AI for hydrology

Accurate and continuous streamflow monitoring can support effective water resource management, yet substantial gaps persist in national streamgage networks due to operational challenges and costs. Non‐contact, image‐based techniques such as Space‐Time Image Velocimetry (STIV) offer a safe, scalable and, cost‐effective alternative for measuring surface flow velocity in rivers, which allows for streamflow calculation, especially during hazardous flood events. Traditional STIV methods, however, often require manual parameter tuning and struggle with complex flow conditions. This study introduces Auto‐STIV, a novel deep learning framework that fully automates the estimation of both flow direction and velocity magnitude from video imagery. The proposed approach utilizes a two‐stage Convolutional Neural Network model. The first stage predicts the primary flow direction by analyzing a stack of Space‐Time Images (STIs) generated from a series of candidate orientations. The second stage estimates the velocity magnitude from the STI corresponding to the predicted direction. The model was trained on a comprehensive data set collected under diverse, controlled laboratory conditions and evaluated using real‐world footage from the Colorado River. Results demonstrate that Auto‐STIV outperforms the traditional autocorrelation‐based STIV method, exhibiting higher accuracy and greater robustness in the presence of challenging environmental factors such as variable lighting, low‐contrast surface textures, and complex flow patterns. This research underscores the potential of deep learning to enhance the capabilities of image‐based velocimetry, that may support more reliable, scalable, and automated streamflow monitoring systems.

Concurrent floods and concurrent droughts in nearby catchments pose challenges to risk assessment and water management. Climate change is affecting extremely high and low discharge, but the complex interplay between changes in individual catchments, and in the dependence across catchments, make it difficult to provide accurate assessments of the occurrence probabilities of concurrent extremes. In this work, we use a contemporary statistical deep learning model to capture concurrent river floods and droughts in four catchments in the Upper Danube basin, based on discharge simulated by a hydrological model driven with large ensemble climate model output. The statistical model is able to accurately capture the tail behavior of the simulated discharge, which we assess by making use of the large available sample. We subsequently use our statistical model to study changes in joint tail behavior of discharge over time, finding that both compound flooding and drought‐like conditions are becoming increasingly likely toward the end of the 21st century under a high‐emission scenario. In particular, our results highlight that changes in the dependence structure of extremes strongly contribute to the detected changes, an aspect that would be difficult to capture with traditional approaches. This work paves the way for highly flexible, general inference on compound extremes in hydrological applications, and demonstrates key advantages of using statistical deep learning in this setting.

Groundwater is one of the most valuable natural resources, yet it is increasingly under pressure from climate change and human activities. Shallow groundwater depth (GWD) is a key indicator for evaluating groundwater availability and flood hazards, yet the measurements are often too sparse to capture its variability both in time and space. In this study, we evaluated hybrid data‐driven machine learning (ML) models that combine the strengths of existing hydrological simulations with the predictive power of ML methods for improved GWD estimation in Denmark. The hybrid model reduced mean bias of hydrological simulations from 1.72 to 0.22 m and decreased the Root Mean Square Error by about 1.64 m at national scale. Furthermore, we demonstrated that increasing the availability of GWD observations over time and space enhances model performance. We also illustrated the flexibility and effectiveness of the hybrid approach for GWD estimation across scales, and results showed that the hybrid model developed with coarse spatial resolutions can be easily used for high‐resolution GWD estimation without significant decrease of accuracy. This study highlights the potential of hybrid modeling as a practical solution for improving groundwater quantification accuracy and shows avenues for higher‐resolution estimates.

Groundwater depletion poses a challenge to water security, but the coarse spatial resolution of Gravity Recovery and Climate Experiment (GRACE) observations limits the characterization of groundwater variability at regional and local scales. Here, we present a 1 km monthly global groundwater storage anomaly (GWSA) dataset covering April 2002–December 2020. The dataset was generated using Singular Spectrum Analysis (SSA) for temporal gap filling, 19 high-resolution hydroclimatic and geophysical predictors, and aquifer-stratified machine-learning downscaling. Extreme Gradient Boosting (XGBoost) and deep neural networks (DNN) were evaluated as candidate models, and XGBoost was selected for dataset production based on internal, cross-scale, and in-situ evaluations. When aggregated to the original GRACE scale, the 1 km product agreed closely with the original GWSA ( R 2 = 0.972, RMSE = 2.10 cm). Independent evaluation used a global archive of 152 069 groundwater-monitoring wells and a 1° spatial-block calibration–validation design. Effective specific yield ( S yeff ) was estimated from calibration blocks and transferred to held-out blocks before comparison at a common 0.25° grid. For XGBoost, the held-out evaluation yielded R 2 = 0.538, RMSE = 0.378 cm yr −1 , and Bias = 0.106 cm yr −1 . The association between modeled and observation-derived trends remained similar under ± 30 % perturbations in S yeff , although amplitude-related metrics were more sensitive to the conversion coefficient. Standardized time-series comparisons and monthly evaluations of representative aquifers showed better agreement for long-term variability than for seasonal variability. Comparisons with published regional estimates showed close agreement in the North China Plain, weaker depletion magnitudes in northwestern India and California, and a trend-sign mismatch in the High Plains Aquifer during 2003–2013. Experiments using CSR, JPL, H08, and WaterGAP also showed that reconstruction performance and predictor importance depend on the target product. The dataset provides kilometre-scale estimates of groundwater-storage variability under the large-scale GRACE constraint. Local interpretation of the reconstructed spatial patterns requires additional groundwater-level, abstraction, and hydrogeological observations. The 1 km monthly global GWSA dataset is openly available from the National…

