Median Nash-Sutcliffe Efficiency Research Articles

Physically-based modelling of UK river flows under climate change

This study presents the model setup and results from the first calibrated, physically-based, spatially-distributed hydrological modelling of combined land cover and climate change impacts on a large sample of UK river catchments. The SHETRAN hydrological model was automatically calibrated for 698 UK catchments then driven by the 12 regional climate model projections from UKCP18, combined with urban development and natural flood management scenarios. The automatic calibration of SHETRAN produces a median Nash-Sutcliffe efficiency value of 0.82 with 581 catchments having a value greater than 0.7. 24 summary metrics were calculated to capture changes to important aspects of the flow regime. The UKCP18 realisations in SHETRAN indicate that a warming climate will cause river flows, on average, to decrease. These decreases are simulated to be greatest in the south and east of the UK, with droughts becoming longer and more severe. While high flows also decrease on average, an increased number of extremes are exhibited, implying a greater number of extreme flood events in the future, particularly in the north and west of the UK. In the urban development scenarios, for flood events there is an increase in flow with the increased urbanization, with the 1 in 3-year peak flow event showing the greatest increase. The natural flood management scenarios consider the effect of increasing woodland and adding surface water storage ponds. The inclusion of these features produces a complex response but overall, the modelling shows a reduction in low, median, and high flows, although the more extreme the flow event the smaller the percentage change in flow. Simulated timeseries and summary metric datasets are freely available on the CEDA archive.

Physical mechanism-based simulation methods for soil and ground heat flux from observations at multi-sites

It is important to obtain the soil heat flux at a specific depth (Gz) and at the ground’s surface (G0) to understand the surface energy balance and climate change. However, previous evaluations of Gz and G0 simulations based on multiple sites have been relatively insufficient. We evaluated the accuracy of using physical mechanism-based gradient and calorimetry methods to simulate Gz and changes in heat storage (ΔS) based on observations collected at 28 sites. Three methods for simulating G0 were also evaluated using triple collocation (TC) and extended triple collocation (ETC) metrics. The ability of the gradient method to simulate Gz was strong, particularly for the lower soil layer, with a median Nash-Sutcliffe efficiency (NSE), correlation coefficient (CC), and index of agreement for the evaluated sites of 0.76, 0.86, and 0.91, respectively. The accuracy of the calorimetry method for simulating ΔS was generally lower than that of the gradient method for determining Gz. Overall, the combination of gradient analysis and calorimetry showed the best performance in simulating G0. The TC evaluation showed that the combination of gradient and calorimetry had a strong ability to simulate G0 in most land cover types (root mean square error (RMSE) <4 W/m2) but this ability was lost in the snow cover type (RMSE > 10 W/m2). When evaluated by ETC, the three methods had strong effects only on alpine meadow, floodland, and shrub types (CC > 0.70). Appropriate evaluation indices were recommended based on the application of the model.

A process-driven deep learning hydrological model for daily rainfall-runoff simulation

Although deep learning (DL) models, especially long-short-term memory (LSTM), demonstrate greater accuracy than process-based models in rainfall-runoff simulation, the predictions from process-based models are more physical than DL models. The main reason is that DL models have almost no process understanding capabilities like process-based models beyond their fitting capability. In this study, we developed a process-driven DL model under a unified DL architecture to improve the process awareness of DL models. To implement the model, a conceptual hydrological model (EXP-HYDRO) is implanted into a recurrent neural network (RNN) cell as a process driver for providing multi-sub-process variables related to the runoff process, and an Entity-Aware LSTM (EA-LSTM) cell is incorporated as a post-processor layer, resulting in the Process-driven RNN-EA-LSTM (PRNN-EA-LSTM). Under the assistance of the process driver, the model performance of PRNN-EA-LSTM on the 531 catchments from the Catchment Attributes and Meteorology for Large-sample Studies dataset is more robust than the pure DL model, and better than using only EXP-HYDRO as the input of EA-LSTM (i.e., EXP-HYDRO-EA-LSTM). Specifically, the median Nash-Sutcliffe efficiency (NSE) of PRNN-EA-LSTM in local and regional simulation is 0.03 and 0.02 higher than LSTM and 0.01 higher than EXP-HYDRO-EA-LSTM. Additionally, PRNN-EA-LSTM significantly enhances the low flow simulations and reduces the catchments number with negative NSE. This study demonstrates that process-based models can help DL models better represent the rainfall-runoff relationship under a unified architecture. Consequently, integrating the adaptability of process-based models into the DL architecture is anticipated to bolster the process understanding capabilities of DL models.

