For operational water management in lowlands and polders (for instance, in the Netherlands), lowland hydrological models are used for flow prediction, often as an input for a real‐time control system ...to steer water with pumps and weirs to keep water levels within acceptable bounds. Therefore, proper initialization of these models is essential. The ensemble Kalman filter (EnKF) has been widely used due to its relative simplicity and robustness, while the unscented Kalman filter (UKF) has received little attention in the operational context. Here, we test both UKF and EnKF using a lowland lumped hydrological model. The results of a reforecast experiment in an operational context using an hourly time step show that when using nine ensemble members, both filters can improve the accuracy of the forecast by updating the state of a lumped hydrological model (Wageningen Lowland Runoff Simulator, WALRUS) based on the observed discharge, while UKF has achieved better performance than EnKF. Additionally, we show that an increase in the ensemble members does not necessarily mean a significant increase in performance. WALRUS model with either UKF or EnKF could be considered for hydrological forecasting for supporting water management of polders and lowlands, with UKF being the computationally leaner option.
Key Points
The ensemble Kalman filter and unscented Kalman filter are utilized to improve forecast skill of a hydrological model by state estimation
Both approaches have improved the forecast performance, with unscented Kalman filter being the computationally leaner option
The unscented Kalman filter can be considered as a new and effective option to assimilate observations into low‐dimensional models
Radar rainfall nowcasting, the process of statistically extrapolating the most recent rainfall observation, is increasingly used for very short range rainfall forecasting (less than 6 hr ahead). We ...performed a large‐sample analysis of 1,533 events, systematically selected for 4 event durations and 12 lowland catchments (6.5–957 km2), to determine the predictive skill of nowcasting. Four algorithms are tested and compared with Eulerian Persistence: Rainymotion Sparse, Rainymotion DenseRotation, Pysteps deterministic, and Pysteps probabilistic with 20 ensemble members. We focus on the dependency of nowcast skill on event duration, season, catchment size, and location. Maximum skillful lead times increase for longer event durations, due to the more persistent character of these events. For all four event durations, Pysteps deterministic attains the longest average decorrelation times, with 25 min for 1‐hr durations, 40 min for 3 hr, 56 min for 6 hr, and 116 min for 24 hr. During winter, with more persistent stratiform precipitation, we find three times lower mean absolute errors than for convective summer precipitation. Higher skill is also found after spatially upscaling the forecast. Catchment location matters too: Given the prevailing storm movement, two times higher skillful lead times are found downwind than upwind toward the edge of the domain. In most cases, Pysteps algorithms outperform the Rainymotion benchmark algorithms. We speculate that most errors originate from growth and dissipation processes which are not or only partially (stochastically) accounted for.
Plain Language Summary
Early warning systems are a key instrument for timely responses to flood risk. These warnings depend on accurate weather forecasts. Most numerical weather prediction models have trouble to accurately forecast the timing and especially the location of rainfall on time scales of less than 6 hr. This time frame is important for decisions in small, mountainous, polder, and urban basins, where river discharge responds quickly to rainfall. Radar rainfall nowcasting, the process of statistically extrapolating the most recent rainfall observations, has the potential to provide forecasts up to 6 hr. We considered 1,533 rainfall events to evaluate the quality of five nowcasting methods. Nowcasts are better for longer‐lasting rain storms with relatively low intensity, which typically occur in winter, than for short, intensive storms, which typically occur in summer. Forecasts are useful up to 2 hr ahead for events of 1 day, while this is only 25 min for 1‐hr events. Moreover, nowcasts are better for larger basins and in the downwind direction with respect to the prevailing storm movement. Hence, nowcasts are useful, but improvements are needed to be able to forecast longer ahead.
Key Points
A large‐sample analysis of 1,533 events is used to determine the skill of radar rainfall nowcasts
Evaluation takes place with four nowcasting algorithms from the Pysteps and Rainymotion libraries
Nowcast skill is found to depend on event duration, season, catchment size, and location
Display omitted
•ECMWF ensemble reforecasts of precipitation and temperature were tested for biases.•An attempt was made to reduce these biases through statistical post-processing.•This resulted in ...modest improvements in the quality of the forcing ensembles.•The effect on streamflow ensembles was explored by verifying against simulated flow.•At all spatial scales considered, the improvements in streamflow quality were modest.
The ECMWF temperature and precipitation ensemble reforecasts are evaluated for biases in the mean, spread and forecast probabilities, and how these biases propagate to streamflow ensemble forecasts. The forcing ensembles are subsequently post-processed to reduce bias and increase skill, and to investigate whether this leads to improved streamflow ensemble forecasts. Multiple post-processing techniques are used: quantile-to-quantile transform, linear regression with an assumption of bivariate normality and logistic regression. Both the raw and post-processed ensembles are run through a hydrologic model of the river Rhine to create streamflow ensembles. The results are compared using multiple verification metrics and skill scores: relative mean error, Brier skill score and its decompositions, mean continuous ranked probability skill score and its decomposition, and the ROC score. Verification of the streamflow ensembles is performed at multiple spatial scales: relatively small headwater basins, large tributaries and the Rhine outlet at Lobith. The streamflow ensembles are verified against simulated streamflow, in order to isolate the effects of biases in the forcing ensembles and any improvements therein. The results indicate that the forcing ensembles contain significant biases, and that these cascade to the streamflow ensembles. Some of the bias in the forcing ensembles is unconditional in nature; this was resolved by a simple quantile-to-quantile transform. Improvements in conditional bias and skill of the forcing ensembles vary with forecast lead time, amount, and spatial scale, but are generally moderate. The translation to streamflow forecast skill is further muted, and several explanations are considered, including limitations in the modelling of the space–time covariability of the forcing ensembles and the presence of storages.