Predictive stream temperature models are commonly used to address spatial and temporal gaps in observational data because stream temperature information is essential for environmental management decisions. National models capture broad patterns driven by climate, topography, and seasonality, but often lack the spatial resolution to simulate fine‐scale variation in groundwater discharge, riparian shading, or tributary inflows. In contrast, watershed models may simulate finer‐resolution patterns but miss regional drivers or be limited by scarcity of observations. To generate watershed‐extent, fine‐resolution stream temperature predictions, we developed 18 multiscale machine‐learning models that leverage (a) a broad‐extent, coarse‐resolution stream temperature model and (b) finer‐resolution observations and predictor data. The coarse‐resolution model simulates daily stream temperature across the conterminous United States on the National Hydrologic Geospatial Fabric (13.7 km median reach length; 865,364 km total). The fine‐resolution models simulate daily stream temperature on the National Hydrography Dataset Plus version 2 (1.5 km median; 5,192,808 km total). We compared multiscale models to seven purely broad‐ (national) or watershed‐extent models, testing them in eight watersheds (436–985 km 2 ) with differing hydrologic settings and under realistic data availability scenarios. Especially when observations within a watershed were scarce, multiscale models generated more accurate predictions than national or watershed‐specific models. A recurrent graph convolution neural network pretrained on coarse predictions and fine‐tuned on watershed‐specific observations was most accurate across watersheds and data scenarios. These results highlight that leveraging both coarse, national results and watershed‐specific data can increase model accuracies, particularly when observations within a watershed are spatially or temporally limited.

Designing reservoir operation policies that can adapt to non‐stationary hydroclimatic conditions is a major challenge for water resources management under climate change. This study develops an adaptive dynamic operation policy (ADOP), in which Gaussian radial basis functions (RBFs) are coupled with a gated recurrent unit (GRU)‐based parameterization module that maps meteorological information to time‐varying RBF weights. Using the reservoir system in the Qingjiang River Basin, China, as a case study, ADOP is optimized using an expanded historical hydroclimatic ensemble within an evolutionary multi‐objective direct policy search (EMODPS) framework and is evaluated under both an independent historical test period and 12 GCM‐based climate scenarios. A static operation policy (SOP) and a scenario‐specific re‐optimized ideal operation policy (IOP) are used as benchmarks. Results show that ADOP outperforms SOP under both historical training and test conditions. Under future scenarios, ADOP achieves higher hypervolume than SOP in 10 of 12 scenarios, indicating the advantage of the time‐varying parameterization in out‐of‐sample conditions. Although both historically trained policies retain a performance gap relative to the re‐optimized IOP, ADOP partially mitigates this gap. The water‐level and release trajectories for representative years further demonstrate that the advantage of ADOP is associated with improved storage–release coordination. Moreover, a diagnostic comparison supports the value of the GRU‐based module, as adding meteorological inputs to a static RBF policy does not reproduce ADOP's gains. The proposed time‐varying policy provides a practical pathway for improving the adaptability of historically trained reservoir operation policies under uncertain future hydroclimatic conditions.

The internal recharge dynamics of karst aquifers are difficult to resolve with traditional outlet‐based indicators, owing to their heterogeneous structure and nonlinear response. To address this, we propose a non‐invasive anomaly detection framework based on a Temporal Convolutional AutoEncoder (TCAE), which integrates high‐frequency electrical conductivity (EC) and temperature ( T ) signals to characterize recharge mechanisms. The study was conducted in a karst basin in Guilin, China, using 15‐min EC, T , and rainfall data from three spring types: conduit‐dominated, fracture‐matrix mixed, and shallow‐fracture controlled. The model accurately reconstructed non‐event EC across sites ( R 2 = 0.980–0.998) and quantified rainfall‐triggered anomalies using peak (AE peak ) and cumulative (AE AUC ) anomaly energy. AE peak showed strong internal consistency with event‐scale EC change ( r = 0.897–0.923), while AE AUC was significantly correlated with cumulative outlet‐discharge response ( r = 0.590–0.857). Mixed‐effects analysis of Ca 2+ and HCO 3 − concentrations, together with environmental isotope evidence from δ 13 C DIC , 222 Rn, and δ 18 O, provided complementary support for the physical interpretation of AE‐based recharge‐type classification. Recharge events were classified into four types based on the maximum event‐scale changes in EC and T : cold dilution, warm dilution, cold concentration, and warm concentration. Kruskal–Wallis test, and K ‐means clustering revealed distinct patterns across spring types: rapid recharge in conduit systems, buffered responses in fracture‐matrix systems, and threshold activation in shallow fracture systems. This study presents the first application of TCAE in karst hydrology. The dual‐parameter EC–T framework offers a real‐time, high‐resolution, and low‐cost tool for characterizing dynamic recharge, supporting sustainable management and early warning in karst aquifers.