Investigating sources of variability in closing the terrestrial water balance with remote sensing

Abstract. Remote sensing (RS) data are becoming an increasingly important source of information for water resource management as they provide spatially distributed data on water availability and use. However, in order to guide appropriate use of the data, it is important to understand the impact of the uncertainties of RS data on water resource studies. Previous studies have shown that the degree of closure of the water balance from remote sensing data is highly variable across basins and that different RS products vary in their levels of accuracy depending on climatological and geographical conditions. In this paper, we analyzed the water-balance-derived runoff from global RS products for 931 catchments across the globe. We compared time series of runoff estimated through a simplified water balance equation using three precipitation (CHIRPS, GPM, and TRMM), five evapotranspiration (MODIS, SSEBop, GLEAM, CMRSET, and SEBS), and three water storage change (GRACE-CSR, GRACE-JPL, and GRACE-GFZ) RS datasets with monthly in situ discharge data for the period 2003–2016. Results were analyzed through the lens of 10 quantifiable catchment characteristics in order to investigate correlations between catchment characteristics and the quality of RS-based water balance estimates of runoff and whether specific products performed better than others under certain conditions. The median Nash–Sutcliffe efficiency (NSE) for all gauges and all product combinations was −0.02, and only 44.9 % of the time series reached a positive NSE. A positive NSE could be obtained for 73.7 % of stations with at least one product combination, while the overall best-performing product combination was positive for 58.4 % of stations. This confirms previous findings that the best-performing products cannot be globally established. When investigating the results by catchment characteristic, all combinations tended to show similar correlations between catchment characteristics and the quality of estimated runoff, with the exception of combinations using MODIS evapotranspiration, for which the correlation was frequently reversed. The combinations with the GPM precipitation product generally performed worse than the CHIRPS and TRMM data. However, this can be attributed to the fact that the GPM data are available at higher latitudes compared to the other products, where performance is generally poorer. When removing high-latitude stations, this difference was eliminated, and GPM and TRMM showed similar performance. The results show the highest positive correlation between highly seasonal rainfall and runoff NSE. On the other hand, increasing snow cover, altitude, and latitude decreased the ability of the RS products to close the water balance. The catchment's dominant climate zone was also found to be correlated with time series performance, with the tropical areas providing the highest (median NSE = 0.11) and arid areas the lowest (median NSE = −0.09) NSE values. No correlation was found between catchment area and runoff NSE. The results highlight the importance of further studies on the uncertainties of the different data products and how these interact when combining them, as well as of new approaches to using the data rather than simple water-balance-type approaches. Efforts to improve specific satellite products can also be better targeted using the results of this study.

Retracted: Spatiotemporal convolutional long short-term memory for regional streamflow predictions

Rainfall-runoff (RR) modelling is a challenging task in hydrology, especially at the regional scale. This work presents an approach to simultaneously predict daily streamflow in 86 catchments across the US using a sequential CNN-LSTM deep learning architecture. The model effectively incorporates both spatial and temporal information, leveraging the CNN to encode spatial patterns and the LSTM to learn their temporal relations. For training, a year-long spatially distributed input with precipitation, maximum temperature, and minimum temperature for each day was used to predict one-day streamflow. The trained CNN-LSTM model was further fine-tuned for three local sub-clusters of the 86 stations, assessing the significance of fine-tuning in model performance. The CNN-LSTM model, post fine-tuning, exhibited strong predictive capabilities with a median Nash-Sutcliffe efficiency (NSE) of 0.62 over the test period. Remarkably, 65% of the 86 stations achieved NSE values greater than 0.6. The performance of the model was also compared to different deep learning models trained using a similar setup (CNN, LSTM, ANN). An LSTM model was also developed and trained individually to predict for each of the stations using local data. The CNN-LSTM model outperformed all the models which was trained regionally, and achieved a comparable performance to the local LSTM model. Fine-tuning improved the performance of all models during the test period. The results highlight the potential of the CNN-LSTM approach for regional RR modelling by effectively capturing complex spatiotemporal patterns inherent in the RR process.

