐based risk reduction strategies.

An important challenge with Machine Learning (ML) is its transferability; that is, whether an ML model trained on one set of data can be applied to a second set of data without requiring full retraining of the model. Transfer Learning (TL) addresses this challenge by transferring knowledge learned in the source domain (the data it was trained on) to the target domain (a second set of data that is statistically different but related, which the model was not trained on). This study investigates the use of TL for street‐scale nuisance flood forecasting by exploring whether an ML model trained for one set of streets can effectively forecast flooding for another set of streets in the same city. TL is explored using a Long Short‐Term Memory (LSTM) model trained on data for the flood‐prone streets of Norfolk City, Virginia. The results show that full‐weight retraining proved most effective and minimal retraining of only the output layer was insufficient. The advantage of TL was most pronounced when target data was limited, meaning data collected at the new water depth sensor location included generally less than 18 flood events. As the target data increased beyond 18 flood events, the benefit of TL diminished relative to training an ML model directly on the local flood events. These findings can assist cities as they implement street‐scale flood sensing systems to create accurate forecasts for new sensing locations that do not yet have sufficient data records to train a local ML model.

Accurate, timely, and actionable flood forecasting has potential for reducing flood risk but remains challenging due to uncertainties in meteorological forecasts, poor hydrological observations, and increasing forecast errors with lead time. While Long Short‐Term Memory (LSTM) models have surpassed conventional hydrological forecasting models, they often struggle to capture the timing and magnitude of extreme floods due to long‐range dependencies and static feature importance. Here, we propose a sequential LSTM with Multi‐Head Attention (SeqLSTM‐MHA) to improve 1–3 day ahead flood forecasts, explicitly using historical hydrometeorological and upstream hydrological observations. SeqLSTM‐MHA attained an NSE of 0.85–0.81 for water level and 0.74–0.71 for streamflow corresponding to t +1 to t +3 day lead, outperforming benchmark models (traditional LSTMs, bidirectional LSTM, CatBoost, and multiple linear regression). The ablation analysis shows that notable improvements at longer lead times ( t +2 and t +3) arise from the interaction of sequential forecasting, hierarchical upstream representation, and attention‐based temporal weighting rather than merely increased model capacity. Our model demonstrates strong skill in forecasting extremes (NSE >0.6) and correctly captures more than 80% of peak streamflow during the extreme events exceeding the 95th percentile threshold. Time‐resolved attention analysis shows that recent rainfall primarily drives immediate responses, while upstream flow propagation becomes more important at longer lead times. Overall, our findings demonstrate the potential of attention‐based LSTM for operational flood forecasting, particularly for high‐impact events.

Although objective functions (OFs) are widely discussed in the literature, many modelling studies still default to a few common metrics, without much consideration of their relative strengths and weaknesses. This paper systematically investigates the impact of OF choice on the representation of various streamflow characteristics across 47 conceptual models and 10 hydro-climatically diverse catchments selected from the CARAVAN dataset. We use eight different OFs for calibration, including the Kling–Gupta efficiency (KGE), Nash–Sutcliffe efficiency (NSE), a variant of each that uses a streamflow transformation, and four more recently proposed metrics. We further account for parameter uncertainty by retaining the top 500 parameter sets for each combination of model, basin and OF. We evaluate the representation of 15 hydrological signatures that capture a relevant selection of streamflow characteristics to determine generalisable strengths and weaknesses of individual OFs across different models and hydroclimatic conditions. A random forest importance analysis showed that OF choice is often more influential than model structure choice in determining signature representation, with catchment characteristics remaining the dominant control. While certain signatures, particularly those related to flow variability, are relatively insensitive to OF choice, other signatures exhibit large performance shifts across different OFs. As other studies have already noted, no single OF performs best across all signatures. Our consistent multi-model and multi-catchment experiment now provides generalisable guidance on which OF best reproduces which aspect of streamflow behaviour. This helps to better select objective functions most appropriate for a specific modelling purpose and provides insights into which metric combinations can cover a broader range of flow characteristics.