The influence of reservoirs on the water cycle introduces significant uncertainty for hydrologic prediction. The representation of reservoirs in hydrologic models ideally must be accurate, interpretable, and transferable across sites. Recent studies have highlighted the potential for data‐driven methods, including long short‐term memory (LSTM) networks, to accurately capture reservoir releases. However, the performance of LSTM models of reservoir releases has not yet been diagnosed on a large‐sample data set to understand whether their accuracy is physically justified. This study evaluates the ability of LSTMs to represent reservoir release policies across the continental U.S., leveraging the recently developed ResOpsUS data set. In particular, we focus on whether LSTMs can infer a latent storage representation without being provided observed storage as input. We find that in many cases the LSTM succeeds in encoding physically interpretable storage dynamics in its internal states, despite measurement errors in the data. However, the model accuracy is only weakly related to the strength of the learned storage representation; instead, it is log‐linearly related to the degree of regulation. In addition, LSTMs struggle to generalize storage dynamics and learned policies in space with pooled training. This study contributes to the growing literature on interpreting deep learning models of human‐hydrologic systems.

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.

Water within the root zone regulates water, energy, and carbon exchanges at the land‐atmosphere interface, yet the dynamically active portion of subsurface water that ecosystems mobilize for evapotranspiration (ET) remains unclear, because it represents only part of the total subsurface storage. To address this gap, we develop a catchment‐scale, ecosystem‐oriented hybrid machine learning model that integrates a parsimonious ecohydrological structure with data‐driven flexibility to infer time‐varying active root zone storage (a S rz ) as a latent state variable. Without prescribing rooting depth or soil layers, the model learns adaptive water partitioning from multi‐source observations of ET, runoff, snow, and terrestrial water storage, enabling diagnosis of both dynamic a S rz and its realized capacity. Applied across hundreds of U.S. catchments over the period 2001–2020, the framework reveals coherent spatial and hydroclimatic patterns, with a S rz peaking in transitional regions where atmospheric demand and water supply are balanced. The realized capacity reflects how belowground water access varies with long‐term hydroclimatic and substrate constraints. We further show that defining storage as a realized quantity through system dynamics improves stability and interpretability under varying observational and structural conditions. Overall, this study demonstrates how hybrid modeling can be used to diagnose ecosystem‐accessible water storage as a latent state, providing a physically consistent and observation‐informed pathway to quantify belowground processes that remain inaccessible to direct measurement and control key land‐atmosphere interactions.

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.

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.

One of the most important skills of hydrological models is to simulate timing and magnitude of flood events. Long Short-Term Memory (LSTM) networks are currently among the most successful models for streamflow and flood prediction over large regions. In snow-influenced catchments, which typically comprise a minority in large-scale studies, floods are generated by two distinctly different processes, snowmelt and rainfall. The applicability of hydrological models in such regions is therefore dependent on their ability to represent both types of floods. Nevertheless, flood evaluations of LSTM taking different flood-generating processes into account are currently lacking. This study fills this gap by evaluating the ability of LSTM to model flood peak characteristics separately for snowmelt and rainfall generated floods. The trained LSTM model successfully simulated streamflow time series across the 103 evaluated catchments, with average NSE of 0.84 and average KGE of 0.86 over the unseen evaluation period. LSTM exhibited better performance in the majority of the catchments in terms of flood peak timing and magnitude for both rainfall and snowmelt generated floods when compared to the operational hydrological model in the region (HBV) used as a benchmark. LSTM had a 27 pp higher percentage of correctly simulated peak days for rainfall generated floods as compared to snowmelt generated floods, similar to what was found for HBV (29 pp). LSTM's ability to simulate flood peak magnitudes was similar for the different flood types, with percent errors within 20 % for close to half of the events. False alarm rate and probability of detection results were similar for snowmelt and rainfall generated events, both slightly better than those of mixed events. LSTM outperformed HBV in 70 %–91 % of the catchments in terms of flood peak timing and magnitude for snowmelt and rainfall generated floods. The largest improvements in flood peak magnitudes were found for rainfall generated events, in particular for catchments where HBV exhibited high (>40 % ) absolute errors. Overall, our findings bring confidence that LSTM can improve hydrological services in regions subject to both snowmelt and rainfall generated floods.

