Search

Physics-based models (PBMs), such as shallow water equations (SWEs) solvers, have been widely used in flood simulation and river hydraulics analysis. However, they are usually computationally expensive and unsuitable for parameter optimizations that need many runs. An alternative is the machine learning (ML) method, which can be used to construct computationally efficient surrogates for PBMs that can approximate their input-output dynamics. Among many ML techniques, convolutional neural network (CNN) is a prevalent method for image-to-image regressions on structured or regular meshes (e.g., mapping from the boundary conditions to flow solutions of SWEs). However, CNN-based methods have significant limitations because of their raster-image nature. Such methods cannot precisely capture the boundary geometry of obstacles and near-field flow features, which are of paramount importance to fluid dynamics. We introduced an efficient, accurate, and flexible neural network (NN) surrogate model [which is based on deep learning and can make point-to-point (p2p) predictions on unstructured meshes] called NN-p2p. The new method was evaluated and compared against CNN-based methods. NN-p2p improves the accuracy of the near-field flow prediction with a mean relative error of 0.56% for the velocity magnitude around piers with unseen length/width ratios. It also respects conservation laws more strictly than the CNN-based models and performs reasonably well for spatial extrapolation. The surrogate reduces computing time by almost 3-orders of magnitude in comparison with its corresponding PBM. Moreover, as a demonstration of the NN-p2p model's practical applicability, we calculated drag coefficient using NN-p2p for piers of varying length-to-width ratios and obtained a novel linear relationship between the drag coefficient and the logarithmic transformation of the pier's geometry. [ABSTRACT FROM AUTHOR]

Nowcasts, or near-real-time (NRT) forecasts, of soil moisture based on the Soil Moisture Active and Passive (SMAP) mission could provide substantial value for a range of applications including hazards monitoring and agricultural planning. To provide such a NRT forecast with high fidelity, we enhanced a time series deep learning architecture, long short-term memory (LSTM), with a novel data integration (DI) kernel to assimilate the most recent SMAP observations as soon as they become available. The kernel is adaptive in that it can accommodate irregular observational schedules. Testing over the CONUS, this NRT forecast product showcases predictions with unprecedented accuracy when evaluated against subsequent SMAP retrievals. It showed smaller error than NRT forecasts reported in the literature, especially at longer forecast latency. The comparative advantage was due to LSTM's structural improvements, as well as its ability to utilize more input variables and more training data. The DI-LSTM was compared to the original LSTM model that runs without data integration, referred to as the projection model here. We found that the DI procedure removed the autocorrelated effects of forcing errors and errors due to processes not represented in the inputs, for example, irrigation and floodplain/lake inundation, as well as mismatches due to unseen forcing conditions. The effects of this purely data-driven DI kernel are discussed for the first time in the geosciences. Furthermore, this work presents an upper-bound estimate for the random component of the SMAP retrieval error. [ABSTRACT FROM AUTHOR]

Accurate and frequent nitrate estimates can provide valuable information on the nitrate transport dynamics. The study aimed to develop a data-driven modeling framework to estimate daily nitrate concentrations at low-frequency nitrate monitoring sites using the daily nitrate concentration and stream discharge information of a neighboring high-frequency nitrate monitoring site. A Long Short-Term Memory (LSTM) based deep learning (DL) modeling framework was developed to predict daily nitrate concentrations. The DL modeling framework performance was compared with two well-established statistical models, including LOADEST and WRTDS-Kalman, in three selected basins in Iowa, USA: Des Moines, Iowa, and Cedar River. The developed DL model performed well with NSE >0.70 and KGE >0.70 for 67% and 79% nitrate monitoring sites, respectively. DL and WRTDS-Kalman models performed better than the LOADEST in nitrate concentration and load estimation for all low-frequency sites. The average NSE performance of the DL model in daily nitrate estimation is 20% higher than that of the WRTDS-Kalman model at 18 out of 24 sites (75%). The WRTDS-Kalman model showed unrealistic fluctuations in the estimated daily nitrate time series when the model received limited observed nitrate data (less than 50) for simulation. The DL model indicated superior performance in winter months' nitrate prediction (60% of cases) compared to WRTDS-Kalman models (33% of cases). The DL model also better represented the exceedance days from the USEPA maximum contamination level (MCL). Both the DL and WRTDS-Kalman models demonstrated similar performance in annual stream nitrate load estimation, and estimated values are close to actual nitrate loads. [Display omitted] • DL model predicts daily NO 3 at low-frequency sites using nearby high-frequency data • DL and WRTDS models' performance was better than LOADEST in nitrate prediction • DL model estimated contamination exceedance days better than LOADEST and WRTDS • NO 3 data availability (<50 data) limited the WRTDS model performance [ABSTRACT FROM AUTHOR]

Geoscientific models manage myriad and increasingly complex data structures as trans -disciplinary models are integrated. They often incur significant redundancy with cross-cutting tasks. Reflection, the ability of a program to inspect and modify its structure and behavior at runtime, is known as a powerful tool to improve code reusability, abstraction, and separation of concerns. Reflection is rarely adopted in high-performance Geoscientific models, especially with Fortran, where it was previously deemed implausible. Practical constraints of language and legacy often limit us to feather-weight, native-language solutions. We demonstrate the usefulness of a structural-reflection-emulating, dynamically-linked metaObjects, gd . We show real-world examples including data structure self-assembly, effortless input/output (IO) and upgrade to parallel I/O, recursive actions and batch operations. We share gd and a derived module that reproduces MATLAB-like structure in Fortran and C++. We suggest that both a gd representation and a Fortran-native representation are maintained to access the data, each for separate purposes. Embracing emulated reflection allows generically-written codes that are highly re-usable across projects. [ABSTRACT FROM AUTHOR]

The Soil Moisture Active Passive (SMAP) mission has delivered valuable sensing of surface soil moisture since 2015. However, it has a short time span and irregular revisit schedules. Utilizing a state-of-the-art time series deep learning neural network, Long Short-Term Memory (LSTM), we created a system that predicts SMAP level-3 moisture product with atmospheric forcings, model-simulated moisture, and static physiographic attributes as inputs. The system removes most of the bias with model simulations and improves predicted moisture climatology, achieving small test root-mean-square errors (<0.035) and high-correlation coefficients >0.87 for over 75% of Continental United States, including the forested southeast. As the first application of LSTM in hydrology, we show the proposed network avoids overfitting and is robust for both temporal and spatial extrapolation tests. LSTM generalizes well across regions with distinct climates and environmental settings. With high fidelity to SMAP, LSTM shows great potential for hindcasting, data assimilation, and weather forecasting. [ABSTRACT FROM AUTHOR]

