Journal of Hydrologic Engineering

This paper describes techniques used operationally by the National Weather Service (NWS) to prepare gridded multisensor (gauge, radar, and satellite) quantitative precipitation estimates (QPEs) for input into hydrologic forecast models and decision-making systems for river forecasting, flood and flash flood warning, and other hydrologic monitoring purposes. Advanced hydrologic prediction techniques require a spatially continuous representation of the precipitation field, and remote sensor input is critical to achieving this continuity. Although detailed descriptions of individual remote sensor estimation algorithms have been published, this review provides a summary of how the estimates from these various sources are merged into finished products. Emphasis is placed on the Weather Surveillance Radar–1988 Doppler (WSR-88D) Precipitation Processing System (PPS) and the Advanced Weather Interactive Processing System (AWIPS) Multisensor Precipitation Estimator (MPE) algorithms that utilize a combination of in situ rain gauges and remotely sensed measurements to provide a real-time suite of gridded radar and multisensor precipitation products. These two algorithm suites work in series to combine both computer-automated and human-interactive techniques, and they are used routinely at NWS field offices [river forecast centers (RFCs) and weather forecast offices (WFOs)] to support NWS’s broader hydrologic missions. The resulting precipitation products are also available to scientists and engineers outside the NWS; a summary of characteristics and sources of these products is presented.

This paper presents a comprehensive intercomparison analysis of different radar-based multisensor precipitation products generated operationally by the National Weather Service (NWS) Multisensor Precipitation Estimator (MPE) algorithm from the Weather Surveillance Radar–1988 Doppler version and concurrent rain gauge data. The analysis provides close insight into different effects of the increasing degree of complexitzy in the MPE algorithms. First, a gauge-only product produced by the MPE algorithm was assessed. Then six MPE products were analyzed: a radar-only product, a mean-field bias-adjusted product, a local bias-adjusted product, two products that are based on merging the bias-adjusted products with gauge observations, and a final product that includes human intervention by NWS forecasters. Data from a dense, carefully maintained experimental rain gauge cluster are used as an independent surface reference. A number of summary and conditional statistics are applied to the product intercomparisons. The results reported in this paper show that the most effective improvement of the rainfall products comes from applying the mean-field bias adjustment to the radar-only product. The analysis demonstrates that, for the current study site, some best-intended schemes for the optimal merging of radar and rain gauge data processing did not necessarily lead to clear improvements and, in some respects, caused accuracy degradation in the final products. This behavior by the MPE merging schemes is possibly attributed to the rather poor density of operational rain gauges that need to be available in real time for the implementation of such schemes. Future research is required to examine whether this behavior persists in other regions that may have better coverage and availability of operational rain gauges.

The Next-Generation Weather Radar (NEXRAD) system digital storm-total precipitation product (DSP; $2\mathrm{km}\times 1°$ at 4–7-min intervals) is examined for its quality in terms of precipitation quantity and spatial rainfall distribution. Observations from a network of 50 rain gauges in the Upper Guadalupe River Basin are compared to the DSP rainfall estimates from two radars: KEWX at New Braunfels, Texas, and KDFX (14 km east of Brackettville, Texas), for the period of September 2006 to June 2008. The rainfall data comparisons are conducted at three temporal scales: 6 min, 1 h, and storm-total accumulations, and at different distances from the radars, from near ($<50\mathrm{km}$) to middle (50–100 km), far (100–160 km), and very far ranges ($>160\mathrm{km}$). A strong range dependence is found from radar estimates, i.e., radar underestimates at near and very far ranges ($<50$ and $>160\mathrm{km}$), matches well or slightly overestimates at middle ranges (50–100 km), and overestimates at far ranges (100–160 km). The correlation coefficients between paired radar and gauge precipitation estimates are moderate ($r$ from 0.62 to 0.76) for both radars at the three time scales, however, they are spatially different (at the county level, showing a decreasing trend from middle to far ranges, passing west to east through Kerr, Kendall, Comal, and Guadalupe counties) under the KDFX umbrella. Under the KEWX umbrella, there is no trend of spatial correlation coefficients from county to county. Similarly, there is an overall increase of mean relative difference between radar and gauge values from middle to far ranges from the KDFX radar, but mixed results for the KEWX radar. The probabilities of rainfall detection (POD) for both radars are greater than 90% and are always higher than the gauge PODs, which increase from $\sim 10\%$ at the 6-min temporal scale, to $\sim 30\%$ at the hourly scale, to $\sim 70\%$ for the storm-total scale. There is no range dependence for POD. Overall, the DSP product suffers from severe range dependence, and its rainfall estimates agree better with rain gauge estimates during the warm season than during the cold season. However, the DSP’s advantages in high temporal and spatial resolutions and the near real time storm-total rainfall estimations provide great potential for near real time flash flood monitoring and forecasting at local and regional scales, although it still needs more testing and corrections, especially at very close ranges, far ranges, and during the cold season.