MA-SARNet: A one-shot nowcasting framework for SAR image prediction with physical driving forces

Remote sensing imagery is one of the most widely used data sources for large-scale Earth observations with consistent spatial and temporal quality. However, the current usage scenarios of Earth’s surface remote sensing images, such as those generated from Landsat, Sentinel 2, and SAR, are largely limited to retrospective tasks, as they are often used to reconstruct existing phenomena, such as land use change, flood inundation, and wildfire. This study proposes MA-SARNet, a one-shot nowcasting framework built with a modified MA-Net structure and ResNet50 as the backbone, to predict the backscatter values of Synthetic-Aperture Radar (SAR) images using the previous SAR observations, precipitation, soil moisture, and geomorphic data layers as input. The model was trained, validated, and tested with SAR images collected during the catastrophic 2019 Midwest U.S. Floods that affected several states on the Missouri and Mississippi River tributaries. Compared to the benchmark performance, model predictions show an increase of 31.9% and 17.8% for the mean and median AAI (Assemble Accuracy Index) and an increase of 37.9% and 15.1% for the mean and median NSE (Nash-Sutcliffe Efficiency) on the test set. Results showed that the flood extent derived from backscatter predictions has less misclassifications caused by pixel-level noise compared to the flood map derived using the real backscatters. Results from spatial and temporal robustness tests demonstrate that the model has sufficient generalization potential and does not require further fine-tuning to work with new data, and therefore proves its usefulness in real-time Earth surface prediction tasks that are aimed at fast response to and mitigation for upcoming floods on a tight time schedule.

Modeling bioinfiltration surface dynamics through a hybrid geomorphic-infiltration model

Abstract Bioinfiltration systems are an increasingly prevalent mechanism for urban stormwater mitigation. One major challenge for the sustainability of bioinfiltration systems is erosion and channelization due to high bed shear stresses developed during large storm events. Sedimentation within these systems could also impact their performance as fine sediment may clog pathways necessary for infiltration. Understanding the geomorphology, shear stress, and sediment flux in the system can help predict maintenance needs associated with erosion and deposition. The current study introduces a framework for addressing this problem by combining a sediment transport model, FaSTMECH, with the Green-Ampt infiltration model. A comparison of observed and predicted ponding depths shows very good agreement (median Nash–Sutcliffe efficiency coefficient = 0.93) and demonstrates the ability of this novel framework in predicting the hydraulics and morphology within a bioswale bioinfiltration system. The framework introduced in this study opens the door to understanding sediment transport dynamics within a bioswale, which has the potential to advance planning and design to minimize impacts due to excessive erosion or deposition within bioswale bioinfiltration systems.

Enhancing process-based hydrological models with embedded neural networks: A hybrid approach

Deep learning (DL) models have demonstrated exceptional performance in hydrological modeling; however, they are limited by their inability to output untrained hydrological variables and lack of interpretability compared to process-based hydrological models. We propose a hybrid approach that combines the conceptual EXP-Hydro model with embedded neural networks (ENNs), replacing its internal modules while maintaining adherence to hydrological knowledge. The resulting hybrid model can predict untrained hydrological variables without requiring post-processing or pre-training procedures. We tested 15 hybrid models that replace different internal modules across 569 basins in the contiguous United States using the CAMELS dataset. Additional experiments were conducted to generalize hydrological relationships within ENNs and further use them to improve the EXP-Hydro model's performance. Results show that all hybrid scenarios outperform the ordinary EXP-Hydro model, with an optimal median Nash-Sutcliffe efficiency (NSE) of 0.701 in the evaluation period – comparable to state-of-the-art LSTM and conceptual hydrological model featuring an error-correcting post-processor. Reasonable patterns of runoff and snow-related processes are captured by ENNs in respective hybrid models. We further used the runoff (snow-related) pattern to improve the ordinary EXP-Hydro model with median NSE increasing from 0.496 to 0.567 (raising median NSE from 0.601 to 0.677 in snow-influenced region). Our study highlights the potential for using ENNs in enhancing process-based hydrological models' performance while maintaining interpretability within a novel hybrid framework.