The central Amazon forest is projected to experience larger interannual precipitation variability, with uncertain impacts on terrestrial hydrologic fluxes. How surface runoff, groundwater, and evapotranspiration (ET) change as a function of annual precipitation (AP) has large climate and biogeochemical implications. A process-based hydrological model is used to examine the sensitivity of hydrologic budgets and stream discharge Q s generation to AP in an upland Amazon catchment. The authors find that AP strongly controls infiltration, base flow, and surface runoff, but not ET. Hence, AP alone can predict interannual changes in these fluxes except ET. Experiments with perturbed rainfall show the strong control derives from the predominant groundwater component that varies linearly with AP but is insensitive to seasonal rainfall fluctuations. Most rainfall from large storms infiltrates and becomes base flow rather than runoff or ET. Annual baseflow index (BFI; the fraction of stream discharge from base flow) is nearly constant (~0.8) when AP is below ~2500 mm yr−1 and decreases with AP above this value, which represents an inflection point for increased storage-dependent saturation excess. These results indicate that the system is energy limited and groundwater dominated in dry seasons, which implies some resilience of ET to moderate droughts. The results suggest AP is a good predictor for interannual changes in infiltration. Both the seasonal near-surface soil moisture and surface runoff are correlated more strongly to the subsurface fluxes than to precipitation over monthly and annual time scales. Finally, the results confirm the importance of central Amazon groundwater flow and its buffering effect on storms and droughts, implying needed model development in regional to global models. [ABSTRACT FROM AUTHOR]

Soil moisture statistical fractal is an important tool for downscaling remotely-sensed observations and has the potential to play a key role in multi-scale hydrologic modeling. The fractal was first introduced two decades ago, but relatively little is known regarding how its scaling exponents evolve in time in response to climatic forcings. Previous studies have neglected the process of moisture re-distribution due to regional groundwater flow. In this study we used a physically-based surface-subsurface processes model and numerical experiments to elucidate the patterns and controls of fractal temporal evolution in two U.S. Midwest basins. Groundwater flow was found to introduce large-scale spatial structure, thereby reducing the scaling exponents ( τ ), which has implications for the transferability of calibrated parameters to predict τ . However, the groundwater effects depend on complex interactions with other physical controls such as soil texture and land use. The fractal scaling exponents, while in general showing a seasonal mode that correlates with mean moisture content, display hysteresis after storm events that can be divided into three phases, consistent with literature findings: (a) wetting, (b) re-organizing, and (c) dry-down. Modeling experiments clearly show that the hysteresis is attributed to soil texture, whose “patchiness” is the primary contributing factor. We generalized phenomenological rules for the impacts of rainfall, soil texture, groundwater flow, and land use on τ evolution. Grid resolution has a mild influence on the results and there is a strong correlation between predictions of τ from different resolutions. Overall, our results suggest that groundwater flow should be given more consideration in studies of the soil moisture statistical fractal, especially in regions with a shallow water table. [ABSTRACT FROM AUTHOR]

Slope stability prediction is an important task in geotechnical engineering which can be achieved through physics-based or data-driven approaches. Physics-based approaches rely on geotechnical knowledge from soil mechanics, such as limit equilibrium analysis and shear strength theories, to evaluate the stability condition of slopes, and they are often limited to slope-specific analysis. Data-driven approaches predict slope stability conditions based on learned relationships between influencing factors and slope stability conditions from past observations of slope failures (i.e., case histories); they rely on big data which are difficult to obtain. This study examines three easy-to-implement and effective methods to integrate geotechnical engineering domain knowledge into data-driven models for slope stability prediction: hybrid knowledge-data model, knowledge-based model initiation, and knowledge-guided loss function. These models were benchmarked against pure data-driven models and domain knowledge–based models, including a physics-based solution chart and a physics-based empirical model. A compilation of slope stability case histories from the literature was used as the benchmark database, and five-fold cross-validation was employed to evaluate model performance. The model validation results demonstrated that machine learning models outperformed domain knowledge–based models in terms of several evaluation metrics. The three proposed methods were found to outperform both domain knowledge–based models and pure data-driven models. Additionally, the hybrid knowledge-data models and knowledge-guided loss function were found to reduce discrepancies in the predicted slope stability conditions compared with reported factor-of-safety values, leading to a better alignment with the underlying physics related to slope stability. This study provides an initial assessment of the value of coupling domain knowledge and data-driven methods in geotechnical engineering applications using slope stability prediction as an example. [ABSTRACT FROM AUTHOR]

Trans-disciplinary hydrologic models oriented toward practical questions must be accompanied by accurate parameterization techniques. This paper describes the effects of different choices in the integration of various data sources on outcomes of the model Process-based Adaptive Watershed Simulator coupled with the Community Land Model (PAWS + CLM). Using our Hierarchical Stochastic Selection method, the represented land use percentages are much closer to the raw dataset, and lead to a 26% difference in carbon flux from that of the traditional dominant classes method. River bed elevations extracted using a novel algorithm agree well with the groundwater table and significantly increase baseflow contribution to streams relative to a coarse-DEM-based model. The inclusion of additional information in the soil pedotransfer functions drastically shifts ET, Net Primary Production and recharge. These results indicate that judicious treatment of input data has strong impacts on hydrologic and ecosystem fluxes. We emphasize the need to report details of data integration procedures. [ABSTRACT FROM AUTHOR]

Abstract: In this paper, we propose a finite difference AMR-WENO method for hyperbolic conservation laws. The proposed method combines the adaptive mesh refinement (AMR) framework with the high order finite difference weighted essentially non-oscillatory (WENO) method in space and the total variation diminishing (TVD) Runge–Kutta (RK) method in time (WENO-RK) by a high order coupling. Our goal is to realize mesh adaptivity in the AMR framework, while maintaining very high (higher than second) order accuracy of the WENO-RK method in the finite difference setting. The high order coupling of AMR and WENO-RK is accomplished by high order prolongation in both space (WENO interpolation) and time (Hermite interpolation) from coarse to fine grid solutions, and at ghost points. The resulting AMR-WENO method is accurate, robust and efficient, due to the mesh adaptivity and very high order spatial and temporal accuracy. We have experimented with both the third and the fifth order AMR-WENO schemes. We demonstrate the accuracy of the proposed scheme using smooth test problems, and their quality and efficiency using several 1D and 2D nonlinear hyperbolic problems with very challenging initial conditions. The AMR solutions are observed to perform as well as, and in some cases even better than, the corresponding uniform fine grid solutions. We conclude that there is significant improvement of the fifth order AMR-WENO over the third order one, not only in accuracy for smooth problems, but also in its ability in resolving complicated solution structures, due to the very low numerical diffusion of high order schemes. In our work, we found that it is difficult to design a robust AMR-WENO scheme that is both conservative and high order (higher than second order), due to the mass inconsistency of coarse and fine grid solutions at the initial stage in a finite difference scheme. Resolving these issues as well as conducting comprehensive evaluation of computational efficiency constitute our future work. [Copyright &y& Elsevier]

