Assimilation of synthetic GRACE observations to improve model prediction of soil moisture at the cat

1
Assimilation of synthetic GRACE observations to
improve model prediction of soil moisture at the
catchment scale
Kevin M. Ellett1,2, Jeffrey P. Walker1, Andrew W.
Western1, Rodger B. Grayson1, and Matthew Rodell3
1 Department of Civil and Environmental
Engineering, The University of Melbourne,
Australia 2 United States Geological Survey,
Water Resources Discipline, Sacramento,
California, USA3 Hydrological Sciences Branch,
NASA Goddard Space Flight Center, Greenbelt,
Maryland, USA

• 
  Introduction
• What is GRACE? GRACE is a novel twin satellite
  system which measures very precise changes in the
  Earths gravity field.
• GRACE observations should be accurate enough to
  relate gravity changes to the spatial and
  temporal redistribution of mass caused by
  hydrological processes (eg. the infiltration of
  precipitation, evapotranspiration, and ground
  water level changes).
• GRACE provides the first-ever measurements of
  terrestrial water storage changes occurring
  throughout the globe. Such measurements could
  allow
• Closure of the terrestrial water balance and
  analysis of the monthly, seasonal, and
  inter-annual hydrological variability at the
  basin to continental scale.
• A unique assessment of the uncertainty and
  performance of numerical models which simulate
  large-scale hydrological processes.
• Potential improvement of hydrological model
  prediction by way of data assimilation (DA)
  techniques.
• In this poster we use a synthetic twin study
  approach to investigate the potential for GRACE
  to improve catchment-scale model prediction in
  the Murray-Darling Basin of Australia (MDB) by
  way of a variational DA method.

• Challenges to GRACE DA
• Downscaling GRACE observations are only accurate
  when averaged over large areas (ie. hydrological
  basins gt500,000 km2) at monthly to annual time
  periods.
• Decomposition GRACE provides only the vertically
  integrated total terrestrial water storage change
  signal (the sum of soil moisture, ground water,
  surface water and snow/ice).
• Observation Error For GRACE DA to be effective,
  the magnitude of the natural storage change
  signal must be larger than the inherent
  uncertainty in GRACE.

• Methods
• Hydrological Model We developed a model of the
  MDB by modifying the conceptual rainfall-runoff
  code SIMHYD Chiew et al., 2002 to run at the
  basin scale (106 km2), with discretisation into
  26 major catchments (104105 km2). The model was
  run on a daily time step and forced by areal mean
  precipitation and potential evapotranspiration
  calculated from 87 Australian Bureau of
  Meteorology climate stations located throughout
  the basin.
• Synthetic Twin Study Results from the initial
  baseline model run were considered as truth and
  were used to generate a synthetic GRACE data set
  of monthly mean, basin-wide total storage
  changes. The model was then run again for the
  same time period but in a degraded fashion where
  error was introduced into the initial conditions.
  The synthetic GRACE observations were then
  assimilated into the degraded model to determine
  if any improvement could be achieved through DA
  of GRACE-type observations.
• Data Assimilation A variational DA method was
  developed using a 2-month assimilation window and
  an automatic pattern search optimisation routine
  to minimise the difference between the synthetic
  GRACE observations and the basin-wide model
  output of mean monthly soil moisture storage
  change. In this study we assumed that the total
  storage change signal had been decomposed prior
  to DA (using either the model or observations
  such as those shown above). The variational DA
  method then downscales the soil moisture signal
  through the optimisation of the catchment-scale
  model states.

• Results and Conclusions
• Results shown below for the soil moisture model
  states clearly indicate that DA of GRACE-type
  observations can substantially improve model
  prediction at not only the basin scale, but also
  the catchment scale, due to the effective
  downscaling by the variation DA method.
• The initial basin-wide mean monthly soil moisture
  storage change predicted by the degraded model
  from January 2001 to February 2001 was negative
  40mm, versus negative 1mm from the synthetic
  GRACE observation. DA of the GRACE observation
  reduced the root mean square error (RMSE) from
  38mm to about 3mm for the MDB, and approximately
  the same amount for the Murrumbidgee and
  Condamine catchments. A smaller average RMSE
  reduction of 16mm was observed in the Goulburn
  catchment due to its lower areal contribution to
  the signal (ie. poorer constraint).
• Our current research involves a more rigorous
  assessment of the advantages and limitations of
  GRACE DA and focuses on improving the model and
  the DA optimisation algorithm.

For more information please visit our website at
http//www.civenv.unimelb.edu.au/jwalker/data/gsm
/hydrograce.html