Correcting real-time radar-rainfall estimates for mean field systematic errors (bias) is normally accomplished through gauge-based adjustment procedures. This study explores two auxiliary data sources derived from cloud-to-ground (CG) lightning measurements and storm tracking applied on radar images in terms of providing microphysical information useful for improving the efficiency of gauge-based bias adjustment techniques. The CG information is used to classify storms into thunderstorms (T-storms) versus showers (i.e., storms without lightning) and the tracking algorithm is used to classify storms according to their stage of maturity (i.e., growing, maturing, and decaying). Data for this study are based on high-resolution radar-rainfall estimates (2-km spatial grid resolution at 15-min intervals) available over the South Florida Water Management District for a period of 11 months, along with corresponding rain gauge measurements from 120 gauges and CG occurrences from the National Lightning Detection Network. The radar error analysis for T-storms versus showers and for the different storm maturity stages indicate that storm tracking and CG contain significant microphysical information that can improve radar-rainfall estimation. It is shown that radar rain estimates tend to underestimate convective rainfall, primarily associated with the growing stage of the storms or the occurrences of CG lightning; showers and storms at mature or decay stages are shown to be better represented by the standard reflectivity-rainfall ($\mathit{Z}$-$R$) relationship used for convective and tropical storms in Florida. Results from this study indicate that information regarding storm maturity stage derived from tracking radar images, and to a lesser extent, CG observations, can be used to reduce variability in the $\mathit{Z}$-$R$ conversion and consequently improve accuracy in real-time radar-rainfall estimation.

Distributed rainfall information is necessary for making operational hydrologic predictions and retrospective studies. Rainfall measurements from radar require conversion of reflectivity into rainfall rates, but contain systematic errors (bias). Default $Z$-$R$ relationships that are commonly used are not adapted to particular radars and vary by season. Improved rainfall estimation may be obtained for a given season and radar installation by deriving $Z$-$R$ relationships. These derived $Z$-$R$ relationships are used before prospective storms cross over gauges within a network, i.e., before near real-time bias corrections are possible. This research focuses on the derivation of $Z$-$R$ relationships for rainfall estimation in the South Florida Water Management District. These relationships are derived through seasonal characterization of gauge and Next-Generation Radar (NEXRAD) system observations. Verification of the derived $Z$-$R$ relationships is evaluated using rain gauges withheld for cross validation. Methods used in forecasting include persistence, seasonal trends, autoregressive, and Kalman filter methods. Results demonstrate that radar-specific $Z$-$R$ relationships exhibit better efficiency than standard $Z$-$R$ relationships, and among the forecasting methods tested, all were advantageous over persistence with varying degrees of accuracy.