Abstract: Process-based watershed models are useful tools for understanding the impacts of natural and anthropogenic influences on water resources and for predicting water and solute fluxes exported from watersheds to receiving water bodies. The applicability of process-based hydrologic models has been previously limited to small catchments and short time frames. Computational demands, especially the solution to the three-dimensional subsurface flow domain, continue to pose significant constraints. This paper documents the mathematical development, numerical testing and the initial application of a new distributed hydrologic model PAWS (Process-based Adaptive Watershed Simulator). The model solves the governing equations for the major hydrologic processes efficiently so that large scale applications become relevant. PAWS evaluates the integrated hydrologic response of the surface–subsurface system using a novel non-iterative method that couples runoff and groundwater flow to vadose zone processes approximating the 3D Richards equation. The method is computationally efficient and produces physically consistent solutions. All flow components have been independently verified using analytical solutions and experimental data where applicable. The model is applied to a medium-sized watershed in Michigan (1169km2) achieving high performance metrics in terms of streamflow prediction at two gages during the calibration and verification periods. PAWS uses public databases as input and possesses full capability to interact with GIS datasets. Future papers will describe applications to other watersheds and the development and application of fate and transport modules. [ABSTRACT FROM AUTHOR]

Abstract: The longitudinal dispersion coefficient (D) is an important parameter needed to describe the transport of solutes in rivers and streams. The dispersion coefficient is generally estimated from tracer studies but the method can be expensive and time consuming, especially for large rivers. A number of empirical relations are available to estimate the dispersion coefficient; however, these relations are known to produce estimates within an order of magnitude of the tracer value. The focus of this paper is on using the shear-flow dispersion theory to directly estimate the dispersion coefficient from velocity measurements obtained using an Acoustic Doppler Current Profiler (ADCP). Using tracer and hydrodynamic data collected within the same river reaches, we examined conditions under which the ADCP and tracer methods produced similar results. Since dead zones / transient storage (TS) are known to influence the dispersion coefficient, we assessed the relative importance of dead zones in different stream reaches using two tracer-based approaches: (1) TS modeling which explicitly accounts for dead zones and (2) the advection–dispersion equation (ADE) which does not have separate terms for dead zones. Dispersion coefficients based on the ADE tend to be relatively high as they describe some of the effects of dead zones as well. Results based on the ADCP method were found to be in good agreement with the ADE estimates indicating that storage zones play an important role in the estimated dispersion coefficients, especially at high flows. For the river sites examined in this paper, the tracer estimates of dispersion were close to the median values of the ADCP estimates obtained from multiple datasets within a reach. The ADCP method appears to be an excellent alternative to the traditional tracer-based method if care is taken to avoid spurious data and multiple datasets are used to compute a distance-weighted average or other appropriate measure that represents reach-averaged conditions. [Copyright &y& Elsevier]

Abstract: The success of transient storage (TS) modeling for natural streams depends, in part, on the ability to describe the dispersion process accurately. Evidence based on stream tracer data shows that solute transport processes often do not follow the classical second-order dispersion model (e.g., early breakthrough and faster than Fickian travel times were observed). While models based on space-fractional dispersion are a promising alternative, different definitions of fractional derivatives exist in the literature. Unlike integer-order derivatives, fractional derivatives represent convolutions of concentration with long-range spatial correlation and numerical approximations can produce dense matrices. Therefore issues of both accuracy and computational efficiency need to be examined to successfully identify model parameters for natural streams. In this paper, we first compare the performance of several numerical approaches for solving the space-fractional dispersion equation. We examine three different numerical approaches to approximate the space-fractional derivatives including: (a) a fully-implicit scheme based on the shifted Grünwald–Letnikov (GL) approximation (b) a three-point implicit representation based on the GL formula and (c) a three-point implicit scheme based on mass conservation and the Caputo definition of the fractional derivative. We then use an operator-splitting technique to evaluate a TS model based on space-fractional dispersion (the FSTS model) and test the model against analytical solutions and stream tracer data. A sequence acceleration method (Richardson extrapolation) significantly improves the performance of all schemes examined. Results indicate that the fully-implicit GL method with Richardson extrapolation produces the most accurate solutions while the three-point implicit GL scheme has a stringent time-step restriction to produce acceptable solutions. The three-point implicit scheme based on the Caputo derivative produces accurate solutions in a fraction of the time taken by the fully-implicit GL method and represents the best trade-off between accuracy and computational efficiency for practical applications. The scheme is suitable for parameter estimation and is used to successfully describe tracer data in a natural stream. [Copyright &y& Elsevier]

This document is a closure to a study on applying knowledge-guided machine learning to slope stability prediction. The study discussed methods for integrating geotechnical domain knowledge into machine learning models to enhance slope stability predictions, particularly when dealing with small data sets. The closure addresses concerns raised by discussers, including errors in the data set, oversimplification of slope conditions, and the effectiveness of machine learning models trained on sparse data sets. The authors recommend the establishment of a high-quality benchmark database for geotechnical engineering problems to address these challenges. They also discuss the use of domain knowledge-based models to generate simulated data and serve as additional input features in machine learning models. The authors emphasize the importance of incorporating domain knowledge to improve model performance and navigate the challenges posed by limited data sets in geotechnical engineering. [Extracted from the article]

Accurate hydrologic modeling is vital to characterizing how the terrestrial water cycle responds to climate change. Pure deep learning (DL) models have been shown to outperform process-based ones while remaining difficult to interpret. More recently, differentiable physics-informed machine learning models with a physical backbone can systematically integrate physical equations and DL, predicting untrained variables and processes with high performance. However, it is unclear if such models are competitive for global-scale applications with a simple backbone. Therefore, we use – for the first time at this scale – differentiable hydrologic models (full name δ HBV-globe1.0-hydroDL, shortened to δ HBV here) to simulate the rainfall–runoff processes for 3753 basins around the world. Moreover, we compare the δ HBV models to a purely data-driven long short-term memory (LSTM) model to examine their strengths and limitations. Both LSTM and the δ HBV models provide competitive daily hydrologic simulation capabilities in global basins, with median Kling–Gupta efficiency values close to or higher than 0.7 (and 0.78 with LSTM for a subset of 1675 basins with long-term discharge records), significantly outperforming traditional models. Moreover, regionalized differentiable models demonstrated stronger spatial generalization ability (median KGE 0.64) than a traditional parameter regionalization approach (median KGE 0.46) and even LSTM for ungauged region tests across continents. Nevertheless, relative to LSTM, the differentiable model was hampered by structural deficiencies for cold or polar regions, highly arid regions, and basins with significant human impacts. This study also sets the benchmark for hydrologic estimates around the world and builds a foundation for improving global hydrologic simulations. [ABSTRACT FROM AUTHOR]