Accurate representation of raindrop size distribution (DSD) is crucial for understanding precipitation microphysics and improving quantitative precipitation forecasting. Conventional methods in numerical models often prescribe the DSD shape parameter as a constant or derive it from empirical relationships based solely on disdrometer observations, limiting their ability to capture DSD variability driven by dynamic, thermodynamic, and microphysical processes. This study develops a machine learning (ML) framework integrating disdrometer observations with environmental atmospheric variables to retrieve DSD parameters. Trained on data from 11 stations in Shanghai during 2019–2021, three ML models outperform conventional methods, effectively reducing biases and capturing observed DSD variability. They remain robust in independent heavy‐rainfall and tropical cyclone validations during 2022–2025, providing reliable DSD and rain rate estimates. Feature analysis and ablation experiments further underscore the critical role of environmental factors in improving retrieval accuracy, offering a data‐driven pathway to improve DSD parameterizations in numerical models.

Inter‐basin groundwater flow (IGF) is a critical process that links material and energy exchange between basins. Based on the source and flow direction of groundwater, IGF can be separated into export IGF and import IGF, which respectively denote the lateral outflow and inflow of groundwater from and into a basin. However, traditional hydrological modeling based on net IGF obscures the bidirectional exchange processes between basins. Furthermore, export and import IGF are difficult to measure directly, and existing methods to estimate them (e.g., hydrochemical tracing and groundwater modeling) are often costly and data‐demanding. Therefore, this study innovatively proposes a method that integrates the Long Short‐Term Memory (LSTM) with the ranked Storage Age Selection (rSAS) approach to partition export and import IGF. This method does not rely on detailed hydrogeological characterization and regional groundwater model. The feasibility of the proposed method was evaluated through a case study in the Lower Hafren basin, demonstrating its ability to quantify chloride transport processes within a unified water‐age balance framework, allowing a constrained estimation of unobserved inter‐basin groundwater flows. The results indicate that export and import IGF exhibit significantly different statistical characteristics compared to the net IGF. At the monthly timescale, the magnitudes of export IGF and import IGF are over two orders of magnitude greater than that of net IGF. And net IGF conceals the true seasonal fluctuation patterns inherent to export and import IGF. Furthermore, the uncertainties in the IGF partitioning were quantified and controlled through Bayesian uncertainty analysis and the LM‐CMA‐ES algorithm.

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.

River discharge estimation (RDE) methods are traditionally prone to high costs, vulnerability to extreme events, and are constrained by real‐time data availability. Current remote sensing‐based RDE approaches mainly extract basic river features such as width and water surface elevation, but rarely exploit the full spectral information from satellite imagery. Most studies also lack proper uncertainty assessment, which is essential for reliable water management decisions. This study proposes a machine learning (ML) approach for RDE that overcomes traditional limitations due to geometry, shadows and vegetated banks. Data are analyzed from 12 United States Geological Survey (USGS) hydrometric stations, spanning four river size classes: stream, small, medium, and large rivers. For each station, mean spectral band values are extracted from multiple polygons along the river channel and paired with temporally coincident in situ discharge measurements. This integrated data set is used to train five ML models—Extreme Gradient Boosting (XGBoost), Multilayer Perceptron (MLP), and three Bayesian Neural Networks (BNNs)—establishing a robust estimation framework that quantifies uncertainty. Models are trained and tested on 2016–2023 data, then evaluated on 2024–2025 observations to assess temporal transferability. Test (evaluation) results show strong performance with Kling‐Gupta Efficiency (KGE) ranging from 0.60 (0.35) for streams to 0.96 (0.92) for medium and large rivers. The approach maintains accuracy under both normal and extreme flow conditions and is robust to shadowed areas and complex channel geometries. By integrating satellite observations with ML, the proposed method enables real‐time RDE essential for streamflow forecasting, water allocation, and reservoir operation.

Drought is a highly consequential natural disaster that may likely increase in both severity and extent across the conterminous United States (CONUS). The mechanisms affecting the propagation of drought from the atmosphere to streamflow are complex and interactive, making the prediction of streamflow drought difficult with current modeling approaches. Machine learning is an emerging tool in the field of hydrology that may be well-suited to prediction of streamflow drought across large and topographically diverse areas. Here, we train and analyze 3198 random forest models at U.S. Geological Survey streamgages to understand common meteorological drivers of streamflow drought and to define physiographic characteristics of basins sensitive to these drivers. We also develop a novel dynamic regionalization approach using donor gages to predict daily streamflow drought at pseudo-ungaged locations. Our results show that teleconnections, temperature, evaporative demand, and snow–water equivalent are important drivers of streamflow drought in the West, Southwest, and Northern Rocky Mountains (Northern Rockies) regions of the United States, and precipitation and soil moisture are primary drivers of streamflow drought in the Northeast, Southeast, and the Northwest regions. Prediction using dynamic regionalization shows comparable performance to at-site models. This method may be applicable to other hydrologic prediction problems requiring transferability to ungaged locations.