Accurate hydrological prediction in alpine regions remains challenging due to complex cryospheric processes and limited observational records. Traditional hydrological models are often subject to structural uncertainty, whereas purely data‐driven deep learning (DL) models may lack physical interpretability under non‐stationary climate conditions. Here, we present a distributed differentiable framework for alpine basins that combines bottom‐up physical constraints with top‐down global learning to test frozen‐ground structural hypotheses while mitigating the effects of local data scarcity. Specifically, we developed differentiable Coupled Routing and Excess STorage model (dCREST), a distributed differentiable physics‐informed model that enables the joint diagnosis of model structure, parameter dynamics, and spatial transferability. Using the Headwater Area of the Yellow River (HAYR) as a test bed, we investigated the hydrological role of frozen‐ground processes and the value of global pretraining for runoff simulation in data‐scarce alpine regions. Results indicate that conventional single‐layer soil structures are unable to adequately capture permafrost‐controlled hydrological dynamics. In contrast, an explicit dual‐layer soil structure coupled with dynamic cryospheric parameterizations substantially improves the simulation of seasonal runoff partitioning and long‐term hydroclimatic responses. Furthermore, pre‐training on globally distributed catchments enables dCREST to learn transferable rainfall‐runoff relationships, enhancing modeling performance despite limited local observations. This transfer learning strategy increases the average Nash–Sutcliffe efficiency (NSE) from 0.35 in the current operational system to 0.63. Our findings demonstrate that differentiable hydrological modeling provides a powerful framework for simultaneously diagnosing model structural deficiencies and improving runoff prediction in data‐scarce alpine basins through physically constrained transfer learning.

Airborne electromagnetic (AEM) surveys offer rapid, cost‐effective subsurface imaging, yet converting their electrical resistivity (ER) models into physically meaningful hydraulic property fields for groundwater models remains a challenge. We develop and demonstrate a data‐driven workflow for an unconsolidated sedimentary aquifer system in Scott Valley, northern California, USA, that incorporates AEM ER data with borehole logs to build and calibrate a geologically heterogeneous groundwater‐surface water model. ER and texture observations are first combined through consensus clustering into five meta‐texture classes; a probabilistic ER‐texture transform then converts ER data to cell‐scale texture probabilities. These probabilities and borehole data are combined using Texture2Par to create a three‐dimensional texture model, which is translated to grid‐scale hydraulic conductivity and storage via power‐law averaging. During calibration, we parameterized the ER‐texture distributions, essentially allowing parameter estimation to adjust the estimated textures along the AEM flight lines. The texture‐based groundwater‐surface water model attains the same high goodness‐of‐fit as the previous zonal calibration (Fort Jones streamflow Nash‐Sutcliffe efficiency = 0.84; groundwater heads r 2 = 0.98), with more geologically plausible heterogeneity and improved simulation of groundwater‐driven seasonally low streamflow. The proposed ER‐to‐texture workflow provides an adaptable workflow for embedding AEM information into basin‐scale groundwater models.

Soil freezing is a major cold region process that influences hydrological response of northern catchments, in particular during winter rainfall and snowmelt events. In land surface models, frozen soil infiltration is difficult to represent because soil structure and hydropedological processes vary at scales finer than the model grid. This is particularly true in operational modeling, where physical process integration must balance performance improvements against computational efficiency and complexity. In this study, we propose a new configuration of the Soil, Vegetation, and Snow (SVS) model used within the operational prediction systems of Environment and Climate Change Canada (ECCC) that enhances frozen soil infiltration by reducing both surface runoff and sub-surface lateral flow. We assessed the effects of this new configuration (Fr-Inf) on streamflow simulations at more than 580 hydrometric stations located in the Great-Lakes and Saint-Lawrence domain over a five-year period. Fr-Inf significantly improves the Kling-Gupta Efficiency (KGE) compared to the default soil freezing configuration (Fr; ΔKGE = 0.28) but slightly underperforms the configuration of SVS without frozen soil (noFr; ΔKGE = −0.07). Strong degradations relative to the no freezing configuration (ΔKGE < −0.5) are observed only at 4 stations with Fr-Inf (< 1 %) as opposed to 172 stations under the Fr configuration (33 %), highlighting the robustness of the approach. To ensure that the proposed change is also acceptable in the context of operational numerical weather prediction, an evaluation of its impact on soil freezing depth as well as screen-level temperature and dew point temperature predictions is performed against in-situ observations. These results support the potential operational implementation of soil freezing at ECCC for numerical weather and streamflow prediction.