Predicting discharge in contiguously data‐scarce or ungauged regions is needed for quantifying the global hydrologic cycle. We show that prediction in ungauged regions (PUR) has major, underrecognized uncertainty and is drastically more difficult than previous problems where basins can be represented by neighboring or similar basins (known as prediction in ungauged basins). While deep neural networks demonstrated stellar performance for streamflow predictions, performance nonetheless declined for PUR, benchmarked here with a new stringent region‐based holdout test on a US data set with 671 basins. We tested approaches to reduce such errors, leveraging deep network's flexibility to integrate "soft" data, such as satellite‐based soil moisture product, or daily flow distributions which improved low flow simulations. A novel input‐selection ensemble improved average performance and greatly reduced catastrophic failures. Despite challenges, deep networks showed stronger performance metrics for PUR than traditional hydrologic models. They appear competitive for geoscientific modeling even in data‐scarce settings. Plain Language Summary: Large areas of land exist on all continents where we do not have access to daily streamflow rate measurements, but predictions are still needed to plan for a changing climate. In such a scenario, if we apply hydrologic models calibrated elsewhere, how reliable are they? Previously, hydrologists studied the problem of prediction in ungauged basins (PUB), where a basin needing predictions can be represented by neighboring or similar basins. There has not been ample recognition that prediction in large, contiguously data‐sparse regions, which we call prediction in ungauged regions (PUR), is a substantially more difficult problem than PUB. Using a new benchmark problem with 671 US basins, we highlight this danger using machine learning models: performance dropped significantly from PUB to PUR scenarios. We proposed multiple approaches to improve model robustness, including a novel input‐selection ensemble method which combines models with different input options, as well as assimilating more widely available data. Despite the challenges with PUR, the machine learning model appears to be the best available tool for PUB and PUR applications. Perhaps counterintuitively, deep networks seem to be highly competitive even in data‐scarce settings, and geoscientific domains should carefully consider the extrapolation performance of their models. Key Points: We highlight increased model error with predictions in contiguously data‐scarce (or ungauged) regions (PUR) and propose a streamflow benchmarkWe introduce methods to mitigate error, including an input‐selection ensemble and integrating satellite data or flow distributionsWith these methods, our machine learning PUR model presents metrics competitive with basin‐by‐basin calibrated hydrologic models in the conterminous United States [ABSTRACT FROM AUTHOR]

Machine learning methods can serve as highly efficient and more objective generators for data-consistent hypotheses that explain how hydrologic systems function. However, they alone cannot differentiate between causal and associative relationships. The hypotheses, however, can be further assessed by the physics and causal relationships encoded in process-based models in a data-centric Bayesian learning framework, as opposed to the typical model-centric one. As a demonstration case, we show how a classification tree algorithm generated competing hypotheses explaining the gradients in storage-streamflow correlation relationships found in basins in the eastern continental US. With the help of numerical experiments using a process-based model, we then rejected soil texture and terrain slope and validated large soil thickness as the factor that causes high correlations between streamflow and storage. Based on our work in soil moisture modeling with deep learning and a trans-disciplinary literature review, we then provide an outlook of how this framework could be migrated to deep-learning-based effort, opening up a new research avenue. [ABSTRACT FROM AUTHOR]

Recently, recurrent deep networks have been shown to be a powerful tool in harnessing newly available big data for accurate long-term soil moisture predictions. However, to be useful in forecasting applications, they need more than just accuracy - the deep networks must provide model uncertainty estimates as well. Uncertainty for statistical and machine learning models can be classified into two categories: aleatoric and epistemic. Aleatoric uncertainty is the inherent uncertainty due to measurement error and noise in the data - it can be thought of like the error obtainable by a correctly specified model with infinite training data. Epistemic uncertainty results from insufficient training data and/or model misspecification. In this study, we adapt and evaluate an efficient uncertainty estimation framework that was recently proposed in the deep learning community. This method simultaneously estimates both aleatoric and epistemic uncertainty. The aleatoric estimate is provided directly by the deep network by adding an additional output unit. The epistemic component is estimated by Monte Carlo Dropout (MCD), which has two interpretations: the variability observed in an ensemble of deep networks, and an uncertainty estimate derived from a Gaussian Process trained with variational inference. Neither interpretation is exact (both rely on approximations whose errors are not quantified); however, Monte Carlo dropout has shown promise in estimating uncertainty in computer vision tasks for non-recurrent deep networks. Here, we carefully evaluate the quality of the aleatoric and epistemic uncertainty measures through a series of controlled experiments for soil moisture prediction. Using recurrent deep learning models trained on regional soil moisture time series, we show that the estimated uncertainties are indeed predictive of model errors. The MCD estimate of epistemic uncertainty clearly rises in response to cases that are different from the training data. Thus it can be used to detect when a model is being applied to a dataset for which the training data were not representative. Meanwhile, the aleatoric uncertainty estimate is accurate for representative training data, but, as expected, can be flawed if training data are not representative (a situation that can be detected Monte Carlo Dropout). However, we found that the two uncertainty estimates are not independent, suggesting that the network's aleatoric uncertainty estimate is also correlated with true epistemic uncertainty, e.g., the network learns not to trust its own predictions for cases that differ from the majority of the training data. [ABSTRACT FROM AUTHOR]