The decoupling coefficient (Ω), a dimensionless metric that quantifies the balance between stomatal conductance ( G s ) and aerodynamic conductance ( G a ) in regulating evapotranspiration (ET), is key to understanding vegetation‐atmosphere interactions, which are fundamental to ecohydrological cycles and ecosystem functioning. However, the spatiotemporal patterns of Ω across different plant functional types (PFTs) and climatic gradients remain poorly understood, particularly in heterogeneous land surfaces under climate change. Here, we quantified Ω at 52 AmeriFlux sites across four PFTs in the contiguous U.S. (CONUS) from 2001 to 2024 and developed a high‐resolution Ω modeling framework that integrates eXtreme Gradient Boosting (XGBoost) with satellite observations and reanalysis data to produce spatially explicit 8‐day Ω maps at 0.1° resolution. This framework reveals a significant increasing trend in continental‐mean Ω (1.0 × 10 −3 yr −1 , p < 0.05), indicating progressively weaker vegetation‐atmosphere coupling, whereby ET becomes relatively insensitive to G s and increasingly constrained by available energy. We also found that Ω in water‐limited regions show highly volatile and asymmetric responses to drought, shifting toward stronger coupling in dry periods and toward stronger decoupling in wet periods, whereas Ω in energy‐limited regions remain strongly buffered with a stable coupling state. Our study reveals the spatiotemporal variability of Ω in CONUS, shedding light on the pivotal role of vegetation in modulating water fluxes under shifting climatic regimes and drought stress.

Rapid oasisification is widely pursued to strengthen regional livelihoods and ecological security in arid endorheic basins, yet it proceeds under mounting water scarcity and ecological trade‐offs. Whether such greening actually enhances ecosystem resilience—and through which components—remains unclear. Here we couple Pressure–State–Response (PSR) accounting with explainable machine learning (XGBoost–TreeSHAP) to characterize how pressure, ecosystem state, response capacity, and composite resilience (ER) co‐vary across contrasting landscapes in the Tarim River Basin (2000–2020). Oasis cover expanded from 12.06% to 14.32% and basin‐wide ER rose modestly from 0.50 to 0.54. However, this gain coincided with improvement in the Response component across 86.9% of the basin, whereas Pressure increased across 75.6%—most acutely along the mainstem corridor (98.8%)—and only 65.8% of the basin showed net ecological improvement. Composite resilience thus appears response‐compensated rather than pressure‐relieved. TreeSHAP attribution revealed contrasting predictor structures: net primary productivity (NPP) was the leading contributor in oasis models (26%–43%), while evaporative‐demand and moisture‐related variables contributed more in non‐oasis models. PSR quadrant analysis further localized higher‐pressure, lower‐state conditions along the mainstem. These results indicate that rising composite resilience can mask persistent pressure in water‐limited inland river systems, and that coordinated monitoring of pressure, state, and response capacity is needed to support integrated water–land management under intensifying climatic and human forcing.

Access to high-resolution long-term earth system projections is essential for advancing research in hydrology, agriculture, disaster management, and adaptation planning. This Data Descriptor presents an Artificial Intelligence (AI)-refined ensemble of Earth system model (ESM) projections for the conterminous United States at 1/24° (~4 km) spatial resolution. The dataset was generated by downscaling ten Coupled Models Intercomparison Project phase 6 (CMIP6) Earth system models (ESMs) under two emissions scenarios (SSP245 and SSP585) for an 80-year period (1980–2059). Two AI-driven methods, Super-Resolution Convolutional Neural Networks (SRCNN) and Super-Resolution Generative Adversarial Networks (SRGAN), were applied to produce high-resolution daily precipitation and minimum/maximum temperature. This descriptor documents the input datasets, processing workflow, bias-correction steps, file structure, variables, spatial and temporal coverage, and technical validation of the released data. Validation analyses compare the generated products with Daymet observations and existing downscaled datasets to characterize spatial patterns, biases, temporal consistency, and differences among products. The dataset provides daily high-resolution historical and future earth system projections that can support regional, impact assessment, and related applications, and details the AI downscaling framework, training protocols, and evaluation strategies to ensure reproducibility and usability.

Soil moisture is an essential climate variable that controls the exchange of water and energy between land and atmosphere, directly impacting land surface feedbacks. Satellite‐derived soil moisture provides valuable large‐scale information, but long‐term applications are limited by sensor lifespans, retrieval noise, and systematic biases. Earth system reanalyses rely on data assimilation to merge these observations with model dynamics. The mentioned limitations of the observations can degrade assimilation performance by propagating artifacts into model states. Here, we present a European daily soil moisture reanalysis for 2003–2023 that explicitly addresses these limitations by pre‐processing satellite observations prior to assimilation. We assimilate a newly developed Deep Learning (DL)‐enhanced AMSR‐E/2 soil moisture data set that provides a seamless daily record from 2003 to 2023 by reducing retrieval artifacts while preserving spatiotemporal consistency. Residual systematic biases are mitigated through climatological bias correction, and observation uncertainty is characterized using spatially distributed uncertainty estimates derived from Triple Collocation Analysis. The processed observations are subsequently assimilated into the encore Community Land Model (eCLM). This study evaluates whether DL‐enhanced observations improve reanalysis skill when assimilated into a continental‐scale ensemble land surface DA system. Evaluation against independent in situ measurements shows that assimilation of the DL‐enhanced AMSR‐E/2 reduces PBIAS from 28.68% to 3.28% and RMSE from 0.100 to 0.082 relative to open‐loop simulation, and outperforms assimilation of the original AMSR‐E/2 (PBIAS 15.13%, RMSE 0.094 ). Representing observation uncertainty as a spatially distributed field rather than a uniform value further improved temporal correlation at 66% of validation stations in a dedicated 3‐year experiment (2016–2018). Improvements are consistent across most in situ networks, with the largest gains observed in semi‐arid and transitional climate zones. At the same time, seasonal benefits are most pronounced during spring, autumn, and summer, where model dry‐down and vegetation dynamics are better represented. These improvements result in better preservation of hydrological signals and more physically consistent soil moisture dynamics. Our results demonstrate that…