Salinity Modeling Using Deep Learning with Data Augmentation and Transfer Learning

Salinity management in estuarine systems is crucial for developing effective water-management strategies to maintain compliance and understand the impact of salt intrusion on water quality and availability. Understanding the temporal and spatial variations of salinity is a keystone of salinity-management practices. Process-based numerical models have been traditionally used to estimate the variations in salinity in estuarine environments. Advances in data-driven models (e.g., deep learning models) make them effective and efficient alternatives to process-based models. However, a discernible research gap exists in applying these advanced techniques to salinity modeling. The current study seeks to address this gap by exploring the innovative use of deep learning with data augmentation and transfer learning in salinity modeling, exemplified at 23 key salinity locations in the Sacramento–San Joaquin Delta which is the hub of the water-supply system of California. Historical, simulated (via a hydrodynamics and water quality model), and perturbed (to create a range of hydroclimatic and operational scenarios for data-augmentation purposes) flow, and salinity data are used to train a baseline multi-layer perceptron (MLP) and a deep learning Residual Long-Short-Term Memory (Res-LSTM) network. Four other deep learning models including LSTM, Residual Network (ResNet), Gated Recurrent Unit (GRU), and Residual GRU (Res-GRU) are also examined. Results indicate that models pre-trained using augmented data demonstrate improved performance over models trained from scratch using only historical data (e.g., median Nash–Sutcliffe efficiency increased from around 0.5 to above 0.9). Moreover, the five deep learning models further boost the salinity estimation performance in comparison with the baseline MLP model, though the performance of the latter is acceptable. The models trained using augmented data are then (a) used to develop a web-based Salinity Dashboard (Dashboard) tool that allows the users (including those with no machine learning background) to quickly screen multiple management scenarios by altering inputs and visualizing the resulting salinity simulations interactively, and (b) transferred and adapted to estimate observed salinity. The study shows that transfer learning results more accurately replicate the observations compared to their counterparts from models trained from scratch without knowledge learned and transferred from augmented data (e.g., median Nash–Sutcliffe efficiency increased from around 0.4 to above 0.9). Overall, the study illustrates that deep learning models, particularly when pre-trained using augmented data, are promising supplements to existing process-based models in estuarine salinity modeling, while the Dashboard enables user engagement with those pre-trained models to inform decision-making efficiently and effectively.

How to enhance hydrological predictions in hydrologically distinct watersheds of the Indian subcontinent?

AbstractAccurate hydrological predictions are required to prepare for the impacts of climate change, especially in India, which experiences frequent floods and droughts. However, the complex hydrological processes of its distinct watersheds and limited data make it challenging to deliver highly‐performant hydrologic predictions using conventional models. Moreover, it remains uncertain where the limits of predictability are and whether recently‐popular deep learning approaches can offer significant improvements. Here, we tested the first instance of the hydrologic model based on long short‐term memory (LSTM) for 55 Indian watersheds, using a new dataset comprising forcing, attributes, and discharge data. Our results show that the LSTM model provides much‐improved performance compared to conventional models in India, providing a median Nash‐Sutcliffe efficiency (NSE) of 0.56. The LSTM model trained on all the watersheds is more favourable to those trained on individual or homogeneous watersheds, as it benefits from a broader range of hydrological processes and patterns in the input data. However, the LSTM model performs poorly for non‐perennial, large, and semi‐arid climate zone watersheds due to its inability to simulate the complex hydrological processes specific to these environments. Integrating lagged observations with the LSTM model (referred to as DI‐LSTM) improved the predictions in such watersheds and enhanced the median NSE to 0.76 by capturing the temporal dependencies and historical patterns that influence hydrological processes. Overall, the contrast of model performance across watersheds suggests major limitations could be associated with the quality of forcing data, and the slow flow or groundwater processes are highly important in the Indian subcontinent. Notably, both LSTM and DI‐LSTM models performed reasonably well for predictions in ungauged watersheds. The findings of this study demonstrate that data‐sparse countries, too, can benefit from big‐data deep learning and point out further avenues toward model improvements.