• A novel approach combining an integrated surface/subsurface model and a groundwater parameter-estimation model, was able to better constrain the model. • The groundwater system may be nearing, if not exceeding, its sustainable groundwater production capacity and the speed of recovery is not indicative of sustainability. • Results from using conventionally-assumed uniform recharge distort calibrated K fields and impacts assessment The quantification of recharge and trans-valley underflow is needed in arid regions to estimate the impacts of new water withdrawals on the water table. However, for mountainous desert areas, such estimates are highly challenging, due to data scarcity, heterogeneous soils, and long residence times. Conventional assessment employs isolated groundwater models configured with simplified uniform estimates of recharge. Here, we employed a data-constrained surface-subsurface process model to provide an ensemble of spatially distributed recharge and underflow estimates using perturbed parameters. Then, the Model-Independent Parameter Estimation and Uncertainty Quantification (PEST) package was used to calibrate the aquifer hydraulic conductivity field in MODFLOW for this ensemble and reject implausible recharge values. This novel dual-model approach, broadly applicable to mountainous arid regions, was designed to maximally exploit available data sources. It can assimilate groundwater head observations, reject unrealistic parameters, and narrow the range of estimated drawdowns due to pumping. We applied this approach to the Chuckwalla basin in California, USA to determine natural recharge. Simulated recharge concentrates along alluvial fans at the mountain fronts and ephemeral washes where run-off water infiltrates. If an evenly distributed recharge was employed as in conventional studies, it would result in regional biases in estimated drawdown and larger uncertainty bounds. We also note that the speed of groundwater recovery does not guarantee sustainability: heavy pumping induces large hydraulic gradients that initially recover quickly when pumping is halted, but the system may not ultimately recover to pre-pumping levels. [ABSTRACT FROM AUTHOR]

The Soil Moisture Active Passive (SMAP) mission measures important soil moisture data globally. SMAP’s products might not always perform better than land surface models (LSM) when evaluated against in situ measurements. However, we hypothesize that SMAP presents added value for long-term soil moisture estimation in a data fusion setting as evaluated by in situ data. Here, with the help of a time series deep learning (DL) method, we created a seamlessly extended SMAP data set to test this hypothesis and, importantly, gauge whether such benefits extend to years beyond SMAP’s limited lifespan. We first show that the DL model, called long short-term memory (LSTM), can extrapolate SMAP for several years and the results are similar to the training period. We obtained prolongation results with low-performance degradation where SMAP itself matches well with in situ data. Interannual trends of root-zone soil moisture are surprisingly well captured by LSTM. In some cases, LSTM’s performance is limited by SMAP, whose main issue appears to be its shallow sensing depth. Despite this limitation, a simple average between LSTM and an LSM Noah frequently outperforms Noah alone. Moreover, Noah combined with LSTM is more skillful than when it is combined with another LSM. Over sparsely instrumented sites, the Noah–LSTM combination shows a stronger edge. Our results verified the value of LSTM-extended SMAP data. Moreover, DL is completely data driven and does not require structural assumptions. As such, it has its unique potential for long-term projections and may be applied synergistically with other model-data integration techniques. [ABSTRACT FROM AUTHOR]

Accurate prediction of snow water equivalent (SWE) can be valuable for water resource managers. Recently, deep learning methods such as long short-term memory (LSTM) have exhibited high accuracy in simulating hydrologic variables and can integrate lagged observations to improve prediction, but their benefits were not clear for SWE simulations. Here we tested an LSTM network with data integration (DI) for SWE in the western United States to integrate 30-day-lagged or 7-day-lagged observations of either SWE or satellite-observed snow cover fraction (SCF) to improve future predictions. SCF proved beneficial only for shallow-snow sites during snowmelt, while lagged SWE integration significantly improved prediction accuracy for both shallow- and deep-snow sites. The median Nash–Sutcliffe model efficiency coefficient (NSE) in temporal testing improved from 0.92 to 0.97 with 30-day-lagged SWE integration, and root-mean-square error (RMSE) and the difference between estimated and observed peak SWE values dmax were reduced by 41% and 57%, respectively. DI effectively mitigated accumulated model and forcing errors that would otherwise be persistent. Moreover, by applying DI to different observations (30-day-lagged, 7-day-lagged), we revealed the spatial distribution of errors with different persistent lengths. For example, integrating 30-day-lagged SWE was ineffective for ephemeral snow sites in the southwestern United States, but significantly reduced monthly-scale biases for regions with stable seasonal snowpack such as high-elevation sites in California. These biases are likely attributable to large interannual variability in snowfall or site-specific snow redistribution patterns that can accumulate to impactful levels over time for nonephemeral sites. These results set up benchmark levels and provide guidance for future model improvement strategies. [ABSTRACT FROM AUTHOR]

Climate change threatens our ability to grow food for an ever-increasing population. There is a need for high-quality soil moisture predictions in under-monitored regions like Africa. However, it is unclear if soil moisture processes are globally similar enough to allow our models trained on available in situ data to maintain accuracy in unmonitored regions. We present a multitask long short-term memory (LSTM) model that learns simultaneously from global satellite-based data and in situ soil moisture data. This model is evaluated in both random spatial holdout mode and continental holdout mode (trained on some continents, tested on a different one). The model compared favorably to current land surface models, satellite products, and a candidate machine learning model, reaching a global median correlation of 0.792 for the random spatial holdout test. It behaved surprisingly well in Africa and Australia, showing high correlation even when we excluded their sites from the training set, but it performed relatively poorly in Alaska where rapid changes are occurring. In all but one continent (Asia), the multitask model in the worst-case scenario test performed better than the soil moisture active passive (SMAP) 9 km product. Factorial analysis has shown that the LSTM model's accuracy varies with terrain aspect, resulting in lower performance for dry and south-facing slopes or wet and north-facing slopes. This knowledge helps us apply the model while understanding its limitations. This model is being integrated into an operational agricultural assistance application which currently provides information to 13 million African farmers. [ABSTRACT FROM AUTHOR]

• Machine learning models were trained for predicting mean NO3 in 2061 US rivers. • Five watershed indexes can explain 69% of spatial variability in NO3 concentrations. • Nitrogen input rates and urban areas are top drivers, followed by climate (MAP, MAT). • The models can predict mean NO3 concentrations in chemically-ungauged rivers. Nitrate is one of the most widespread and persistent pollutants in our time. Our understanding of nitrate dynamics has advanced substantially in the past decades, although its predominant drivers across gradients of climate, land use, and geology have remained elusive. Here we collated nitrate data from 2061 rivers along with 32 watershed characteristic indexes and developed machine learning models to reconstruct long-term mean (multi-year average) nitrate concentrations in the contiguous United States (CONUS). The trained models show similarly satisfactory model performance and can predict nitrate concentrations in chemically-ungauged places with about 70% accuracy. Further analysis revealed that five (out of 32) indexes (drivers) can explain about 70% of spatial variations in mean nitrate concentrations. The five influential drivers are nitrogen application rates N rate and urban area A urban % (human drivers), mean annual precipitation and temperature (climate drivers), and sand percent Sand% (soil property driver). Nitrate concentrations in undeveloped sites are primarily modulated by climate and soil property; they decrease with increasing mean discharge and Sand%. Nitrate concentrations in agriculture and urban sites increase with N rate and A urban % until reaching their apparent maxima around 10,000 kg/km2/yr and around 25%, respectively. Results indicate that nitrate concentrations may remain similar or increase with growing human population. In addition, nitrate concentrations can increase even without human input, as warming escalates water demand and reduces mean discharge in many places. These results allude to a conceptual model that highlights the impacts of distinct drivers: while human drivers predominate nitrogen input to land and rivers, climate drivers and soil properties modulate its transport and transformation, the balance of which determine long-term mean concentrations. Such mechanism-based insights and forecasting capabilities are essential for water management as we expect changing climate and growing agriculture and urbanization. [Display omitted] [ABSTRACT FROM AUTHOR]

