Correcting bias in radar Z–R relationships due to uncertainty in point rain gauge networks
Introduction
Rainfall is one of the most important inputs to hydrological analysis and modelling. Any error in this input will propagate through the model and will introduce uncertainty in subsequent predictions (Hossain et al., 2004, Morin et al., 2005, Pessoa et al., 1993, Sharif et al., 2002, Vieux and Bedient, 1998, Vivoni et al., 2007, Zhu et al., 2013). One source of uncertainty and possible errors is the temporal and spatial variability of rainfall (AghaKouchak et al., 2010, Faurès et al., 1995, Goodrich et al., 1995, Shah et al., 1996). Therefore, accurately measuring and predicting the spatial and temporal distribution of rainfall is an important challenge in hydrology.
Remote sensing provides an increasingly important source for spatial estimation of precipitation. Using measurements of reflectivity, weather radars can produce estimates of precipitation over large geographic areas, and therefore provide information about rainfall patterns at high temporal and spatial resolution. As a result, the potential applications of weather radars for hydrologic modelling have been the subject of extensive research (Bonnifait et al., 2009, Cole and Moore, 2008, Cole and Moore, 2009, Collier, 2009, Keblouti et al., 2013, Looper and Vieux, 2012, Viviroli et al., 2009).
Traditionally, radar reflectivities Z (mm6/m3) are converted to ground rainfall rates R (mm/h) using a power–law relationship (Z = ARb) known as the Z–R relationship. The relationship between radar reflectivity and rainfall rate depends on the nature of rainfall, and in particular the drop size distribution (DSD) (Chumchean et al., 2006a, Chumchean et al., 2008, Islam et al., 2012, Uijlenhoet and Pomeroy, 2001). One way to calibrate the Z–R relationship is through the use of disdrometers which can explicitly measure the DSD and therefore the relationship with the radar reflectivity (Ochou et al., 2011, Verrier et al., 2013). However in many parts of the world, including Australia, gauge based calibration of the Z–R relationships is routinely used (Bringi et al., 2011, Rendon et al., 2013). Furthermore, in developing countries where disdrometer data is not available and even sub-daily rainfall measurements are infrequent, daily rainfall gauges are the sole source of information on ground rainfall estimates (Mapiam et al., 2009). In these cases gauge densities are also likely to be very low compared to developed countries such as the United States of America and Europe. Thus errors in the recorded gauge rainfall can bias the Z–R relationship, resulting in errors in the radar rainfall estimates. A biased Z–R relationship may be due to an inadequate consideration of the spatial subgrid scale gauge rainfall variability and its representation through the handful of gauges that are available for use (Ciach et al., 2007, Jordan et al., 2003) and therefore methods that take account of gauge uncertainty can be very useful.
The performance of the radar rainfall estimates is evaluated by comparing them with point gauge measurement at the ground (Anagnostou et al., 1999, Habib and Krajewski, 2002). These point gauge measurements are therefore considered as the “ground truth” (Lebel and Amani, 1999, Wolff et al., 2005). However, there is a spatial discrepancy between the two data sources, since radar rainfall estimates are provided as spatial averages with a resolution 1–4 km2 (Krajewski and Smith, 2002). In contrast, whilst rain gauges measure precipitation at fixed point locations, these are often too sparse to properly represent the spatial variability and areal structure of rainfall (Morrissey et al., 1995, Rodríguez-Iturbe and Mejía, 1974). Furthermore, the uncertainty associated with spatial averages over a single radar grid cell is considerable, and is a function of the number of gauges that are within the cell being considered. In addition, rain gauge measurements can contain a variety of errors including wind effects, evaporation and mechanical tipping bucket errors (Groisman and Legates, 1994). An important question is whether the parameters estimated for the Z–R relationship would remain the same if consideration was given to the nature of the errors associated with spatial averaging of the gauge rainfall over the radar grid scale. We argue that if these errors are significant (and vary with space as they will if the rainfall and the gauge density are not uniform across the network), the estimated parameters will have a bias with respect to the true parameters that ought to be used.
This research aims to address whether the A parameter in the radar Z–R relationship needs to change to account for uncertainty in the point gauge rainfall network. It has been reported that the parameter A carries most of the variability in the Z–R relationship, whereas the uncertainty in b can be seen as second-order (Chumchean et al., 2003, Chumchean et al., 2006b, Steiner et al., 1999). Marshall and Palmer (1948) proposed the power law relation, Z = 200 R1.6 with specified values for A equal to 200 and b equal to 1.6. Since then, several studies have been conducted to find appropriate A and b parameter values in different settings. Given the need for consistent estimates of rainfall, for operational purposes the Australian Bureau of Meteorology (BOM) specifies a fixed value for b equal to 1.53 for most of the weather radars that form its rainfall measuring network. Fixing b for operational networks has also previously been used by other researchers (Steiner and Smith, 2004, Verrier et al., 2013). We use the Simulation Extrapolation method (SIMEX) (Cook and Stefanski, 1994) to investigate the significance of point gauge uncertainty on the Z–R relationship parameter A. The SIMEX method involves developing a relationship between the gauge uncertainty distribution and input gauge rainfall, which can then be used to assess the bias in the parameters of the Z–R relationship.
