better performance over longer timescales as they dynamically resolve the largescale atmospheric processes. Radar-based precipitation forecasting is known as
precipitation nowcasting. Nowcasting models are based on the extrapolation of
radar rainfall scans to track the motion of precipitation cells with a forecasting
lead time of a few hours. Radar-based precipitation nowcasting has a higher performance than NWP forecasts for the first few hours of the forecasts, but NWP forecasts
have a better performance at longer forecasting lead times. Radar nowcasting can be
however very valuable for flash flood forecasting in urban areas or hydrological
forecasting in large catchments.
Nowcasting aims to tracking the movement of storms to extrapolate the radar
rainfall field into the future with a forecasting lead time of a few hours [131–
133]. Radar-based nowcasting methods include Tracking radar echoes by correlation
(TREC and COTREC methods), tracking the centroids of rain cells, use of wind
fields from NWP forecasts to advect the precipitation field, the Variational Echo
Tracking (VET) method and optical flow techniques. In the TREC method, the
advection vector for each block is determined using a correlation method, but gaps
appear where blocks diverge. The COTREC algorithm minimizing the divergence of
the velocities of adjacent blocks to avoid the issues observed in the TREC method.
The Variational Echo Tracking (VET) method is similar to the COTREC method,
but also takes into account radial velocities from Doppler radar. Bowler et al. [131]
developed a method to compute the advection field using optical flow techniques
that has shown better performance in cases involving embedded convection. This
method assumes that features in a sequence of radar scans only change shape, but do
not change in size or intensity. Radar echo-tracking techniques assume Lagrangian
persistence along the direction and speed of movement of storms. For forecasts
beyond 5 min, nowcasting systems that rely on Lagrangian persistence are subjected
to errors because storms evolve and change direction [87, 88]. Larger scale precipitation features have higher Lagrangian temporal autocorrelation and may persist
longer than smaller scale features in the forecast. In fact, the predictability of
precipitation systems depends on the size of their scale, with small-scale features
being less persistent and less predictable [134]. Also, smaller scale precipitation
features evolve faster than larger-scale features. For instance, isolated thunderstorms
smaller in size (e.g. 10 km) will undergo considerable evolution over a short-time
scale (e.g. 30 min) [135], whereas larger storm systems (e.g. 100 km in size) will
evolve over several hours. This evolution is difficult to forecast and consequently the
skill of the forecast decreases rapidly with lead time. Large-scale precipitation
patterns can be forecasted by extrapolation whereas locally generated precipitation
(e.g. due to convection) is less predictable. Convective storm initiation is therefore
an important area of research in radar nowcasting.
Radar-based extrapolation techniques generally do not account for precipitation
growth and decay [136]. However, precipitation systems evolve with time, and
therefore growth and decay of precipitation processes become important resulting
in a decrease of forecasting skill with lead time. In fact, uncertainties in radar-based
forecasts are due to errors in the original radar rainfall field, uncertainties due to the
temporal evolution of the velocity field (i.e. assumption that the rainfall trajectories
Precipitation Measurement with Weather Radars
249
precipitation nowcasting. Nowcasting models are based on the extrapolation of
radar rainfall scans to track the motion of precipitation cells with a forecasting
lead time of a few hours. Radar-based precipitation nowcasting has a higher performance than NWP forecasts for the first few hours of the forecasts, but NWP forecasts
have a better performance at longer forecasting lead times. Radar nowcasting can be
however very valuable for flash flood forecasting in urban areas or hydrological
forecasting in large catchments.
Nowcasting aims to tracking the movement of storms to extrapolate the radar
rainfall field into the future with a forecasting lead time of a few hours [131–
133]. Radar-based nowcasting methods include Tracking radar echoes by correlation
(TREC and COTREC methods), tracking the centroids of rain cells, use of wind
fields from NWP forecasts to advect the precipitation field, the Variational Echo
Tracking (VET) method and optical flow techniques. In the TREC method, the
advection vector for each block is determined using a correlation method, but gaps
appear where blocks diverge. The COTREC algorithm minimizing the divergence of
the velocities of adjacent blocks to avoid the issues observed in the TREC method.
The Variational Echo Tracking (VET) method is similar to the COTREC method,
but also takes into account radial velocities from Doppler radar. Bowler et al. [131]
developed a method to compute the advection field using optical flow techniques
that has shown better performance in cases involving embedded convection. This
method assumes that features in a sequence of radar scans only change shape, but do
not change in size or intensity. Radar echo-tracking techniques assume Lagrangian
persistence along the direction and speed of movement of storms. For forecasts
beyond 5 min, nowcasting systems that rely on Lagrangian persistence are subjected
to errors because storms evolve and change direction [87, 88]. Larger scale precipitation features have higher Lagrangian temporal autocorrelation and may persist
longer than smaller scale features in the forecast. In fact, the predictability of
precipitation systems depends on the size of their scale, with small-scale features
being less persistent and less predictable [134]. Also, smaller scale precipitation
features evolve faster than larger-scale features. For instance, isolated thunderstorms
smaller in size (e.g. 10 km) will undergo considerable evolution over a short-time
scale (e.g. 30 min) [135], whereas larger storm systems (e.g. 100 km in size) will
evolve over several hours. This evolution is difficult to forecast and consequently the
skill of the forecast decreases rapidly with lead time. Large-scale precipitation
patterns can be forecasted by extrapolation whereas locally generated precipitation
(e.g. due to convection) is less predictable. Convective storm initiation is therefore
an important area of research in radar nowcasting.
Radar-based extrapolation techniques generally do not account for precipitation
growth and decay [136]. However, precipitation systems evolve with time, and
therefore growth and decay of precipitation processes become important resulting
in a decrease of forecasting skill with lead time. In fact, uncertainties in radar-based
forecasts are due to errors in the original radar rainfall field, uncertainties due to the
temporal evolution of the velocity field (i.e. assumption that the rainfall trajectories
Precipitation Measurement with Weather Radars
249