Understanding water fluxes and mixing dynamics in the unsaturated zone is crucial for ecohydrological studies. Yet, observations of soil moisture and soil water isotopes remain limited in spatial extent and temporal resolution, restricting modeling studies on the critical zone. A data‐driven Artificial Intelligence (AI) approach was applied in a first attempt to simulate daily soil moisture and soil water isotopes (δ 2 H, δ 18 O) across soil profiles of a mixed land use catchment using parsimonious climate and vegetation predictors. A sequential model combined a Long Short‐Term Memory network for soil moisture and a Random Forest model for soil water isotope simulation, using LSTM‐predicted soil moisture as input. Training used ca. 2 years of daily soil moisture and one year of monthly soil water isotopes from different depths (0–100 cm) under seven land uses in the Demnitzer Millcreek catchment (66 km 2 ), NE Germany. The AI model captured seasonal and depth‐dependent hydrological patterns, outperforming a process‐based model previously applied in the catchment. Soil moisture simulations showed Kling‐Gupta Efficiency of 0.75–0.92 (RMSE: 1.81%–4.50%), while isotopes simulations achieved 0.80–0.88 (RMSE: 4.3–5.8‰ for δ 2 H, 0.6–0.9‰ for δ 18 O). However, the performance of AI model depended on the study period and catchment, showing overestimation in drier test periods and underestimation in wetter validation periods. Also, in this evapotranspiration‐dominated site, the model relied heavily on indirect climate predictors (e.g., relative humidity), derived from direct drivers (e.g., precipitation), reducing performance in short‐term hydrological responses. Our sequential model has potential as stand‐alone simulations or as a surrogate for process‐based models.

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.

In the past decade, the Amazon has experienced multiple severe droughts, raising critical questions about how vegetation stress can be detected using remote sensing indicators such as Solar‐Induced Chlorophyll Fluorescence (SIF). In this study, we compare two anomaly detection approaches applied to TROPOSIF data. First a classical pixel wise z‐score method, and second a 3D Convolutional Autoencoder (CAE) that learns spatiotemporal SIF patterns. We evaluate both methods against two drought products, the global Standardised Precipitation‐Evapotranspiration Index (SPEI) metric and Brazil's Monitor de Secas (MSB). Our results show that the CAE identifies broader, more persistent anomalies than the z‐score method and aligns more closely with drought patterns detected by SPEI and MSB, particularly during the peak dry period in September 2024 and the months of delayed recovery that followed. These findings highlight how methodological choices shape the interpretation of vegetation stress and why anomaly detection methodologies should be carefully selected.

We present a 225-year (1799–2024) homogenized daily water level series for the Vistula River in Warsaw, comprising 82 453 observations. The construction of this consistent dataset required adjustments for changes in gauge location, shifts in gauge zero, differences in historical measurement units, and calendar discrepancies between the Julian and Gregorian systems. A small number of missing measurements was completed using the Long Short-Term Memory (LSTM) method, applied to measurements taken at the Warsaw gauge and parallel measurements from downstream water level gauge stations along the Vistula. The resulting dataset offers a robust foundation for long-term hydrological, climatic, and socio-environmental research. The dataset is openly available at Zenodo repository: https://doi.org/10.5281/zenodo.16919654 (Sobechowicz et al., 2025a).

Evaporative fraction (EF), defined as the ratio of latent heat flux to the sum of latent and sensible heat fluxes, is a key metric of surface energy partitioning and an indicator of plant water stress. To investigate the mechanisms underlying EF dynamics, we developed an explainable machine learning (ML) model based on a Long Short-Term Memory (LSTM) architecture that explicitly incorporates memory effects. The model was trained using data from 90 eddy-covariance sites across diverse plant functional types (PFTs) from the ICOS, AmeriFlux, and FLUXNET2015 Tier 1 datasets. Using only routinely available meteorological inputs (e.g., precipitation, incoming shortwave radiation, air temperature, and vapor pressure deficit) together with static site attributes (e.g., PFT and soil properties), the model accurately captured EF dynamics, particularly during post-rainfall pulses and soil moisture dry-down events. Across all test periods, ensemble mean predictions showed good agreement with observations, with a median site-level NSE of 0.63 across sites spanning broad climate and ecosystem gradients. Explainable ML analyses identified precipitation and vapor pressure deficit as the dominant drivers of EF in woody savanna, savanna, open shrubland, and grassland ecosystems, whereas air temperature emerged as the dominant factor in deciduous broadleaf, evergreen needleleaf, and mixed forests. Expected Gradients further revealed substantial variation in memory effect contributions across PFTs, with evergreen broadleaf forests and savannas showing stronger influences from antecedent conditions than grasslands. These memory effects were closely associated with rooting depth, soil water-holding capacity, and plant water-use strategies, which together regulate drought-response timescales. The learned memory patterns also suggest potential for inferring information related to rooting-zone water storage capacity and plant water stress from surface observations. Overall, our findings underscore the critical role of memory effects in EF prediction and highlight their relevance for anticipating vegetation water stress under increasing drought frequency and intensity.

