Geoscience Reference
In-Depth Information
quantifying uncertainties in phase as well as amplitude. In STEPS, the noise serves several
purposes: it enables ensembles of equally likely nowcast solutions to be generated by
perturbing predicted features as they lose skill; it also downscales an NWP forecast,
injecting variance at scales lacking power (variance) relative to the radar.
4.3.3 Treatment of observation errors
Uncertainties in nowcasts of precipitation also derive from errors in the radar observations
and processing. Austin (1987) categorized radar errors into physical biases, measurement
biases and random sampling errors. Historically, much effort has been invested in
improving deterministic estimates of precipitation accumulation at the surface by correcting
physical (e.g. ground clutter and beam blockage) and measurement (e.g. Z-R conversion)
biases. However, more recently, a growing number of researchers have focused their
attentions on the treatment of random sampling errors and how these can be utilized within
stochastic, integrated system frameworks to improve hydro-meteorological nowcasting.
Two main approaches to the modelling of random sampling errors in QPE have been
described in the literature: one entails a statistical description of the difference between the
radar estimates and a reference (e.g. Ciach et al., 2007; Llort et al., 2008; Germann et al.,
2009); a second involves modelling the characteristics of individual sources of error (e.g.
Jordan et al., 2003; Lee & Zawadzki, 2005a, 2005b, 2006; Lee, 2006; Lee et al., 2007). The
challenge with the first approach is the need for a reference field: this is usually derived
from a dense network of rain gauges. The difficulty with the second approach is that the
true error structure of QPEs can vary significantly depending on the meteorological
conditions and is therefore largely unknowable.
Germann et al. (2009) describe a radar ensemble generator using LU decomposition
(factorization) of the radar-gauge error covariance matrix to derive an ensemble of
precipitation fields. Each ensemble member is the sum of the bias corrected,
deterministically derived radar precipitation field and a stochastic perturbation representing
the random error. The stochastic term is generated such that it preserves the correct space-
time error covariances. The authors present the results of the coupling of a real-time
implementation of the radar ensemble generator with a semi-distributed hydrological
model.
Norman et
al
. (2010) implemented several radar ensemble generators and compared their
performance on a selection of case study events using rain gauges. An implementation of
the Germann et al. (2009) scheme was found to be marginally superior to one comprising
separate models of Z-R (Lee et al., 2007) and VPR (Jordan et al., 2003) errors. Pierce et al.
(2011) integrated these two ensemble generators to produce ensembles of radar-based
analyses of surface precipitation rate for input to STEPS. They evaluated the impact of these
ensembles on the performance of STEPS ensemble precipitation nowcasts. Verification
results demonstrated that accounting for QPE errors improved the ensemble spread-skill
relationship in the first hour of the nowcasts.
One alternative to the stochastic QPE and QPN schemes described above is the use of
historical analogues. Panziera et al. (2011) describe an analogue-based heuristic tool for
nowcasting orographically forced precipitation. The system known as Nowcasting of
Orographic Rainfall by means of Analogues, exploits the strong correlation between
