Title: Error Propagation from Radar Rainfall Nowcasting Fields to a FullyDistributed Flood Forecasting Mode
1Error Propagation from Radar Rainfall Nowcasting
Fields to a Fully-Distributed Flood Forecasting
Model
- Enrique R. Vivoni1, Dara Entekhabi2 and Ross N.
Hoffman3 - 1. Department of Earth and Environmental Science
- New Mexico Institute of Mining and Technology
- 2. Department of Civil and Environmental
Engineering Massachusetts Institute of Technology - 3. Atmospheric and Environmental Research, Inc.
ERAD 2006 Conference, Barcelona, Spain
September 21, 2006
2Motivation
Radar nowcasting and distributed watershed
modeling can improve prediction of hydrologic
processes across basin scales.
- Combined Radar QPF-Distributed QFF
- How does rainfall forecast skill translate to
flood forecast skill? - What are the effects of lead time and basin
scale on flood forecast skill? - Does a hydrologic model dampen or amplify
nowcasting errors? Why? - How do errors propagate into the flood
predictions as a function of scale?
Quantitative Precipitation Forecasts (QPFs) using
Radar Nowcasting
Nowcasting of Radar Quantitative Precipitation
Estimate (QPE)
Quantitative Flood Forecasts (QFFs) using
Distributed Hydrologic Modeling
3Combined Rainfall-Flood Forecasting
The distributed QPF and QFF models are combined
using a method denoted as the Interpolation
Forecast Mode.
- Interpolation Forecast Mode
- Multiple QPEs available at a specific time
interval (depending on radar estimation
technique). - Nowcasting QPF generated from available QPEs to
fill in periods with no radar observations. - QPEs Radar Nowcasting QPFs fused according to
lead time prior to forcing input into distributed
model. - Forecast lead time (tL) is varied from 15-min to
3-hr to introduce radar nowcasting errors into
QFF.
Interpolation Forecast Mode
tL
tL
Lead Time
Vivoni et al. (2006)
4STNM Radar Rainfall Nowcasts
Rainfall forecasting using scale-separation
extrapolation allows for predictability in the
space-time distribution of future rain.
Large-scale Features
- STNM Nowcasting Model
- Predictability in rainfall over 0-3 hr over
regional, synoptic scales. - Forecast of space-time rainfall evolution suited
for linear storm events (e.g. squall lines). - Tested over ABRFC over 1998-1999 period using
NEXRAD, WSI data. - Skill is a function of lead time, rainfall
intensity and verification area.
Unfiltered Radar Rainfall
Envelope Motion
MIT Lincoln Lab Storm Tracker Model (STNM)
Van Horne et al. (2006)
Small-scale Features
5Distributed Hydrologic Modeling
TIN-based Real-time Integrated Basin Simulator
(tRIBS) is a fully-distributed model of coupled
hydrologic processes.
- Distributed Hydrologic Modeling
- Coupled vadose and saturated zones with dynamic
water table. - Moisture infiltration waves.
- Soil moisture redistribution.
- Topography-driven lateral fluxes in vadose and
groundwater. - Radiation and energy balance.
- Interception and evaporation.
- Hydrologic and hydraulic routing.
Surface-subsurface hydrologic processes over
complex terrain.
Ivanov et al. (2004a,b)
6Study Area
Radar rainfall over ABRFC used as forcing to
hydrologic model operated over multiple stream
gauges in the Baron Fork, OK.
- NEXRAD-based Rainfall
- WSI (4-km, 15-min) NOWrad
- STNM nowcasting algorithm
- Transformed to UTM 15
- Clipped to Baron Fork basin
- Basin QPFs
- 808, 107 and 65 km2 basins
- 52, 13 and 10 (4 km) radar cells
7Basin Data and Interior Gauges
Soils and vegetation distribution used to
parameterize tRIBS model. Fifteen gauges (range
of A, tC) used for model flood forecasts.
8Hydrometeorological Flood Events
Two major flood events January 4-6, 1998 and
October 5-6, 1998 varied in the basin rainfall
and runoff response.
Oct 98
Jan 98
- Fall Squall Line
- Concentrated rain accumulation.
- Decaying flood wave produced in Dutch Mills.
- Winter Front
- Banded rain accumulation.
- 7-yr flood event at Baron Fork.
