Geoscience Reference
In-Depth Information
errors in the location of rainfall. Fabry and Seed (2009) showed that forecasts of high rain
rates were generally over-predictions and that the performance of advection forecasts in
the recent past is not a good predictor of future performance. The best predictors were
found to be raining fraction and the rate of change in mean areal precipitation over the
forecast domain.
Germann and Zawadzki (2002) demonstrated that filtering the rainfall analysis field with a
64 km, low-pass filter increased Lagrangian life times by between 40 and 60 minutes,
depending on the extent to which small scale features are embedded in larger-scale rain
areas. Germann et al. (2006) state that the upper bound for an advection-based nowcasting
system that does not include growth and dissipation of rainfall is about six hours. The
typical lifetime of a storm is closely related to the scale of the storm, and is often represented
as a power law of the scale (e.g. Marsan et al., 1996; Schertzer et al., 1997; Seed et al., 1999).
Therefore, some nowcasting systems improve the accuracy (in the RMS error sense) of their
predictions by progressively smoothing out the small scale features present in the analysis
field (e.g. Seed, 2003; Turner et al., 2004). This removes features from the nowcast that are
essentially unpredictable.
An alternative way of handling the perishability of the fine scale components in advected
precipitation fields is to model them stochastically. This approach will be discussed in
section 4.3.
4.2.3 Performance of nowcasting algorithms
The Critical Success Index is often used to report the accuracy of nowcasting algorithms
presented in the literature. Ruzanski et al. (2011) found a CSI of approximately 0.5 after
10 minutes at a spatial resolution of 0.5 km. Liang et al. (2010) calculated a CSI of
approximately 0.35 after 60 minutes for echoes in the 15-45 dBZ range at 2 km resolution.
Berenguer et al. (2011) report a CSI for 60 minute forecasts of reflectivity (dBZ) at 1 km
resolution of approximately 0.5 for widespread rainfall, and in the range of 0.1 to 0.3 for
isolated convection. Poli et al. (2008) discovered that the CSI was generally low at the start
and end of a storm, reaching a peak of around 0.4 for 1 km resolution T+60 minute forecasts
of reflectivity greater than 30 dBZ.
Nine nowcasting systems were implemented for the Sydney 2000 Forecast Demonstration
Project (Ebert et al., 2004) and eight nowcasting systems participated in the Beijing 2008
Olympics' Forecast Demonstration Project (Wang et al., 2009; Wilson et al., 2010). Wang et
al. (2009) demonstrated that the overall performance of the nowcasting systems had
improved during the years from 2000 to 2008. They showed that the maximum CSI for
forecasts of hourly precipitation accumulation greater than 1 mm/h increased from 0.2 in
2000 to 0.45 in 2008, although the maximum CSI for rain greater than 10 mm/h was still
only 0.15.
Lee et al. (2009) found that the CSI decreased with increasing rain rate and forecast lead
time: the CSI for 60 minute rainfall forecasts decreased from 0.60 for 0.1 mm/h to 0.2 for
10 mm/h rain rates. Ebert et al. (2004) reported that the CSI for rain greater than 20 mm/h is
essentially zero. This implies that the use of nowcasting techniques to predict the precise
location of extreme rain for flash flood warning may not be viable.