This work provides an analysis of the Weisseritz 2002 flood event in Eastern Germany, which was characterized by intense upslope storm movement and associated orographic enhancement of precipitation. Examination of the event is based on the availability of weather radar observations and post-event surveys to document the peak runoff rates. The analysis is based on the use of a set of statistics that clarify the dependence existing between spatial rainfall distribution, basin morphology, and runoff response. The statistics provide a description of overall spatial rainfall organization at the catchment scale in terms of concentration and dispersion statistics and storm motion, as a function of the flow distance coordinates, i.e., distance to the outlet measured along the flow path. The statistics are based on the observation that runoff routing through branched channel networks imposes an effective averaging of spatial rainfall at equal flow distance, despite the inherent spatial variability. A spatially distributed rainfall-runoff model is used to evaluate the effects of the rainfall spatial variability and storm velocity on flood modeling at various river sections. It is shown that storm velocity exhibits rather moderate values, despite the strong kinematic characteristics of individual storm elements. Consistently with this observation, hydrologic simulations show that orographically controlled rainfall spatial variability and storm motion have an almost negligible effect on flood response modeling. The drainage network structure filters out the effects of spatial rainfall variability for storm event hydrologic response, resulting in Weisseritz exhibiting striking uniformity of response to a storm with strongly varying precipitation. Overall, this shows the value of using the rainfall statistics introduced in this paper to isolate and describe the features of rainfall space-time variability that have significant impacts on runoff simulation.

Precipitation is the primary driver for hydrologic modeling. Because hydrologic models often require long-term, spatially distributed precipitation data sets for calibration and validation, a novel approach was developed to generate spatially distributed precipitation data using an artificial neural network (ANN) for the periods when Next-Generation Weather Radar (NEXRAD) data are either unavailable or the quality of the NEXRAD data is not good. The multilinear regression (MLR) technique was also evaluated for completeness. The study’s focus was the Saugahatchee Creek watershed in southeast Alabama. In the study area, the wet seasons are dominated by frontal precipitations, whereas the dry seasons primarily contain patchy, convective thunderstorms. The basic approach was to train and validate the ANN and MLR models using recent NEXRAD and rain gauge precipitations, and then use the trained model with the rain gauge precipitation data to generate past, spatially distributed precipitation estimates at the NEXRAD grid locations. For the testing period, the ANN-simulated wet season precipitations in all the NEXRAD grids had a Nash-Sutcliffe efficiency greater than or equal to 0.72 and a mass balance error less than or equal to 14%. The same model performance parameters were 0.65 and 17%, respectively, for the dry season. The MLR model did not perform as well as the ANN model. For the MLR model, the wet season mass balance error ranged from 13–48%, whereas the dry season mass balance error ranged from 0.1–36% on the testing data set. An uncalibrated soil and water assessment tool model was used to assess the improvements in stream flow simulations with the ANN-simulated spatially distributed precipitation data. The stream flow simulations using ANN-generated, spatially distributed precipitations were closer to the observed stream flows relative to stream flows generated using the rain gauge precipitations. Overall, the results suggest that the method developed in this study can be used to generate past, spatially distributed precipitations at NEXRAD grid locations.

This paper presents a new forecasting methodology that uses self-learning cellular automata (SLCA) for including variables that consider the spatial dynamics of the mass of precipitation in a radar forecast model. Because the meteorological conditions involve nonlinear dynamic behavior, an automatic learning model is used to aid the cellular automata rules (SLCA). The new methodology is applied to the western part of England (Brue river basin) using NIMROD data. The radar information from 1 month of hourly radar measurements is used. Two models, a regression model tree (MT) and an artificial neural network (ANN) model, are used to learn the dynamics of the spatially local effects within the cellular automation (CA) neighboring areas. A spatial correlation (tracking pattern) reference model is built for comparing the first hour of precipitation forecast. Model results show that the SLCA is more accurate than conventional tracking. Furthermore, it appears that this technique can be extended to include other important atmospheric variables in forecasting processes.

Statistical analyses were performed using five years (1996–2000) of Weather Surveillance Radar–1988 Doppler (WSR-88D) weather-radar rainfall estimates that were produced for the Global-Continental Scale International Project over the entire Mississippi River basin. The project radar rainfall estimates were adjusted using a $Z$-$R$ relation optimized to improve performance at the time scale of individual storms. The accuracy of radar-rainfall estimates were analyzed during the warm season considering a number of factors, including number of overlapping radars, distance from gauge to nearest radar, gauge elevation, and the geographic location of the radar. Results were used to identify optimal radar-rainfall estimation areas (ORREAs) within the Mississippi River basin with high correlation between storm-total radar and rain gauge rainfall. Additionally, estimation of the bias between storm-total radar-rainfall and rain gauge rainfall accumulations in areas without gauges was assessed to identify the appropriateness of applying bias adjustments derived from gauge data at points away from those rain gauges.