Jan 98
Oct 98
BF
DM
Discharge (m3/s)
Discharge (m3/s)
Rainfall (mm/h)
Rainfall (mm/h)
Simulation Hours
Simulation Hours
9Multi-Gauge Model Calibration
January 1998
October 1998
Baron Fork (808 km2)
Dutch Mills (107 km2)
Peacheater Creek (65 km2)
10Rainfall and Runoff Forecasts
Radar nowcasting QPFs and distributed QFFs are
tested in reference to the radar QPE and its
modeled hydrologic response.
January 1998
October 1998
- Multiple QPF and QFF Realizations
- Solid black lines represent QPE Mean Areal
Precipitation (MAP) and Outlet discharge at Baron
Fork. - Thin gray lines are Nowcast QPF MAP and Outlet
discharge for 12 different lead times (tL). - Two events had varying rainfall amounts and
runoff transformations - January Q/P 1.20
- October Q/P 0.24
- January Recurrence 6.75 yr
- October Recurrence 1.43 yr
- January Basin Lag 13.3 hr
- October Basin Lag 15.3 hr
MAP
MAP
Outlet
Outlet
Vivoni et al. (In Press)
11Flood Forecast Skill
Flood forecast skill decreases as a function of
lead time and increases with basin area for the
two storm events.
Lead-Time Dependence
Catchment Scale Dependence
QPE
Increasing Skill
1-hr
Decreasing Skill
2-hr
At 1-hr Lead Time
Vivoni et al. (2006)
12Radar Nowcast Error Propagation
Statistical measures of error propagation show
that nowcasting errors are amplified in the flood
forecast as lead time increases.
- Bias defined as
- where F forecast mean
- O QPE mean
- indicates discharge bias increases more quickly
than rainfall bias. - Mean Absolute Error defined as
- shows that increase in rainfall MAE leads to
higher discharge MAE. - Note the strong impact of the increasing
forecast lead time.
Mean Absolute Error Propagation
Bias Propagation
Slope 1.3 for January 2.6 for
October
Slope 0.099 for January 0.105 for
October
13Error Dependence on Basin Scale
Propagation of radar nowcasting errors is reduced
with increasing catchment scale (area) over range
0.8 to 800 km2.
1-hr Lead Time
2-hr Lead Time
- Bias Ratio defined as
-
- indicates comparable bias for large basins and
large variability in B ratio for small basins. - Mean Absolute Error ratio is
- shows small basins either amplify or dampen
errors, while at large scales errors tend to
cancel out.
Vivoni et al. (In Press)
14Final Remarks
- We have analyzed the propagation of radar
nowcasting errors to distributed flood forecasts
using two forecast models. - The study results reveal
- Increasing the forecast lead time results in
nowcasting errors which are amplified in the
flood forecast at the basin outlet. - Catchment scale controls whether rainfall
forecast errors are strongly amplified or
dampened (in small basins) or effectively
comparable to (in large basins) flood forecast
errors. - Differences in storm characteristics (winter air
mass vs. fall squall line) have a strong effect
on the error propagation characteristics. - To best utilize the distributed nature of the
forecast models, a next step would be utilizing
spatial metrics to assess error propagation from
rainfall to soil moisture fields.
15References
Ivanov, V.Y., Vivoni, E.R., Bras, R.L. and
Entekhabi, D. 2004a. Preserving High-Resolution
Surface and Rainfall Data in Operational-scale
Basin Hydrology A Fully-distributed
Physically-based Approach. Journal of Hydrology.
298(1-4) 80-111. Ivanov, V.Y., Vivoni, E.R.,
Bras, R. L. and Entekhabi, D. 2004b. Catchment
Hydrologic Response with a Fully-distributed
Triangulated Irregular Network Model. Water
Resources Research. 40(11) W11102,
10.1029/2004WR003218. Van Horne, M.P., Vivoni,
E.R., Entekhabi, D., Hoffman, R.N. and Grassotti,
C. 2006. Evaluating the effects of image
filtering in short-term radar rainfall
forecasting for hydrological applications.
Meteorological Applications. 13(3) 289-303.
Vivoni, E.R., Entekhabi, D., Bras, R.L.,
Ivanov, V.Y., Van Horne, M.P., Grassotti, C. and
Hoffman, R.N. 2006. Extending the Predictability
of Hydrometeorological Flood Events using Radar
Rainfall Nowcasting. Journal of Hydrometeorology.
7(4) 660-677. Vivoni, E.R., Entekhabi, D. and
Hoffman, R.N. 2006. Error Propagation from Radar
Rainfall Nowcasting Fields to a Fully-Distributed
Flood Forecasting Model. Journal of Applied
Meteorology and Climatology. (In Press).