Development and evaluation of the channel routing model and parameters within the National Water Model

AbstractThe National Water Model (NWM) was deployed by the National Oceanic and Atmospheric Administration to simulate operational forecasts of hydrologic states across the continental United States. This paper describes the geospatial river network (“hydro‐fabric”), physics, and parameters of the NWM, elucidating the challenges of extrapolating parameters a large scale with limited observations. A set of regression‐based channel geometry parameters are evaluated for a subset of the 2.7 million NWM reaches, and the riverine compound channel scheme is described. Based on the results from regional streamflow experiments within the broader NWM context, the compound channel reduced the root mean squared error by 2% and improved median Nash–Sutcliffe efficiency by 16% compared with a non‐compound formulation. Peak event analysis from 910 peak flow events across 26 basins matched from the US Flash Flood Observation Database revealed that the mean timing error is 3 h lagged behind the observations. The routing time step was also tested, for 5‐min (default, operational setting) and 1‐h increments. The model was computationally stable and able to convey the flood peaks, although the hydrograph shape and peak timing were altered.

Impact of snow distribution modelling for runoff predictions

AbstractSnow in the mountains is essential for the water cycle in cold regions. The complexity of the snow processes in such an environment makes it challenging for accurate snow and runoff predictions. Various snow modelling approaches have been developed, especially to improve snow predictions. In this study, we compared the ability to improve runoff predictions in the Överuman Catchment, Northern Sweden, using different parametric representations of snow distribution. They included a temperature-based method, a snowfall distribution (SF) function based on wind characteristics and a snow depletion curve (DC). Moreover, we assessed the benefit of using distributed snow observations in addition to runoff in the hydrological model calibration. We found that models with the SF function based on wind characteristics better predicted the snow water equivalent (SWE) close to the peak of accumulation than models without this function. For runoff predictions, models with the SF function and the DC showed good performances (median Nash–Sutcliffe efficiency equal to 0.71). Despite differences among the calibration criteria for the different snow process representations, snow observations in model calibration added values for SWE and runoff predictions.

Assimilation of transformed water surface elevation to improve river discharge estimation in a continental-scale river

Abstract. Quantifying continental-scale river discharge is essential for understanding the terrestrial water cycle, but it is susceptible to errors caused by a lack of observations and the limitations of hydrodynamic modeling. Data assimilation (DA) methods are increasingly used to estimate river discharge in combination with emerging river-related remote sensing products (e.g., water surface elevation (WSE), water surface slope, river width, and flood extent). However, directly comparing simulated WSE to satellite altimetry data remains challenging (e.g., because of large biases between simulations and observations or uncertainties in parameters), and large errors can be introduced when satellite observations are assimilated into hydrodynamic models. In this study we performed direct, anomaly, and normalized value assimilation experiments to investigate the capacity of DA to improve river discharge within the current limitations of hydrodynamic modeling. We performed hydrological DA using a physically based empirical localization method applied to the Amazon basin. We used satellite altimetry data from ENVISAT, Jason 1, and Jason 2. Direct DA was the baseline assimilation method and was subject to errors due to biases in the simulated WSE. To overcome these errors, we used anomaly DA as an alternative to direct DA. We found that the modeled and observed WSE distributions differed considerably (e.g., differences in amplitude, seasonal flow variation, and a skewed distribution due to limitations of the hydrodynamic models). Therefore, normalized value DA was performed to improve discharge estimation. River discharge estimates were improved at 24 %, 38 %, and 62 % of stream gauges in the direct, anomaly, and normalized value assimilations relative to simulations without DA. Normalized value assimilation performed best for estimating river discharge given the current limitations of hydrodynamic models. Most gauges within the river reaches covered by satellite observations accurately estimated river discharge, with the Nash–Sutcliffe efficiency (NSE) &gt; 0.6. The amplitudes of WSE variation were improved in the normalized DA experiment. Furthermore, in the Amazon basin, normalized assimilation (median NSE =0.50) improved river discharge estimation compared to open-loop simulation with the global hydrodynamic model (median NSE =0.42). River discharge estimation using direct DA methods was improved by 7 % with calibration of river bathymetry based on NSE. The direct DA approach outperformed the other DA approaches when runoff was considerably biased, but anomaly DA performed best when the river bathymetry was erroneous. The uncertainties in hydrodynamic modeling (e.g., river bottom elevation, river width, simplified floodplain dynamics, and the rectangular cross-section assumption) should be improved to fully realize the advantages of river discharge DA through the assimilation of satellite altimetry. This study contributes to the development of a global river discharge reanalysis product that is consistent spatially and temporally.