Recent advances in hydrologic models have depended on the use of radar rainfall input in physically based, fully distributed models. Previous research conducted for case studies near Houston, Texas, have focused on the use of radar rainfall for large storm events such as tropical storms and hurricanes. A fully distributed model, Vflo, was used to model streamflow during small storm events in the Cypress Creek watershed, near Houston, Texas. Two events were simulated both with rain gauge—corrected radar data and exclusively with rain gauges. Then a third event was modeled exclusively with rain gauge data to assess model performance with only rain gauge data. The modeled streamflow was then compared to the USGS observed streamflow, using peak streamflow, time to peak, and volume streamflow, to evaluate the model performance between radar and rain gauge input. A comparison of the models for the events shows that the radar input results better match the observed streamflow for the streamflow volume and peak streamflow.

The July 2002 floods in South Texas resulted from excessive precipitation caused by a slow moving tropical wave accompanied by an abundance of tropical moisture, which resulted in 12 deaths and nearly $1 billion in damage. South Central Texas is particularly vulnerable to floods because of: (1) its proximity to a moist air source (the Gulf of Mexico); (2) the Balcones Escarpment, which concentrates rainfall runoff; (3) a tendency for synoptic scale features to become cut off and stall over the area; and (4) decaying tropical cyclones stalling over the area. This paper examines the hydrology of the July 2002 floods in three urbanized watersheds in the city of San Antonio and Bexar County, Texas. The physically based, distributed parameter gridded surface subsurface hydrologic analysis hydrologic model was used to simulate the flood over the three watersheds. Various flood control features were included in the simulations. The hydrologic model, driven by the next generation radar multi-sensor precipitation estimator, was able to reasonably simulate runoff in the three watersheds. Differences in the responses of the three watersheds were highlighted with the role of flood-control hydraulic structures.

This paper describes the radar-based estimation of groundwater recharge by means of distributed hydrologic modeling used to estimate components of the hydrologic water balance. Conventional management of water resources has focused on groundwater as a separate constituent of surface water in hydrologic systems, but factors such as precipitation, surface runoff, base flow, evapotranspiration, and water demands determine the change in storage in a stream-aquifer system that varies in both time and space. The temporal and spatial variability of groundwater recharge have been estimated in the eastern outcrop of the Arbuckle-Simpson aquifer, located in Oklahoma. Distributed maps of precipitation from radar were corrected for bias using rain gauges and used as input to a distributed hydrologic model. Distributed grids of infiltration were combined with evapotranspiration to extend groundwater recharge estimates from three years, when streamflow records existed, to 13 years, using archival radar. The results show that better characterization of precipitation and runoff, achieved with bias-corrected radar, produces more reliable estimates of runoff and groundwater recharge than those derived from the use of rain gauge data alone.

On July 17–18, 1996, two mesoscale convective systems (MCSs) passed through northeastern Illinois, causing a record 440-mm total storm rainfall within a 24-h period at Aurora, Illinois, with values exceeding 200 mm throughout a broad area of the region. The storm caused flooding with a return period larger than 100 years at different USGS regional stream gauging locations. The Davenport, Iowa, Weather Surveillance Radar–1988 Doppler (WSR-88D) radar coverage allows high-quality characterization of the storm at fine spatial and temporal scales. Of particular interest is the inter- and intravariability in watershed response to the two pulses of intense rainfall. Spatial distribution of rainfall and the degree of urbanization of the individual basins are the dominant factors determining the magnitude of runoff response. These properties are highly dependent on the extent and history of urbanization. Examination of the annual maximum instantaneous peak discharge and the peaks-over-threshold (POT) time series at three stream gauging stations in Illinois (Blackberry Creek, DuPage River, and Sawmill Creek) over the past 50 years points to the large effect of urbanization on the flood peak distribution in the greater Chicago metropolitan area.