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.

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.

Soil moisture dynamics in vertical profiles result from multiple interacting processes, generating complex temporal behavior and intricate connectivity patterns. However, accurately simulating multi‐depth soil moisture while diagnosing time‐varying vertical connectivity remains challenging. This study develops a Long Short‐Term Memory model with time‐varying spatial attention (LSTM + ATT) to predict multi‐depth soil moisture and quantify vertical hydrological connectivity via attention weights. Evaluated across 19 USCRN (U.S. Climate Reference Network) stations, LSTM + ATT demonstrated superior robustness compared to independent (Ind‐LSTM) and joint (Joint‐LSTM) baselines. While performance was nearly identical at shallow depths (5–20 cm), LSTM + ATT excelled in deeper layers, achieving median NSE values of 0.91 (50 cm) and 0.93 (100 cm), outperforming Ind‐LSTM (0.86–0.89) and Joint‐LSTM (0.90–0.91) while eliminating catastrophic failures (NSE < 0). Learned attention weights revealed physically consistent connectivity patterns, weakening as vertical distance increased. Weight dynamics were strongly linked to soil moisture: influence from upper layers correlated positively ( r : 0.20 to 0.50) with moisture content, while influence from lower layers (50 and 100 cm) correlated negatively ( r : −0.76 to −0.18). During preferential flow, weights reorganized to enhance connectivity with upper layers (increasing 30%–100%) at the expense of self‐influence (decreasing 10%–30%), highlighting distinct hydraulic signatures of preferential pathways. Unlike the implicit relationships in Joint‐LSTM, this dynamic attention framework provides a novel quantitative diagnostic for time‐varying connectivity, effectively bridging the gap between predictive accuracy and process interpretability in subsurface hydrology.

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.

Flood risk is correlated in space and time, challenging insurance systems that rely on diversification across assets. Financial instruments governing flood coverage are typically structured as 1–5‐year contracts, exposing portfolios to climate‐driven risk at interannual‐to‐decadal scales. Yet existing tools address climate risk either through seasonal forecasts extending only months or multidecadal projections misaligned with fiscal horizons, leaving a critical gap in actionable flood risk simulation. We introduce a multisite flood simulation framework combining attention‐based analog retrieval with stochastic generation of multivariate flood frequency, intensity, and duration sequences. Applied to over 100 sites in the Mississippi River Basin, the model produces spatiotemporally coherent flood portfolios conditioned on interannual climate variability. Explainable AI attribution paired with wavelet analysis links simulated clustering to large‐scale climate drivers, yielding physically interpretable flood clusters for portfolio‐scale loss simulation. The framework provides plausible, out‐of‐sample flood risk catalogs for interannual‐to‐decadal insurance risk assessment and financial planning.

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.

Evapotranspiration (ET) is a critical component of the water cycle, influencing climate, agriculture, and water resource management. However, most satellite-derived ET products are limited to daily or coarser temporal resolutions, despite the strong diurnal variability of ET processes. Existing satellite-based ET retrievals are largely restricted to daytime conditions, when nighttime ET is a small but often non-trivial flux. In this study, we introduce the Advanced Baseline Imager Live Imaging of Vegetated Ecosystems ET (ALIVE ET ), a near real-time, 5 min ET estimation framework, leveraging geostationary satellite observations from the GOES-R Advanced Baseline Imager (ABI) and machine learning models under both clear and cloudy conditions. We test Gradient Boosting Regression (GBR) and Long Short-Term Memory (LSTM) models to assess their ability to estimate ET variations across the diurnal cycle. GBR captures daytime ET with an R 2 of 0.74 (normalized RMSE of 0.91) while maintaining low computational cost. For nighttime ET, LSTM models trained on time-series observations perform better, achieving an R 2 of 0.24 (nRMSE of 1.29) by leveraging temporal dependencies in land surface temperature (LST) and past ABI observations. Comparisons against daily ET estimates from the physically-based ALEXI remote sensing model demonstrates good agreement but opportunities for improvement. This study demonstrates the potential of integrating machine learning with geostationary remote sensing to advance high-temporal-resolution ET estimation.

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.

