Geoscience Reference
In-Depth Information
weather or some sort of severity index, but does not necessarily add value to the
identification and tracking of the cells. Rigo et al. (2010) used both 2D and 3D radar
products and total lightning data to identify and track convective storms. A storm track is
defined as a time series of cell positions. Assigning cells to tracks is the most complex aspect
of these algorithms.
All cell tracking algorithms have to deal with cell initiation, mergers, splits, and
terminations - the hatches, matches, and dispatches as it were - and this is often the point of
differentiation between the various approaches. Errors in assigning the correct cell to a track
are a major cause of error when estimating the cell velocity. The size of the object depends
on the threshold that has been selected. Therefore, the predictability of the object decreases
as the threshold is increased since the lifetime of a cell is related to its size. Using a high
threshold to define the cell will make it more difficult to assign a cell to a track. This will
increase the errors when estimating the track velocity. Using a low threshold will increase
the longevity of the tracks, but will tend to limit the ability to forecast the location of the
most severe cells within the storm. The concept of an “object” becomes less useful as the
precipitation becomes more widespread. At some point (depending on the skill of the
tracking algorithm), cell tracking algorithms fail to provide useful forecasts.
TITAN (Dixon & Weiner, 1993) and SCIT (Johnson et al., 1998) are good examples of what
can be achieved in the object-tracking paradigm. Both TITAN and SCIT use the three-
dimensional radar reflectivity data to identify a convective object that is defined by a
reflectivity threshold. In TITAN, the current objects are linked to past objects through
combinatorial optimization. This minimizes the total advection and change in cell volume
between the previous and current time steps. Many other cell tracking algorithms, SCIT
for example, assign the cell that is closest to the forecast location of an active track. Han et
al. (2009) evaluated several extensions to TITAN including improvements in assigning
cells to tracks and using TREC motion vectors to advect the cells. In assigning a cell to a
track, they found that the most significant improvements were due to adding a
requirement that the forecast cell from the track at the previous time step must overlap
with the current cell.
4.1.2 Field-based advection
Field tracking algorithms generally divide a Cartesian grid of radar reflectivity or rain rate
into a number of tiles and then find the advection of the tile that maximizes the cross
correlation (or some other measure of similarity) between successive time steps in the data.
The mean advection vector for each tile containing rain is then calculated by applying some
form of constraint to minimize the divergence of the resulting vectors.
A number of the current field tracking-based nowcasting algorithms use COTREC (Li et al.,
1995) as the basis for deriving the advection vectors. Examples include the system that has
been developed at the Czech Hydrometeorological Institute (Novak, 2007), the Hong Kong
Observatory system, SWIRLS (Li et al., 2000), and the system implemented at the
Guangdong Meteorological Observatory system (Liang et al., 2010). Liang et al. (2010)
determined that the optimum size of the tile was 30 km. Li et al. (2000) evaluated the
performance of an advection scheme on a 93 x 93 grid using a 19 pixel tile: this equates to
20 km on their 256 km x 256 km domain.
Search WWH ::




Custom Search