Moving toward high‐resolution gridded hydrologic models asks for novel parametrization approaches. A high‐resolution conceptual hydrologic model (wflow_sbm) was parameterized for the Rhine basin in ...Europe based on point‐scale (pedo)transfer functions, without further calibration of effective model parameters on discharge. Parameters were estimated on the data resolution, followed by upscaling of parameter fields to the model resolution. The method was tested using a 6‐hourly time step at four model resolutions (1.2, 2.4, 3.6, and 4.8 km), followed by a validation with discharge observations and a comparison with actual evapotranspiration (ETact) estimates from an independent model (DMET Land Surface Analysis Satellite Application Facility). Additionally, the scalability of parameter fields and simulated fluxes was tested. Validation of simulated discharges yielded Kling‐Gupta Efficiency (KGE) values ranging from 0.6 to 0.9, except for the Alps where a volume bias caused lower performance. Catchment‐averaged temporal ETact dynamics were comparable with independent ET estimates (KGE ≈ 0.7), although wflow_sbm model simulations were on average 115 mm yr−1 higher. Spatially, the two models were less in agreement (SPAEF = 0.10), especially around the Rhine valley. Consistent parameter fields were obtained, and by running the model at the different resolutions, preserved ETact fluxes were found across the scales. For recharge, fluxes were less consistent with relative errors around 30% for regions with high drainage densities. However, catchment‐averaged fluxes were better preserved. Routed discharge in headwaters was not consistent across scales, although simulations for the main Rhine River were. Better processing (scale independent) of the river and drainage network may overcome this issue.
Plain Language Summary
Hydrologic models are used for flood and drought predictions. Most models have parameters, and to increase model performance, hydrologists often tune these parameters by calibration. State‐of‐the‐art gridded hydrologic models have parameter sets per grid cell, leading to many parameters and making current calibration procedures far from ideal. Here, we tested the use of well‐known (pedo)transfer functions from literature to estimate these parameter values, something which can reduce the calibration burden. By using parameter‐specific upscaling rules to derive seamless parameter maps for the wflow_sbm model, which explicitly takes subsurface lateral flows into account, this gives a model which is scalable to different grid cell sizes. We assessed the approach on multiple model resolutions, and we found consistent parameter fields and the preservation of vertical fluxes. Only routed discharge, a key output, deteriorates for headwater catchments on coarser resolutions. We attribute this to model structure and the derivation procedure of the river network on different scales, resulting in the loss of lateral flow representation on coarser resolutions. Nevertheless, discharge and evapotranspiration simulations are similar to observations and other models. Hence, regionalization with literature transfer functions and upscaling techniques can further lower the calibration burden and enable predictions in ungauged basins.
Key Points
Seamless distributed parameter maps can be obtained for the gridded hydrologic model wflow_sbm with transfer functions from literature
Application of wflow_sbm with these seamless parameter maps yields simulation results with high KGE and NSE across the Rhine basin
Fluxes matched across model scales for evapotranspiration, but this match was considerably less for fluxes affected by (sub)surface flows
This paper presents a hybrid local‐global sensitivity analysis method termed the Distributed Evaluation of Local Sensitivity Analysis (DELSA), which is used here to identify important and unimportant ...parameters and evaluate how model parameter importance changes as parameter values change. DELSA uses derivative‐based “local” methods to obtain the distribution of parameter sensitivity across the parameter space, which promotes consideration of sensitivity analysis results in the context of simulated dynamics. This work presents DELSA, discusses how it relates to existing methods, and uses two hydrologic test cases to compare its performance with the popular global, variance‐based Sobol' method. The first test case is a simple nonlinear reservoir model with two parameters. The second test case involves five alternative “bucket‐style” hydrologic models with up to 14 parameters applied to a medium‐sized catchment (200 km2) in the Belgian Ardennes. Results show that in both examples, Sobol' and DELSA identify similar important and unimportant parameters, with DELSA enabling more detailed insight at much lower computational cost. For example, in the real‐world problem the time delay in runoff is the most important parameter in all models, but DELSA shows that for about 20% of parameter sets it is not important at all and alternative mechanisms and parameters dominate. Moreover, the time delay was identified as important in regions producing poor model fits, whereas other parameters were identified as more important in regions of the parameter space producing better model fits. The ability to understand how parameter importance varies through parameter space is critical to inform decisions about, for example, additional data collection and model development. The ability to perform such analyses with modest computational requirements provides exciting opportunities to evaluate complicated models as well as many alternative models.
Key Points
Sobol' and DELSA methods showed that they yield similar sensitivity results
DELSA provides more information at lower calculation costs than Sobol'
DELSA is recommended as effective and efficient method for parameter sensitivity