• LSTM predicts daily SSC well using basin-scale atmospheric forcings, attributes, and observed or modeled streamflow. • Whole-CONUS LSTM can provide reliable predictions for ungauged stations without SSC observations. • Local and Whole-CONUS models are more successful in the humid Eastern US (with high SSC-streamflow correlation, low SSC) • We hypothesize low SSC-streamflow correlation is due to seasonality, basin size, land cover, and rainfall heterogeneity. Suspended sediment concentration (SSC) is a crucial indicator for aquatic ecosystems and reservoir management but is challenging to predict at large scales. It is unclear whether SSC is predictable using macroscopic environmental attributes and forcings. This study tested the feasibility of deep-network-based models to predict daily SSC at basin outlets given only basin-averaged forcings, readily-available physiographic attributes, and streamflow (from observation or model). We trained long short-term memory (LSTM) deep networks both separately for each of the 377 sites across the conterminous United States (CONUS) (termed "local models"), and on all the sites collectively (Whole-CONUS). The Whole-CONUS and local models presented median coefficient of determination (R2) values of 0.63 and 0.52, respectively. This comparison agrees with previously acknowledged "data synergy" effects for LSTM models, where more data from more sites can help improve predictions overall. Furthermore, the continental-scale analysis provided a wealth of insights about SSC patterns. Both local and Whole-CONUS models tended to be more successful where SSC-streamflow correlations (R s - q ) were high − typically in the humid Eastern US − and with lower SSC. Low R s - q basins were often found in the arid Southwest with higher SSC. The highly-nonlinear SSC-streamflow relationships seem related to seasonality, basin size, and heterogeneity of land cover and rainfall within the basins, suggesting these basins need to be simulated at higher spatial resolution and may require additional inputs related to SSC induced by seasonality. The Whole-CONUS model also performed well for spatial extrapolation (basins not included in the training dataset, median R2 = 0.55), supporting large-scale modeling efforts. These state-of-the-art results using only minimal inputs suggest data-driven approaches can exploit the natural coevolution of sediment processes and the environment to support sediment modeling. [ABSTRACT FROM AUTHOR]

Deep learning (DL) models trained on hydrologic observations can perform extraordinarily well, but they can inherit deficiencies of the training data, such as limited coverage of in situ data or low resolution/accuracy of satellite data. Here we propose a novel multiscale DL scheme learning simultaneously from satellite and in situ data to predict 9 km daily soil moisture (5 cm depth). Based on spatial cross‐validation over sites in the conterminous United States, the multiscale scheme obtained a median correlation of 0.901 and root‐mean‐square error of 0.034 m3/m3. It outperformed the Soil Moisture Active Passive satellite mission's 9 km product, DL models trained on in situ data alone, and land surface models. Our 9 km product showed better accuracy than previous 1 km satellite downscaling products, highlighting limited impacts of improving resolution. Not only is our product useful for planning against floods, droughts, and pests, our scheme is generically applicable to geoscientific domains with data on multiple scales, breaking the confines of individual data sets. Plain Language Summary: High‐resolution soil moisture data can be of great value for many practical applications, for example, flood forecasting, drought monitoring, crop management, weather forecasting, and better understanding the world's ecosystems. Recently, deep learning (DL) models directly learning patterns from observations have shown strong performance in modeling soil moisture and other environmental factors. Used directly, however, they inherit certain limitations of their training data such as low resolution and low accuracy; in other words, these models are students that cannot exceed their teacher (the data source). For soil moisture, just as for many other environmental variables of interest, observations are available on multiple scales, for example, probes in the ground and satellite‐based data. Here we show that learning from multiple data sources on different scales at the same time can allow DL models to overcome limitations of every single source: the student model can then outperform each teacher. Our multiscale model reported the best metrics for daily soil moisture prediction compared to alternatives and is no longer limited by poor satellite sensing capability in certain regions. This multiscale scheme is broadly applicable to many areas of environmental study. Key Points: We propose a novel multiscale soil moisture model (5 cm depth) learning from both satellite and in situ data, resulting in high accuracyIn situ data provides more information than satellite data, but if used together, they can overcome the limitations of each data sourceThe impacts of input resolution waned below the 9 km grid, and the results established the upper error bound (RMSE ∼0.034) due to resolution [ABSTRACT FROM AUTHOR]

Downward longwave radiation (DLR) at the surface is a key variable of interest in fields, such as hydrology and climate research. However, existing DLR estimation methods and DLR products are still problematic in terms of both accuracy and spatiotemporal resolution. In this article, we propose a deep convolutional neural network (DCNN)-based method to estimate hourly DLR at 5-km spatial resolution from top of atmosphere (TOA) brightness temperature (BT) of the Himawari-8/Advanced Himawari Imager (AHI) thermal channels, combined with near-surface air temperature and dew point temperature of ERA5 and elevation data. Validation results show that the DCNN-based method outperforms popular random forest and multilayer perceptron-based methods and that our proposed scheme integrating multisource data outperforms that only using remote sensing TOA observations or surface meteorological data. Compared with state-of-the-art CERES-SYN and ERA5-land DLR products, the estimated DLR by our proposed DCNN-based method with physical multisource inputs has higher spatiotemporal resolution and accuracy, with correlation coefficient (CC) of 0.95, root-mean-square error (RMSE) of 17.2 W/m2, and mean bias error (MBE) of −0.8 W/m2 in the testing period on the Tibetan Plateau. [ABSTRACT FROM AUTHOR]

