1. Prospects for improved hydrological and agricultural drought prediction: The role of precipitation forecasts
Dennis P. Lettenmaier
Department of Civil and Environmental Engineering
University of Washington
Climate Diagnostics and Prediction Workshop
Tallahassee, FL
October 24, 2007

2. Talk Outline The nature of hydrologic prediction
Hydrologic predictability: the role of precipitation forecasts
Predicting drought recovery – the role of initial conditions
Where are the opportunities for improvement?
Towards a national hydrologic prediction strategy

3. Overview: Hydrologic prediction strategy
month 6-12
1-2 years back
start of month 0
forecast ensemble(s)
model spin-up
climatology ensemble
Observations
Initial conditions update
important point(s):
after bias correcting and downscaling the climate model forecasts, the procedure for producing hydrologic forecasts is as follows:
we spin up the hydrologic model to the start of the forecast using observed met. data (from 2 sources: NCDC cooperator stations through 3-4 months before the start of the forecasts, then LDAS 1/8 degree gridded forcings thereafter).
The GSM forecasts comprise 2 sets of ensembles, one for climatology and one for the forecast. The climatology ensemble yields a distribution of the conditions we’ve seen over the period , while the forecast ensemble yields the distribution of the conditions we might see for the next 6 months. Although the climatology ensemble is nominally unbiased against a simulated climatology based on observed met. data (rather than bias-corrected, downscaled GSM met. forcings), we compare the forecast and GSM climatology so that any unforeseen biases (resulting, perhaps, from the downscaling method) occur in both climatology and forecast. Eventually this cautionary step may be eliminated, and we’ll compare directly to the simulated observed climatology.
at the end of the spin-up period and one month before month 1 (out of 6) of the forecasts, we save the hydrologic model state. The state is then used for initializing the forecast runs. Through the first month, the model runs on observed data to the last date possible, then switches to the forecast data. Usually, we process the observed forcings up through the 15th to 25th of this initialization month, then the forecast forcing data carries the run forward for the remaining days in the month, and throughout the following 6 month forecast period. Note, the state files used for the climatology runs correspond to the spin-up associated with the particular year (out of ) from which the climatology ensemble member is drawn.
the spin-up period captures the antecedent land surface hydrologic conditions for the forecast period: in the Columbia basin, the primary field of interest is snow water equivalent.
forecast products are spatial (distributed soil moisture, runoff, snowpack (swe), etc.), and spatial runoff + baseflow is routed to produce streamflow at specific points, the inflow nodes for a management model, perhaps.
Note that for hydrologic forecasts, benchmark is NOT climatology, but rather hydrologic model forced with climatological resampling (ESP)

5. Results of previous seasonal hydrologic predictability studies for continental U.S.
Wood et al, “A retrospective assessment of NCEP climate model-based ensemble hydrologic forecasting in the western United States”, JGR, 2005
Maurer and Lettenmaier, “Predictability of seasonal runoff in the Mississippi River basin”, JGR, 2003
Wood, “An ensemble-based framework for characterizing sources of uncertainty in hydrologic prediction”

6. Maurer and Lettenmaier - Estimation of Runoff Predictability
Mississippi River Basin
Strong gradient in precipitation and runoff
Winter runoff concentrated in SE
High snowmelt runoff in summer
In continuing this examination of persistence, I focus on the Mississippi River basin, which coincides with GCIP region, a project established in 1992 to demonstrate skill in predicting changes in water resources on time scales up to seasonal, annual, and interannual.
By narrowing the region, I also reduce the computational demands. I addition aggregate to ½-degree (about 50km), so there are 1500 grid cells in the basin.
This area has been heavily studies, so there are many resources to draw from, and interest in the results obtained.
World Climate Research Programme, Global Energy and Water Cycle Experiment Continental-scale International Project GCIP