Achieving Realistic Runoff in the Western United States with a Land Surface Model Forced by Dynamically Downscaled Meteorology

Abstract In this study, we calibrate a regional climate model’s (RCM) underlying land surface model (LSM). In addition to providing a realistic representation of runoff across the hydroclimatically diverse western United States, this is done to take advantage of the RCM’s ability to physically resolve meteorological forcing data in ungauged regions, and to prepare the calibrated hydrologic model for tight coupling, or the ability to represent land surface–atmosphere interactions, with the RCM. Specifically, we use a 9-km resolution meteorological forcing dataset across the western United States, from the fifth generation ECMWF Reanalysis (ERA5) downscaled by the Weather Research Forecasting (WRF) regional climate model, as an offline forcing for Noah-Multiparameterization (Noah-MP). We detail the steps involved in producing an LSM capable of accurately representing runoff, including physical parameterization selection, parameter calibration, and regionalization to ungauged basins. Based on our model evaluation from 1954 to 2021 for 586 basins with daily natural streamflow, the streamflow bias is reduced from 24.2% to 4.4%, and the median daily Nash–Sutcliffe efficiency (NSE) is improved from 0.12 to 0.36. When validating against basins with monthly natural streamflow data, we obtain a similar reduction in bias and a median monthly NSE improvement from 0.18 to 0.56. In this study, we also discover the optimal setup when using a donor-basin method to regionalize parameters to ungauged basins, which can vary by 0.06 NSE for unique designs of this regionalization method. Significance Statement This study provides useful guidance for improving a land surface model to accurately represent runoff across a spatially extensive and hydroclimatically diverse region (the western United States). The land surface model is updated to represent runoff more accurately at gauged basins, and then additionally updated for basins without observational data using a mathematical approach called the donor-basin method. We make use of a regional climate model’s reanalysis-derived meteorological data and its underlying land surface model to achieve realistic runoff. The calibrated land surface model can thus be tightly coupled in subsequent studies in a manner that should more accurately reflect runoff conditions. Findings from this study will serve as a useful reference for the atmospheric (and hydrologic) modeling communities and their ability to represent large-scale hydrology accurately.

Evaluation of Sentinel-3A altimetry over Songhua river Basin

Satellite radar altimetry is an important tool for monitoring water levels of inland water bodies, particularly in poorly gauged or ungauged areas. This study reveals for the first time the performance of water levels derived from Sentinel-3A altimetry in the Songhua River Basin in northeast China. This study compares water levels derived via different retrackers (SAMOSA+, OCOG, Ocean, MWaPP, and NPPR) from altimetry in virtual stations (VSs) to water levels measured by gauges. To improve the influence of sandbars or islands and surrounding land cover, a modified retracker called AMPD-PF (Automatic Multiscale-based Peak Detection retracker using Physically-based model Fitting) is implemented here, by combing the physical modeling of SAR altimeter waveform with a shortest-path algorithm to find the best possible sub-waveform representing the signal of water bodies. The accuracy assessment is based on comparing water level anomalies from 26 VSs and those from the nearest adjacent gauges with an average distance of 8.98 km in the Songhua River Basin from 2016 to 2019. The median of the root-mean-square error (RMSE) differences at the virtual stations ranges from 0.46 to 0.82 m, and the median of the Nash-Sutcliffe efficiency (NSE) among different retracker varies from 0.49 to 0.85 among different retracker. The performance of Sentinel-3A altimetry in the Songhua River Basin is not poor. Results also show a highest accuracy of water levels from AMPD-PF, especially for the non-flood season (January to June and October to December). In addition, we find that, especially for the virtual stations unaffected by the sand bars and islands, AMPD-PF give the best performance with a median RMSE value of 0.33 m and a median NSE value of 0.86. Factor analysis indicates that the variations of these metrics can not be directly related to the river width but might be related to the cultivated land types and the presence of sand bars and islands in the vicinity of rivers over the Songhua River Basin. Moreover, it is confirmed that a near-parallel orientation of the river with respect to the satellite ground track often leads to poorer performance at virtual stations. This suggests that if one is aiming at calibrating a hydrodynamic or hydrological model by combining altimetry, the VSs with the near-parallel orientation are not necessary to be considered when the percentage of the non-parallel crossings of all crossings is big enough.

