Geoscience Reference
In-Depth Information
In summary then, the accuracy of a nowcast depends on the accuracy of the initial radar
derived rainfall field, the degree of spatial organization of the rain, the rain rate, and
forecast lead time. Also, it is likely to be higher in the middle of the storm (in both space and
time) than at the edges.
4.3 Probabilistic techniques
4.3.1 Justification
Given the magnitude of the errors in a 30 minute precipitation nowcast, it is reasonable to
adopt a probabilistic approach to nowcasting and attempt to convey to the users the
uncertainty that is associated with a particular weather situation. As explained earlier,
within the extrapolation nowcast framework, errors can be categorized into those
attributable to the radar observations and processing, inaccuracies in the field of motion
used to advect the observations, and errors arising from assumptions made about the
Lagrangian evolution of the advected precipitation field.
4.3.2 Methods of handling uncertainties
A number of techniques have been developed for modelling nowcast errors with a view to
producing probabilistic precipitation nowcast products. One of the simplest entails time-
lagging a consecutive series of deterministic nowcasts using techniques similar to those
demonstrated in a NWP post-processing context (Mittermaier, 2007). Each member of the
time-lagged ensemble is assigned a weight which is a function of lead time. SWIRLS
generates probabilistic nowcasts using this approach (Wang et al., 2009).
Another approach relies on the assumption that errors in the diagnosed advection velocity
field predominate. Consequently, the probability of exceeding a chosen precipitation
threshold at a given location can be derived from the distribution of precipitation in a
neighbourhood surrounding the forecast location. The neighbourhood size increases with
lead time to reflect to the growth in advection errors (Andersson & Ivarsson, 1991; Schmid et
al., 2000).
Germann and Zawadzki (2004) compared four methods of generating probabilistic
precipitation nowcasts based upon radar extrapolation. They concluded that the most skilful
method was one based upon the local Lagrangian technique. Essentially, this produces an
advection forecast using a semi-Lagrangian backward advection scheme and then uses the
probability distribution of forecast rain rates in some search area centred on a pixel to
calculate the probability of exceeding a threshold at that location. The size of the search area
increases with lead time to reflect the increasing forecast uncertainty. This approach has
since been exploited by others, for example Megenhardt et al
.
(2004), and more recently,
Kober et al. (2011).
Other authors have focused their attentions on modelling errors using stochastic space-time
models. Pegram and Clothier (2001) used a power law model to filter Gaussian distributed
random numbers to generate stochastic realizations of radar precipitation fields in their
String of Beads model (SBM). Noise generation techniques similar to these were combined
with a stochastic model of extrapolation velocity errors in the Short Term Ensemble
Prediction System (STEPS; Bowler et al., 2006) to produce operational precipitation nowcasts