Storm water runoff can significantly affect flooding in urban areas. Flood prediction depends on model structure uncertainties and the accurate determination of rainfall. Three aspects of hydrologic forecasting in real time and hydrologic predictions in off-line modes include the following: (1) distributed model reliability, (2) accuracy of radar-derived rainfall, and (3) scaling of basin input and response. The existing flood alert system (FAS) that is operational for Brays Bayou in Houston, Texas, forms the basis for testing the relative magnitudes of these effects on prediction accuracy. The importance of gauge-corrected radar input was demonstrated through a probabilistic approach and by comparison to events with streamflow observations. The difference in discharge, called dispersion, obtained from corrected and uncorrected radar input scales with drainage area, but at a nonlinear rate, and it differs from storm to storm. An additional comparison was made between the existing flood alert system’s kinematic wave model, Vflo, and the full dynamic wave model, Hydrologic Engineering Center’s River Analysis System (HEC-RAS). Both models showed similar scaling with radar bias correction. Considering that random errors in rainfall rates measured by radar should cancel out over large areas, the decline in forecast skill measured by the critical success index (CSI) was not intuitive. Both empirical observations and the perturbation experiment confirm that predictability decreased with increased drainage area. This article shows the benefit of accurate radar rainfall, but that predictability does not scale linearly with drainage area.

Contemporary sewer system operation heavily relies on urban drainage model results. In a case study near Leuven, Belgium, this research investigates the potential of a cost-effective local area weather radar (LAWR) for providing rainfall input to sewer models. Before rainfall estimation was possible, the radar was calibrated using tipping bucket rain gauge measurements. Various calibration methods were investigated, and a nonlinearly regressed power law function taking range to the radar into account was most suitable in this case study. Rainfall estimations could be slightly improved by rendering the calibration dependent on radar output. Fifty-seven events were simulated using the InfoWorks CS modeling software, and the effect of radar input on model results was investigated. There were large quantitative deviations, but the qualitative course of time series observed in conduits was reproduced well. Although certain events show a higher correspondence between measured and simulated time series when using radar-driven input, the gauge-driven input generally outperforms the former in this case study and the accompanying model.

The Dutch weather radars in De Bilt and Den Helder have only limited coverage in the Dutch-German border area in the northeastern part of the Netherlands, whereas the German radar in Emden is just across the border. For many years, local water authorities have been looking for a better quantitative precipitation estimate (QPE) for this region. Recently, a German water-management consultancy and a Dutch one have jointly taken up the challenge to develop a completely new precipitation radar composite for this part of the Netherlands. The new composite uses the basic polar radar products (volume data) of the two national weather bureaus. The composite should be able to meet precipitation-information requirements of operational water managers; in other words, it should be able to provide QPE in real time. The present case study of an interesting rainfall event demonstrates the capabilities of the new composite tool. The rainfall event was used to evaluate various filtering and correction algorithms. QPE results were verified against independent rain-gauge data. On average, the results of the new composite were found to be similar to the Dutch weather bureau’s QPE for the entire Netherlands. However, the new composite yielded a much better QPE for the northeastern part of the country, as a result of the addition of the Emden radar data. The algorithms developed are ready to be applied in operational water management by water boards and municipalities in the northeastern part of the Netherlands.

Reduction of uncertainty in hydrologic predictions is important for water resources development and management, especially in ungauged basins. The objective of this study was to predict rainfall and runoff in an ungauged basin located in a mountainous region of northern Thailand by using a weather radar to estimate rainfall and a modeling approach to simulate rainfall-runoff processes. The calibrated $Z_e\text{-}R$ relationship ($Z_e=18R^{1.45}$) was used to estimate rainfall over a gauged basin and an assumed ungauged basin. The runoff simulations in the gauged and ungauged basins were conducted using the quasi-distributed hydrological model HEC-HMS. The simulation results in the gauged basin, for both calibration and verification, were in agreement with the observed data. Model parameters in the assumed ungauged basin were estimated using transposition and regionalization techniques. The predicted direct runoff hydrograph in the assumed ungauged basin using radar measured rainfall and the regionalization technique matches well with the measurements at the basin outlet. Based on the study results, the radar rainfall estimates and regionalization of the hydrologic model parameters are promising alternatives for improved hydrologic predictions in ungauged basins.