7. Methods for Determining Runoff Predictability
Indices Characterizing Sources of Predictability:
SOI – An index identifying ENSO phase
AO – An index of phase of the Arctic Oscillation
SM – Soil moisture level (normalized)
SWE – Snow water equivalent (normalized)
Varying Lead Times between IC and Forecast Runoff
Only Use Indices in Persistence Mode
Climate
Land
1) Predictability in the context of this study is that due to linear relationships of runoff with the state of the climate and land surface at some earlier date.
2) Variables are introduced in a manner similar to step-wise regression, except instead of the most significant variables first, the best known are introduced first.
3) SM and SWE represent perfect knowledge of these states, and therefore a maximum obtainable knowledge of initial state.
Forecast
Season
DJF
Initialization Dates for DJF Forecast
Dec 1
Mar 1
Jun 1
Sep 1
Lead-0
Lead-4
Lead-3
Lead -2
Lead 1 D J F M A S O N

8. Methods 2 Multiple linear regression used between IC and runoff
Variance explained (r2) indicates level of predictability
Variables introduced in order
of how well indices represent
current knowledge of state:
SOI/AO
SWE SM
Incremental predictability
r2SOI/AO
r2SWE
Runoff
SOI/AO
SWE

9. Methods 3 Test for Significant Predictability (r2) in 2 steps
Local Significance:
Tested at each grid cell
Accounts for temporal autocorrelation
95% confidence level estimated
Field Significance (Livezey and Chen, 1983):
Tests area showing local significance over entire basin
Accounts for limited sample size, spatial correlation in both predictors and predictand
95% confidence for field significance

10. Total Runoff Predictability
Lead,
months
1.5
4.5
7.5
10.5
13.5
Uses all 4 indices to predict runoff
“X”  no field significance
Predictability deteriorates with time
Significance – first local significance is tested, accounting for autocorrelation in both series. Second, field significance is tested to account for spatial correlation in predictor and predictand fields. All at 95%.
Winter – highest predictability in mountains, but very low runoff – combined with some overestimation by using VIC model. Mild, significant predictability in SE, where most runoff is occurring, with some predictability at 3 season lead.
Summer – significant predictability of high mountain runoff at 4 season lead

11. Predictability due to Climate
Predictors currently available
Moderate levels of r2
Greater influence in winter, in area and lead time
Difficulty in long-lead persistence prediction with climate signals
Generally moderate levels of predictability – this is not a physical forcing of the runoff by the index, but is remotely teleconnected, so weaker predictability is understandable.
This is a level of predictability that is currently achievable.
Even with small areas of moderate predictability in summer at leads of 4 seasons, a lot of runoff is at stake
In a moment of weakness I accept a posteriori statistics, and changed one X to light green, since it is within 1% of significant, and would probably pass a 90% test.
Long-lead predictability of winter runoff in southeast is due to climate influences.

12. Predictability Due to Snow
r2 represents incremental increase
Focus at 1 or more season lead is in Rocky mountains
At level of Mississippi basin, predictability limited to 1-2 seasons
Analysis by sub-area could reveal greater predictability
Persistence from snow only possible through snow melt – limited scale
This corresponds to current use of snow forecasts beginning Jan1, as little information is available before this date.
Predictability represents incremental predictability above that from climate signals

13. Predictability due to Soil Moisture
Widespread predictability at 0 lead (1½ month)
Winter Runoff: little predictability where runoff is high
Summer Runoff: limited predictability to 3 seasons
Predictability at greater than 1 season only in west.
Some prediction of JJA runoff the previous Sept 1 in the west!
What is evident from this is that the coincidence of runoff magnitude and predictability may be a better gauge of the importance of these predictabilities.

14. Wood (2002) Reverse ESP

15. Reverse ESP vs ESP – typical results for the western U.S.
Columbia R. Basin
fcst more impt
ICs more impt
Rio Grande R. Basin

16. Wood et al 2005: Retrospective Assessment: Results using GSM
General finding is that NCEP GSM climate forecasts do not add to skill of ESP forecasts, except…
April GSM forecast with respect to climatology (left) and to ESP (right)

17. Wood et al 2005: Retrospective results for ENSO years
Summary: During strong ENSO events, for some river basins (California, Pacific Northwest) runoff forecasts improved with strong-ENSO composite; but Colorado River, upper Rio Grande River basin RO forecasts worsened.
October GSM forecast w.r.t ESP: unconditional (left) and strong-ENSO (right)

18. USGS streamflow gauges that are used in the evaluation of streamflow predictions
Luo, L. and E. F. Wood (2007): Seasonal Hydrologic Prediction with the VIC Hydrologic Model for the Eastern U.S. Journal of Hydrometeorology. In review.