The paper is organised in seven sections. Section 2 outlines the logic behind the SIMEX approach. Section 3 describes the radar and gauge data used in this study, whilst Section 4 presents the development of an error model for the lowest radar pixel resolution (1 km2). The application of SIMEX method on radar Z–R relationship is presented in Section 5 followed by results and discussion in Section 6. The main findings are summarised in Section 7.
Section snippets
Simulation Extrapolation (SIMEX)
SIMEX is a method for parameter estimation that attempts to ascertain model parameters taking into account the error distribution associated with each predictor variable. It estimates parameter values that should have resulted if the covariates were error-free. The general idea behind the method is that if the error in the predictors causes bias in the parameter estimates, then adding more error should cause the parameter estimates to become even more biased (Benoit et al., 2009). A
Data
In this research, the reflectivity data were obtained from the Australia Bureau of Meteorology for the Terrey Hills radar (Sydney, Australia) during the period from November 2009 to December 2011. The Terrey Hills radar is an S-band Doppler radar with 6 min temporal resolution and 1 km spatial resolution. The radar covers a region of 256 km by 256 km extent, with bandwidth of 1° and wavelength of 10.7 cm. The climatological freezing levels in Sydney are about 2.5 km (Chumchean et al., 2003).
Error model
The rain gauge data cannot be realistically assumed as error-free when calibrating the Z–R relationship. The SIMEX technique allows to account for the various sources of errors in the rainfall totals and to remove any corresponding bias from our estimate of the parameter A. As outlined above, there are two main sources of error in the rain gauge observations with respect to the radar–rainfall relationship. The first one is due to recording errors in the tipping bucket gauge. The second one is
Use of SIMEX to improve the radar Z–R relationship
This section presents the algorithm that was used to estimate the unbiased A parameter using the SIMEX method. Suppose, instead of observing the gauge rainfall R, that we actually observe the erroneous rainfall W where, W = R × δ and δ is the multiplicative error distribution given by Eq. (5) (assuming . Here, σu2 is the error variance of the point gauge rainfall uncertainty obtained from the error model and is the multiple of error variance. In the simulation step, additional
SIMEX results
We have combined gauge measurement uncertainty (σg) and spatial variability uncertainty (σcv) to estimate point gauge uncertainty (σu) (Fig. 4c). We have found that spatial variability uncertainty dominates (Fig. 4b) compared to the gauge measurement uncertainty (Fig. 4a). In addition, the same number of gauges in a larger area gives higher spatial variability uncertainty in the area of interest (Fig. 3).
The results of the SIMEX analysis are shown in Fig. 5. In this case the naive estimate is
Conclusion
Accurate spatial rainfall data is essential to get the best output from hydrological models. Generally, radar rainfall estimation accuracy is evaluated by comparing point gauge rainfall without considering gauge uncertainty. This paper makes a unique contribution as we have identified that point gauge uncertainty is a factor that may affect the radar rainfall estimation process. In this paper we have considered how this uncertainty impacts the Z–R relationship. We have identified that there are
Acknowledgements
The authors gratefully acknowledge the Australian Bureau of Meteorology for providing radar and rain gauge data for this study. The comments of two anonymous reviewers have greatly improved the presentation of the work. Authors acknowledge the Australian Research Council for partial funding for this work.
References (69)
Distributed hydrologic and hydraulic modelling with radar rainfall input: reconstruction of the 8–9 September 2002 catastrophic flood event in the Gard region, France
Adv. Water Resour.
(2009)- et al.
Mitigating parameter bias in hydrological modelling due to uncertainty in covariates
J. Hydrol.
(2007) - et al.
Radar rainfall error variance and its impact on radar rainfall calibration
Phys. Chem. Earth
(2003) - et al.
Correcting of real-time radar rainfall bias using a Kalman filtering approach
J. Hydrol.
(2006) - et al.
An operational approach for classifying storms in real-time radar rainfall estimation
J. Hydrol.
(2008) - et al.
Hydrological modelling using rain gauge-and radar-based estimators of areal rainfall
J. Hydrol.
(2008) - et al.
Distributed hydrological modelling using weather radar in gauged and ungauged basins
Adv. Water Resour.
(2009) - et al.
Impact of small-scale spatial rainfall variability on runoff modeling
J. Hydrol.
(1995) - et al.