India has one of the largest networks of large dams in the world, where reservoirs play a crucial role in regulating monsoon‐driven water availability for irrigation, hydropower, flood control, and water supply. However, long‐term daily reservoir storage observations are available only from the year 2000 onwards. In this study, we reconstruct long‐term daily live storage for major reservoirs across India by integrating climate variables, hydrological model simulations, and machine learning approaches. We used model‐simulated reservoir storage as a baseline predictor and applied Random Forest (RF) and XGBoost (XGB) with meteorological and hydrological variables, substantially improving accuracy compared to the raw simulations. The trained models over the observation period were then applied to the pre‐2000 period to reconstruct reservoir storage. Using the reconstructed data, we examined long‐term changes in reservoir storage variability and the timing of peak variability in reservoir storage during both the summer monsoon (June–September) filling and dry‐season release periods. We show that despite increasing reservoir capacity, normalized storage shows a moderate long‐term decline, reflecting an increase in water withdrawals. Pre‐monsoon storage exhibits larger and more spatially variable trends than post‐monsoon storage, while the frequency of low‐storage conditions has declined and high‐storage states have become more common. Peak storage variability typically occurs during late July‐mid August and around January, corresponding to peak monsoon inflows and irrigation demand. These findings provide insights into evolving reservoir storage dynamics and their implications for water resources management under changing hydroclimatic conditions in India.

Accurately understanding and predicting the evolution process of meteorological elements in the Luanhe River basin is of vital importance for ensuring water security in the Luanhe River basin. Based on the observation data from meteorological and hydrological stations, this study analyzed the spatiotemporal evolution characteristics of key meteorological elements such as temperature and precipitation in the Luanhe River basin. By using machine learning and CMIP6 data, combined with statistical downscaling and error-correction techniques, it predicted the trends of temperature and precipitation in the basin from 2025 to 2049. The results show that the annual average temperature, precipitation, and runoff in the Luanhe River basin have shown a fluctuating upward trend from 1980 to 2023. The average temperature rises by 0.33°C every 10 years, and the average precipitation increases by 23.73 mm every 10 years. In the future, with the increase of greenhouse gas emission scenarios, both the annual average temperature and the annual cumulative precipitation in the Luanhe River basin will show an upward trend. Under the Shared Socioeconomic Pathway 5-8.5 (SSP585) scenario, the predicted annual average temperature is 10.15°C and the annual cumulative precipitation is 712.06 mm. The complementary use of machine learning and CMIP6 provides mutually supportive evidence for interpreting future climate change tendencies. Significance Statement This study aims to assess future tendencies of temperature and precipitation in the Luanhe River basin, a critical water source for the Beijing–Tianjin–Hebei region. Our key finding is that both temperature and rainfall in the basin will continue to rise until 2049, with a temperature increase of up to 0.33°C and a precipitation increase of up to 23.73 mm decade −1 . These changes matter because they directly impact regional water security, flood risks, and ecological health. Our complementary approach, using both climate models and machine learning, provides additional evidence for interpreting climate impacts, which can help policymakers design better water management and climate adaptation strategies for this and similar watersheds.

Accurate estimates of snow water equivalent (SWE) are critical for hydrologic forecasting and water resource management in snow-dominated regions. While in situ observations, such as those from the U.S. Department of Agriculture Snowpack Telemetry (SNOTEL) network, provide direct measurements, their sparse spatial coverage limits their broader applicability. Numerical models such as Noah LSM with multiparameterization options (Noah-MP) provide spatially continuous SWE estimates but are computationally demanding. This study develops a deep learning–based emulator to reproduce SWE simulated by Noah-MP using a convolutional long short-term memory (ConvLSTM) network trained on meteorological forcings from the Weather Research and Forecasting Model and augmented with static physiographic variables and remotely sensed snow cover and leaf area index data. Incorporating spatial coupling substantially improves performance relative to a pixelwise long short-term memory baseline, increasing the fraction of grid cells with nonnegative Nash–Sutcliffe efficiency (NSE) by 15%–27% during the test period. The emulator reproduces large-scale SWE patterns consistent with both Noah-MP and Snow Data Assimilation System (SNODAS) and maintains competitive NSE and root-mean-square error performance relative to SNOTEL observations. Seasonal sensitivity experiments yield high R 2 values up to 0.96 during accumulation and 0.97 during melt. Once trained, the emulator generates SWE predictions faster than the parent physics-based simulation, enabling rapid, high-resolution estimates that support real-time monitoring and ensemble forecasting. Significance Statement Communities across the western United States rely on mountain snow as a significant source of freshwater, yet it remains difficult to track how much snow is stored at any given time. Ground measurements are sparse, and the numerical models used to estimate snowpack can take hours to run. We show that a data-driven approach can mimic the behavior of these complex models while producing results almost instantly. By capturing how weather and landscape features shape snowpack across the Colorado Rockies, the system achieves performance comparable to that of a widely used numerical model, with improved agreement with ground measurements at many locations. This capability provides a faster, more practical tool for water planning,…