19. RPS: 0~1 with 0 being the perfect forecast
The evaluation of streamflow predictions over selected gauges. The ranked probability score (RPS) for monthly streamflow for the first three months are examined against the offline simulation. The bars are for CFS, CFS+DEMETER and ESP from the left to the right, respectively.
RPS: 0~1 with 0 being the perfect forecast
3 tercels, below normal, normal and above normal with probability of 1/3 each.
Luo, L. and E. F. Wood (2007): Seasonal Hydrologic Prediction with the VIC Hydrologic Model for the Eastern U.S. Journal of Hydrometeorology. In review.

20. Hydrologic prediction during droughts

21. Historic archive: UW SW Monitor

22. Drought recovery – the concept
Real-time applications!
Drought recovery probability described by soil moisture percentiles: (a) Current drought area (based on August 1933); and for different lead times, maps showing the probability (in each grid cell experiencing drought) that soil moisture percentiles will recover. (b) The grid cell-specific recovery probabilities are derived from real-time soil moisture simulations up to the current date, after which simulations are driven by ensemble climate forecasts based on a variety of sources -- e.g., ESP, climate index-conditioned ESP, and the CPC seasonal climate outlooks

23. Initial soil moisture percentiles 2/2006
1 month lead, forecast for 3/2006
3 month lead, forecast for 5/2006
6 month lead, forecast for 8/2006)

24. California-Arizona drought
Feb
Mar
Apr
May
Jun
Jul
Aug

25. Texas drought
Feb
Mar
Apr
May
Jun
Jul
Aug

26. Initial Condition

27. One month lead -- observed

28. Three month lead -- observed

29. Six month lead -- observed

30. Comparison of soil moisture forecasts (ensemble mean of monthly average precipitation expressed as the percentile value within the climatological distribution) from three forecast approaches and observations for summer 1988
Luo, L. and E. F. Wood (2007): Seasonal Hydrologic Prediction with the VIC Hydrologic Model for the Eastern U.S. Journal of Hydrometeorology. In review.

31. Opportunities for improving the initial conditions for seasonal hydrologic prediction
Soil moisture
Surface observations
Remote sensing
Snow
Areal extent
Water equivalent
Other (streamflow?)

33. State soil moisture networks – Illinois (19 stations) and Oklahoma mesonet (~60 stations)

34. NRCS SNOTEL network (~700 stations)

35. UW West-wide forecast system streamflow forecast points
Soil moisture nowcast
Streamflow forecast points

38. local scale weather inputs
MODIS updating of snow covered area
local scale weather inputs
Initial Conditions:
soil moisture,
snowpack
Hydrologic model spin up
Hydrologic simulation
Ensemble Forecast:
streamflow, soil moisture, snowpack, runoff
NCDC met. station obs. up to 2-4 months from current
LDAS/other real-time met. forcings for remaining spin-up
MODIS Update
1-2 years back
25th Day of Month 0
End of Month
Change in Snowcover as a Result of MODIS Update for April 1, 2004 Forecast
Snowcover before MODIS update
Snowcover after MODIS update

39. Unadjusted vs adjusted forecast errors, , for reservoir inflow volumes (left plot) and reservoir storage (right)

40. Passive microwave remote sensing for snow water equivalent
In principle, attractive since it measures the “right” variable (water equivalent rather than extent)
AMSR-E product probably is best current generation, but numerous problems (mostly generic):
Coarse resolution (~ km)
Saturation at mm SWE
Requires dry snowpacks (algorithms fail if there is liquid water in the pack)
Algorithms unreliable for mixed pixels (especially forest)
Signature is highly sensitive to grain size (and other snow microphysical properties)

41. Concluding thoughts
Hydrologic prediction skill at S/I lead times comes mostly from initial conditions.
Hence more focus on data assimilation, and its implications for forecast skill in a climatological context, needs more attention.
For drought, ESP may be the most viable method for drought recovery prediction.
The role of model error in hydrologic predictions needs more focus – how do we best weight land models in multimodel ensemble?

42. Soil moisture from UW west-wide forecast system and surface water monitor
reconstruction Aug 1934
real-time 10/13/07