WaterBench-Iowa: a large-scale benchmark dataset for data-driven streamflow forecasting

Abstract. This study proposes a comprehensive benchmark dataset for streamflow forecasting, WaterBench-Iowa, that follows FAIR (findability, accessibility, interoperability, and reuse) data principles and is prepared with a focus on convenience for utilizing in data-driven and machine learning studies, and provides benchmark performance for state of art deep learning architectures on the dataset for comparative analysis. By aggregating the datasets of streamflow, precipitation, watershed area, slope, soil types, and evapotranspiration from federal agencies and state organizations (i.e., NASA, NOAA, USGS, and Iowa Flood Center), we provided the WaterBench-Iowa for hourly streamflow forecast studies. This dataset has a high temporal and spatial resolution with rich metadata and relational information, which can be used for a variety of deep learning and machine learning research. We defined a sample benchmark task of predicting the hourly streamflow for the next 5 d for future comparative studies, and provided benchmark results on this task with sample linear regression and deep learning models, including long short-term memory (LSTM), gated recurrent units (GRU), and sequence-to-sequence (S2S). Our benchmark model results show a median Nash-Sutcliffe efficiency (NSE) of 0.74 and a median Kling-Gupta efficiency (KGE) of 0.79 among 125 watersheds for the 120 h ahead streamflow prediction task. WaterBench-Iowa makes up for the lack of unified benchmarks in earth science research and can be accessed at Zenodo https://doi.org/10.5281/zenodo.7087806 (Demir et al., 2022a).

Accuracy of five ground heat flux empirical simulation methods in the surface-energy-balance-based remote-sensing evapotranspiration models

Abstract. Based on the assessment from 230 flux site observations, intra-day and daytime ground heat flux (G) accounted for 19.2 % and 28.8 % of the corresponding net radiation, respectively. This indicates that G plays an important role in remote-sensing (RS) energy-balance-based evapotranspiration (ET) models. The G empirical estimation methods have been evaluated at many individual sites, while there have been relatively few multi-site evaluation studies. The accuracy of the five empirical G simulation methods in the surface-energy-balance-based RS–ET models was evaluated using half-hourly observations. The linear coefficient (LC) method and the two methods embedded with the normalized difference vegetation index (NDVI) were able to accurately simulate a half-hourly G series at most sites. The mean and median Nash–Sutcliffe efficiency (NSE) values of all sites were generally higher than 0.50 in each half-hour period. The accuracy of each method varied significantly at different sites and at half-hour intervals. The highest accuracy was exhibited during 06:00–07:00 LST (all times hereafter are LST), followed by the period of 17:00–18:00. There were 92 % (211/230) sites with an NSE of the LC method greater than 0.50 at 06:30. It showed a slightly higher accuracy during nighttime periods than during daytime periods. The lowest accuracy was observed during the period of 10:00–15:30. The sites with an NSE exceeding 0.50 only accounted for 51 % (118/230) and 43 % (100/230) at 10:30 and 13:30, respectively. The accuracy of the model was generally higher in Northern Hemisphere sites than in Southern Hemisphere sites. In general, the highest and lowest accuracies were observed at the high- and low-latitude sites, respectively. The performance of the LC method and the methods embedded with NDVI were generally satisfactory at the Eurasian and North American sites, with the NSE values of most sites exceeding 0.70. Conversely, it exhibited relatively poor performance at the African, South American, and Oceanian sites, especially the African sites. Both the temporal and spatial distributions of the accuracy of the G simulation were positively correlated with the correlation between G and the net radiation. Although the G simulation methods accurately simulated the G series at most sites and time periods, their performance was poor at some sites and time periods. The application of RS ET datasets covering these sites requires caution. Further improvement of G simulations at these sites and time periods is recommended for the RS ET modelers. In addition, variable parameters are recommended in empirical methods of G simulation to improve accuracy. Instead of the Rn, finding another variable that has a physical connection and strong correlation with G might be a more efficient solution for the improvement, since the weak correlation between G and Rn is the main reason for the poor performance at these regions.