Measurement and analysis of small-scale convective storm rainfall variability
J. Hydrol.
(1995) - et al.
Characteristics of raindrop spectra as normalized gamma distribution from a Joss–Waldvogel disdrometer
Atmosph. Res.
(2012)
Spatial variability of rainfall: variations within a single radar pixel
Atmosph. Res.
Radar hydrology: rainfall estimation
Adv. Water Resour.
Experimental and numerical studies of small-scale rainfall measurements and variability
Water Sci. Technol.
An assessment of distributed flash flood forecasting accuracy using radar and rain gauge input for a physics-based distributed hydrologic model
J. Hydrol.
Rainfall uncertainty in hydrological modelling: an evaluation of multiplicative error models
J. Hydrol.
Quantification of the spatial variability of rainfall based on a dense network of rain gauges
Atmosph. Res.
Modelling the effects of spatial variability in rainfall on catchment response. 2. Experiments with distributed and lumped models
J. Hydrol.
Empirically-based modeling of spatial sampling uncertainties associated with rainfall measurements by rain gauges
Adv. Water Resour.
An introduction to the hydrological modelling system PREVAH and its pre-and post-processing-tools
Environ. Model. Software
Errors and fluctuations of rain gauge estimates of areal rainfall
J. Hydrol.
Modeling radar rainfall estimation uncertainties: random error model
J. Hydrol. Eng.
Uncertainty quantification of mean-areal radar–rainfall estimates
J. Atmosph. Oceanic Technol.
Treating words as data with error: uncertainty in text statements of policy positions
Am. J. Political Sci.
Rainfall estimation with an operational polarimetric C-band radar in the United Kingdom: comparison with a gauge network and error analysis
J. Hydrometeorol.
Asymptotics for the SIMEX estimator in nonlinear measurement error models
J. Am. Statist. Assoc.
Nonparametric regression in the presence of measurement error
Biometrika
A simulation based approach for representation of rainfall uncertainty in conceptual rainfall runoff models
Hydrol. Res. Lett.
Application of scaling in radar reflectivity for correcting range-dependent bias in climatological radar rainfall estimates
J. Atmosph. Oceanic Technol.
An integrated approach to error correction for real-time radar–rainfall estimation
J. Atmosph. Oceanic Technol.
Local random errors in tipping-bucket rain gauge measurements
J. Atmosph. Oceanic Technol.
Product-error-driven uncertainty model for probabilistic quantitative precipitation estimation with NEXRAD data
J. Hydrometeorol.
On the propagation of uncertainty in weather radar estimates of rainfall through hydrological models
Meteorol. Appl.
Simulation–extrapolation estimation in parametric measurement error models
J. Am. Statist. Assoc.
The accuracy of United States precipitation data
Bull. Am. Meteorol. Soc.
Cited by (30)
A new approach to weather radar adjustment for heavy rainfall events using ANFIS-PSO
2023, Journal of HydrologySensitivity of Australian roof drainage structures to design rainfall variability and climatic change
2019, Building and EnvironmentA Bayesian partial pooling approach to mean field bias correction of weather radar rainfall estimates: Application to Osungsan weather radar in South Korea
2018, Journal of HydrologyCitation Excerpt :In addition to systematic error, there are many different types of errors associated with the radar measurement, including nonlinearity between the echo intensity and the rain drops (Berne and Uijlenhoet, 2005; Kummerow, 1998; Lafont and Guillemet, 2004; Sassi et al., 2014), errors between the precipitation intensity and the radar reflectivity (Anagnostou et al., 1998; Steiner et al., 1999; Villarini and Krajewski, 2010a, 2010b), errors associated with the curvature of the earth’s surface in a remote location from the radar site (Joss et al., 1990; Yoo et al., 2014; Yoo et al., 2010) and range-dependent bias related to systematic bias in space (Borga and Tonelli, 2000; Chumchean et al., 2004; Fulton et al., 1998; Seo et al., 2000; Zhang et al., 2014). A number of approaches to bias correction have been explored (Chumchean et al., 2004; Chumchean et al., 2006; Dai et al., 2015; Goudenhoofdt and Delobbe, 2009; Harrison et al., 2000; Hasan et al., 2014; Rabiei and Haberlandt, 2015; Seo and Breidenbach, 2002a; Thorndahl et al., 2014; Todini, 2001). The gauge-radar (GR) ratio approach has been the most commonly used method for correcting the systematic bias of radar precipitation data.
Comparison of random forests and support vector machine for real-time radar-derived rainfall forecasting
2017, Journal of HydrologyCitation Excerpt :Future work may factor in the uncertainty associated with grid-based radar-derived rainfalls, e.g., uncertainty in point rain gauge networks (Hasan et al., 2014), to improve the proposed models for enhancing the simulated streamflow.