• First time Transformer achieved the same performance as LSTM on CAMELS dataset. • Unmodified vanilla Transformer could not reach LSTM's performance. • A non-recurrent connection was added to the Transformer, supporting parallelism. • Transformers may have scale advantages for larger datasets. • LSTMs and Transformers are likely nearing the prediction limits of the dataset. For a number of years since their introduction to hydrology, recurrent neural networks like long short-term memory (LSTM) networks have proven remarkably difficult to surpass in terms of daily hydrograph metrics on community-shared benchmarks. Outside of hydrology, Transformers have now become the model of choice for sequential prediction tasks, making it a curious architecture to investigate for application to hydrology. Here, we first show that a vanilla (basic) Transformer architecture is not competitive against LSTM on the widely benchmarked CAMELS streamflow dataset, and lagged especially prominently for the high-flow metrics, perhaps due to the lack of memory mechanisms. However, a recurrence-free variant of the Transformer model obtained mixed comparisons with LSTM, producing very slightly higher Kling-Gupta efficiency coefficients (KGE), along with other metrics. The lack of advantages for the vanilla Transformer network is linked to the nature of hydrologic processes. Additionally, similar to LSTM, the Transformer can also merge multiple meteorological forcing datasets to improve model performance. Therefore, the modified Transformer represents a rare competitive architecture to LSTM in rigorous benchmarks. Valuable lessons were learned: (1) the basic Transformer architecture is not suitable for hydrologic modeling; (2) the recurrence-free modification is beneficial, so future work should continue to test such modifications; and (3) the performance of state-of-the-art models may be close to the prediction limits of the dataset. As a non-recurrent model, the Transformer may bear scale advantages for learning from bigger datasets and storing knowledge. This work lays the groundwork for future explorations into pretraining models, serving as a foundational benchmark that underscores the potential benefits in hydrology. [ABSTRACT FROM AUTHOR]

• Formulate two application-orientated objective functions for streamflow forecasting. • Develop an objective function-based ensemble model framework. • The proposed ensemble model achieves a better overall forecasting performance. • Highlight the importance of objective function for model establishment. The objective function plays an important role in hydrological model calibrations/training, since it largely determines the values of the model parameters and consequently influences the model performance. In this study, we establish two application-orientated objective functions, namely high flow balance error (HFBE) and mean squared percentage error (MSPE), for the forecasts of high flows and low flows, respectively. We examine the strengths and weaknesses of these streamflow forecast models trained with HFBE, MSPE and mean square error (MSE). Furthermore, we develop an objective function-based ensemble model (OEM) framework that can integrate the models trained with different objective functions. Our results in 273 catchments over USA show that the models trained on MSE have obvious underestimation in high-flow prediction. The models trained on HFBE can alleviate this underestimation and thus perform remarkably better for high-flow forecast. In addition, the models trained on MSPE outperform the other two models in low-flow forecast, but with an expense of the deterioration in the forecasting performance for high-flow. By incorporating the models trained on HFBE, MSPE and MSE, our proposed OEM performs well under all streamflow levels, with a median KGE of 0.96 and a median logNSE of 0.95. OEM realistically captures the mean and the variability of the observational streamflow under different scenarios with a variety of hydrometeorological conditions. This study highlights the necessity of applying objective functions that are appropriate for the modeling goal and the potential of ensemble learning methods for multi-objective optimization in hydrological modeling. [ABSTRACT FROM AUTHOR]

The 2013 federal Colorado River Basin Water Supply and Demand Study projected the water imbalance between future supply and demand to increase. The Colorado water supply (WS) exemplifies a pronounced quasi-decadal oscillation (QDO) of 10-20 years throughout its historical record; however, this QDO feature is unaccounted for in the climate models used to project the future WS. Adjacent to the Colorado River, the large watershed of the Great Salt Lake (GSL) in Utah records the hydrologicQDOsignal in its water surface, leading previous studies to explore the cause of decadal fluctuations in the lake elevation and assess predictability. This study reports a remarkable coherence between the Colorado WS and the GSL elevation at the 10-20-yr time scale. Analysis of precipitation and terrestrial water storage anomalies suggests a crossbasin connection in the climate and hydrometeorological variations of the Colorado WS and the GSL. The 160-yr-long and well-kept GSL elevation record makes it an effective indicator for the Colorado WS. [ABSTRACT FROM AUTHOR]

• The effectiveness of DL and TL in predicting alpine river discharge was tested. • DL and TL models were pretrained using data from 671 catchments across the U.S. • Conformance to rainfall-runoff principles determines the training observations. • Climate forcing data influence streamflow prediction more than hydrogeological attributes. • DL and TL models were ineffective in areas with abundant groundwater contribution. The Tibetan Plateau (TP) is an important Asian water tower for livelihood, irrigation, hydropower, and downstream ecosystems. Hydrological predictions in the TP have long been limited by sparse and discontinuous observations. Transfer learning (TL) technology may improve hydrological predictions by pre-training deep learning (DL) models on data-rich areas and then applying them to data-limited areas. However, the extent to which the DL and TL models work in alpine regions and the physical knowledge they provide remain unclear. Models were pretrained using data from 671 catchments across the U.S. and fine-tuned using data from four basins around the TP. Our results show that streamflow data with different temporal resolutions (monthly or daily) have little effect on discharge predictions when using DL. The number of discharge observations needed to enable acceptable performance of the DL models (Nash-Sutcliffe Efficiency coefficient greater than 0.6) depends on the hydrological characteristics of the catchments, in particular, how closely they conform to a general rainfall-runoff system. Climate forcing data determine the performance of streamflow prediction, whereas other attributes (i.e., soil and geology) have less significant impacts on prediction. The effectiveness of the DL and TL models is limited because groundwater contributes significantly to river discharge. This study provides an updated understanding of the application of DL and TL to quantify hydrological changes in the global cryosphere environment. [ABSTRACT FROM AUTHOR]

• Transfer learning can realize the benefits of integrating recent discharge during streamflow prediction. • Addition of a physically based model output as an input can improve the streamflow prediction. • Large hydrological datasets like CAMELS contain more information than a hydrological model with a physical foundation. • Testing the source model trained on large datasets can provide satisfactory streamflow predictions for moderately monitored regions. Recently, it has been demonstrated that the use of lagged discharge in long short-term memory (LSTM) networks represents an effective method for streamflow prediction, so-called, data integration (DI). However, it is uncertain if a transfer learning (TL) model, which did not include recent discharge when trained on the source region, can reap the benefits of including recent discharge data. This study investigated the ability of TL to provide daily streamflow predictions for a few watersheds in a snow-dominated region (target) while transferring the knowledge acquired from the conterminous United States (source) based on an LSTM architecture. The performance of the TL model was compared with that of a physically based model (PBM), HYDROTEL. Additionally, testing the source model on the target region demonstrated the performance of the model used in TL. The approaches applied to improve the accuracy of TL included use of: (i) DI of recent observed flows and simulated flows (HYDROTEL) to improve predictions, (ii) different meteorological and physiographic variables from the source and target datasets, (iii) incremental numbers of training watersheds from the target region. Testing the source model with DI at the target region produced test-period median Kling-Gupta-efficiency (KGE) and Nash-Sutcliffe-log-model-efficiency (Nash-log) values of 0.837 and 0.870, respectively. The best performance was achieved with the TL model using DI with a maximum number of watersheds (median values of 0.953 and 0.942 for Nash-log and KGE, respectively). Moreover, including PBM simulated flows improved predictions, reducing the variability of performance metrics. Our findings show that TL and additional new procedures could significantly enhance streamflow predictions even when DI models are available. [ABSTRACT FROM AUTHOR]