Improving Rainfall‐Runoff Model Reliability Under Nonstationarity of Model Parameters: A Hypothesis Testing Based Framework

AbstractRainfall‐runoff models are used to project the impact of climate change on water resources. Often, models are parameterized over historical climate prior to simulating runoff under changed climates. An increasing body of literature indicates that this assumption of model parameter invariance to climate can introduce significant simulation biases as projected climate deviates from historical climate. One way to consider the effect of climate on model parameterization is to regionalize model parameters using historical data and use that information to project parameters’ probable distribution under a changed climate. These approaches, termed trading space‐for‐time (TSFT), have shown promising results when simulating mean annual runoff, but their ability to reproduce time varying daily runoff response remains unexplored. Here, we design experiments to identify possible ways to account for the influence of climate on model parameters by building upon a recently developed Whittaker‐biome‐based TSFT framework. We perform 35 experiments, each representing different hypotheses regarding how model parameters may be altered with climate. We test these hypotheses for seven watersheds across the conterminous United States that undergo significant climate change between 1980–1990 and 1999–2009. Our results show that updated parameters for the best experiment increased median Nash–Sutcliffe Efficiency for the two most climate‐impacted watersheds in our analyses from 0.37 to 0.35 to 0.64 and 0.49, respectively. Improvements in simulation performance for watersheds transitioning from the boreal forest and the woodland/shrubland biomes to the temperate seasonal forest biome were attributed to altered snow and routing parameters, respectively.

A Framework for Mechanistic Flood Inundation Forecasting at the Metropolitan Scale

AbstractUrban flooding from extreme precipitation and storm surge is a growing threat to cities, and detailed forecasts of urban inundation are needed for emergency response. We present a mechanistic framework to simulate flood inundation over metropolitan‐wide areas at fine resolution (3 m). A dual‐grid shallow‐water model is used to overcome computational bottlenecks, and an application to Hurricane Harvey focused on pluvial flooding provides a multi‐dimensional assessment of predictive skill. A hindcast model is shown to simulate peak stage across 41 stream gages with a mean absolute error (MAE) of 0.63 m, and hourly stage levels over a 5‐day period with a median MAE and Nash‐Sutcliffe Efficiency (NSE) of 0.74 m and 0.55, respectively. Peak flood level across 228 high water marks (HWMs) were captured with an MAE of 0.69 m. A forecast model forced by Quantitative Precipitation Forecast data is shown to be only marginally less accurate than the hindcast model. Peak stage is simulated with an MAE of 0.86 m, hourly stage is captured with a median MAE and NSE of 0.90 m and 0.41, respectively, and HWMs are captured with an MAE of 0.77 m. The forecast system also achieves hit rates of 90% and 73% predicting distress calls and FEMA damage claims, respectively, based on simulated flood depth. These results demonstrate the potential to operationally forecast pluvial flood inundation in the U.S. with the timeliness and accuracy needed for early warning, and we also highlight future research needs.

Comparative analysis of water quality prediction performance based on LSTM in the Haihe River Basin, China.

As the most water shortage and water polluted area in China, the water quality prediction is of utmost needed and important in Haihe River Basin for its water resource management. The long short-term memory (LSTM) has been a widely used tool for water quality forecast in recent years. The performance and adaptability of LSTM for water quality prediction of different indicators needs to be discussed before it adopted in a specific basin. However, literature contains very few studies on the comparative analysis of the various prediction accuracy of different water quality indicators and the causes, especially in Haihe River Basin. In this study, LSTM was employed to predict biochemical oxygen demand (BOD), permanganate index (CODMn), dissolved oxygen (DO), ammonia nitrogen (NH3-N), total phosphorus (TP), hydrogen ion concentration (pH), and chemical oxygen demand digested by potassium dichromate (CODCr). According to results under 24 different input conditions, it is demonstrated that LSTMs present better predicting on BOD, CODMn, CODCr, and TP (median Nash-Sutcliffe efficiency reaching 0.766, 0.835, 0.837, and 0.711, respectively) than NH3-N, DO, and pH (median Nash-Sutcliffe efficiency of 0.638, 0.625, and 0.229, respectively). Besides, the performance of LSTM to predict water quality is linearly related to the maximum value of temporal autocorrelation and cross-correlation coefficients of water quality indicators calculated by maximal information coefficient with the coefficients of determination of 0.79 to approximately 0.80. This study would provide new knowledge and support for the practical application and improvement of the LSTM in water quality prediction.