Soil Moisture Data
Assimilation Using the 1D Kalman Filter
|
|
|
1,2Jeffrey P. Walker and 2Paul
R. Houser |
|
|
|
1Goddard Earth Sciences and
Technology Center 2NASA/Goddard Space Flight Center |
|
|
|
http:land.gsfc.nasa.gov/~cejpw |
The OPE3 Field
Site
Model Forcing and
Parameters
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Vegetation parameters derived from
relationships with corn height. |
|
Meteorological forcing data measured in
the field with an automatic weather station. |
Soil Moisture Data
Simulation Results
Plans ?
|
|
|
Modify model to have a 10cm surface
soil moisture to be compatible with observations. |
|
|
|
Calibrate model to soil moisture and
runoff. |
|
|
|
Assimilate surface (10cm) soil moisture
for simulation with poor initial condition in both calibrated and
uncalibrated model. Compare with no assimilation. |
Slide 7
Slide 8
Slide 9
Motivation
|
|
|
|
Gain experience with the 1D Kalman
filter in a series of controlled experiments. |
|
What are the soil moisture needs for
NSIPP and what are the defensible requirements of a remote sensing mission
for measurement of surface soil moisture? |
|
Polarization – horizontal |
|
Wavelength – L-band |
|
Look Angle – < 50° |
|
Observation Accuracy – ? |
|
Temporal Resolution – ? |
|
Spatial Resolution – ? |
Methodology Overview
|
|
|
Use a LSM to generate a “truth” data
set that provides both surface soil moisture “observations” and evaluation
data. |
|
Degrade the land surface forcing data
and initial conditions to simulate uncertainties in this data (assume a
perfect model). |
|
Run the LSM with degraded data. |
|
Run the LSM with degraded data and
assimilate the “observations” with various levels of error imposed using the
extended Kalman filter. |
|
Compare with the “truth”. |
The Catchment-Based LSM
Data
|
|
|
|
Model Input Data |
|
Atmospheric forcing data and soil and
vegetation properties were taken from ISLSCP-1. |
|
Initial Conditions |
|
Spin-up catchment-based LSM for 1987
using ISLSCP forcing data. |
|
Surface Soil Moisture Observations |
|
Surface (2cm) soil moisture data output
every 3 days from the “truth” run. |
|
Evaluation Data |
|
Surface, root zone and total profile
soil moisture data output each day plus average evapotranspiration data
output each 10 days. |
Degraded Data
|
|
|
Zero mean normally distributed
perturbations with standard deviations given below added to the initial
conditions, forcing and obs data. |
Effect of Observation
Error: 1
|
|
|
Time Series Histogram |
|
|
|
Precipitation (mm/day) Profile Soil Moisture
(%v/v) |
|
|
Effect of Observation
Error: 2
Effect of Observation
Error: 3
|
|
|
Evapotranspiration RMS Error (mm/day) |
|
|
|
|
|
|
|
|
|
|
|
|
|
Evapotranspiration Bias (mm/day) |
Effect of Observation
Error: 4
|
|
|
Profile Soil Moisture RMS Error (v/v) |
|
|
|
|
|
|
|
|
|
|
|
|
|
Profile Soil Moisture Bias (v/v) |
Effect of Observation
Error: 5
|
|
|
|
|
|
|
|
|
Precip Bias (mm/day)
Soil Depth (mm) Avg Pro Soil Moist (v/v) |
Effect of Forcing Bias: 1
|
|
|
Time Series Histogram |
|
Precipitation (mm/day) Profile Soil Moisture (%v/v) |
Effect of Observation
Error: 1
|
|
|
Time Series Histogram |
|
Precipitation (mm/day) Profile Soil Moisture (%v/v) |
Effect of Forcing Bias: 2
|
|
|
Time Series Histogram |
|
Precipitation (mm/day) Profile Soil Moisture (%v/v) |
Effect of Forcing Bias: 3
Effect of Temporal
Resolution
Spatial Dissagregation
Effect of Spatial
Resolution
Conclusions
|
|
|
Typically observations of surface soil
moisture must have an accuracy of better than 5% v/v. |
|
If model covariance forecasts are
imperfect and observation errors are large there may be a slight degradation
of the model forecast. |
|
To achieve the best results, it is
important that the model, model forcing and observations be unbiased, or the
bias modeled as part of the assimilation framework. |
Conclusions
|
|
|
Daily observations achieved the best
results. |
|
The greatest impact of temporal
resolution was for 1 to 5 days, with greater time between observations having
a marginal degradation. |
|
Spatial resolution less than the model
resolution achieved the best results. Greater resolution produced only
slightly worse results. |
|
Observations at half the spatial
resolution of the model is a good compromise. |
Slide 29
Slide 30