Summary Process-based hydrological models have a long history dating back to the 1960s. Criticized by some as over-parameterized, overly complex, and difficult to use, a more nuanced view is that these tools are necessary in many situations and, in a certain class of problems, they are the most appropriate type of hydrological model. This is especially the case in situations where knowledge of flow paths or distributed state variables and/or preservation of physical constraints is important. Examples of this include: spatiotemporal variability of soil moisture, groundwater flow and runoff generation, sediment and contaminant transport, or when feedbacks among various Earth’s system processes or understanding the impacts of climate non-stationarity are of primary concern. These are situations where process-based models excel and other models are unverifiable. This article presents this pragmatic view in the context of existing literature to justify the approach where applicable and necessary. We review how improvements in data availability, computational resources and algorithms have made detailed hydrological simulations a reality. Avenues for the future of process-based hydrological models are presented suggesting their use as virtual laboratories, for design purposes, and with a powerful treatment of uncertainty. [ABSTRACT FROM AUTHOR]

Glacier hydrology has profound implications for socio-economic development and nature conservation in arid Central Asia. Process-based hydrological models, which are the traditional tools used to simulate glacier melting, have made considerable contributions to advance our understanding of glacio-hydrology. Simultaneously, deep learning (DL) models have achieved excellent performance in many complex tasks and provide high accuracy. However, it is uncertain whether glacio-hydrological studies can benefit from the application of DL models. In this study, to help us assess water resource change for glacier-influenced regions, we used DL models to simulate glacio-hydrological processes in the Urumqi Glacier No. 1 in northwest China. First, we proposed a newly DL model called Exogenous Regularization Network (ERNet), which focuses on the relationship between exogenous (temperature and precipitation) and endogenous (runoff) variables, balancing the roles of different variables in the simulation process. Second, we compared ERNet with a stacked long short-term memory (LSTM) model and a process-based glacio-hydrology model, FLEX G. Experiments showed that compared with the other two models, ERNet not only performed well in runoff and peak flow simulations but also displayed superior transferability. Third, given that the DL model is data-driven, we experimentally compared the importance of air temperature and precipitation to glacial runoff processes. The results show that air temperature plays a dominant role in glacier runoff generation. We believe that the proposed model provides a useful predictive tool and that the results shed light on the future implication in cold region hydrology. • We proposed a new DL model called the Exogenous Regularization Network (ERNet). • ERNet performed well in runoff and peak flow simulation, and transferability test. • We analysed the importance of meteorological inputs for runoff simulation. • We provided some practical advices for the usage of DL models in glacio-hydrology. [ABSTRACT FROM AUTHOR]

• Synthetic samples are added to LSTM by previously undiscussed physical mechanisms. • Using extreme events to improve flood peaks and avoid negative streamflow. • Proposed PHY-LSTM outperforms conventional one both in local and regional models. • Physics-based monotonic relationships are upheld in the PHY-LSTM. Deep learning methods have recently shown a broad application prospect in rainfall-runoff modeling. However, the lack of physical mechanism becomes a major limitation in using machine learning methods that rely on the available labeled observations. To address this issue, the study proposes that synthetic samples are added to train the deep learning network by using three previously undiscussed physical mechanisms as follows: (1) extreme heavy rainfalls when the soil water is saturated, (2) long-duration rainless events when soil water is exhausted, and (3) the monotonic relationship between rainfall and runoff. A physics-guided Long Short-Term Memory (LSTM) network, shortly named PHY-LSTM, is then formulated. PHY-LSTM network is trained on 531 basins of the Catchment Attributes and Meteorology for Large-sample Studies (CAMELS) dataset, indicating that the performance is significantly improved compared to conventional LSTM. Specifically, the mean Nash-Sutcliffe Efficiency (NSE) improves from 0.52 to 0.61 from the daily simulations during the testing period in local models. It is demonstrated that synthetic samples can effectively improve the simulation of flood peaks and reduce the number of negative streamflow, and strong monotonicity is still maintained even if a slight disturbance is involved in the training dataset. The proposed PHY-LSTM shows that physical mechanisms are very useful to improve efficiencies of the data-driven rainfall-runoff model. [ABSTRACT FROM AUTHOR]

• LSTM achieved state-of-the-art performance for modeling basins with reservoirs. • Reservoir types, capacity-to-runoff ratio (dor) and diversion control streamflow. • LSTM performed well for basins with reservoirs that store about a month of flow. • It is crucial to include basins with reservoirs in the training set. • Large- dor basins with certain types of dams are more difficult for LSTM. A large fraction of major waterways have dams influencing streamflow, which must be accounted for in large-scale hydrologic modeling. However, daily streamflow prediction for basins with dams is challenging for various modeling approaches, especially at large scales. Here we examined which types of dammed basins could be well represented by long short-term memory (LSTM) models using readily-available information, and delineated the remaining challenges. We analyzed data from 3557 basins (83% dammed) over the contiguous United States and noted strong impacts of reservoir purposes, degree of regulation (dor), and diversion on streamflow modeling. While a model trained on a widely-used reference-basin dataset performed poorly for non-reference basins, the model trained on the whole dataset presented a median Nash-Sutcliffe efficiency coefficient (NSE) of 0.74. The zero- dor , small- dor (with storage of approximately a month of average streamflow or less), and large- dor basins were found to have distinct behaviors, so migrating models between categories yielded catastrophic results, which means we must not treat small- dor basins as reference ones. However, training with pooled data from different sets yielded optimal median NSEs of 0.72, 0.79, and 0.64 for these respective groups, noticeably stronger than existing models. These results support a coherent modeling strategy where smaller dams (storing about a month of average streamflow or less) are modeled implicitly as part of basin rainfall-runoff processes; then, large- dor reservoirs of certain types can be represented explicitly. However, dammed basins must be present in the training dataset. Future work should examine separate modeling of large reservoirs for fire protection and irrigation, hydroelectric power generation, and flood control. [ABSTRACT FROM AUTHOR]