Assimilating eddy covariance data with a particle filter
This is one of several projects concerned with improving land surface model predictions through data assimilation approaches. Land surface models try to predict exchanges of water, energy and trace gases between the land surface and the lower atmosphere, as well as storage and transport of water and energy on the land surface. Land surface models are used for modeling the lower boundary condition of regional and global circulation models and play therefore an essential role both for weather forecasting and climate predictions. Predictions with land surface models are affected by several sources of uncertainty:
- Uncertainty with respect to model forcings. Especially the uncertainty with regard to the precipitation amounts at a certain location plays an important role. Precipitation shows a strong spatio-temporal variability which is not captured well by rain gauge networks. The combination of rainfall radar and rain gauge data can improve the characterization of spatio-temporal variability, but uncertainty remains significant.
- Model parameter uncertainty. The uncertainty with respect to soil hydraulic parameters like porosity or saturated hydraulic conductivity and vegetation parameters like stomata resistance or rooting depth contributes significantly to the overall model prediction uncertainty. Given the fact that grid scales of land surface models are very large, it is very difficult to derive meaningful values for these parameters.
- Model structure uncertainty. Land surface models simplify the complex processes at the land surface: For example in CLM, the soil respiration is described with an empirical approach. Another example is neglecting lateral exchange fluxes of water and energy between grid cells. A third example is the fact that vegetation development is described with a static approach, neglecting the role of water stress or early/late phenology due to warm/cold conditions in spring.
- Uncertainty of initial conditions. The initial state of the system is not exactly known and subject to uncertainty as well.
These uncertainties can be reduced with help of measurement data. In an operational setting, measurement data can be assimilated with data assimilation methods to update states (for example soil moisture contents, soil temperature, leaf area index) and also model parameters (especially soil hydraulic and vegetation parameters). The updating of model parameters also has an impact for long-term climate predictions. An improved model parameterization should also improve the quality of the prediction of long-term exchange fluxes of water, energy and greenhouse gases between the land and the atmosphere.
This project focuses on the use of eddy covariance data to improve land surface model predictions. The potential of these data, also for parameter estimation, is investigated. In this study, the particle filter is used for data assimilation. The particle filter is a data assimilation methodology which is more robust for strongly non-linear models and non-Gaussian probability density functions than the Ensemble Kalman Filter. The different steps in this study (recent past and nearby future) are:
- In order to assimilate eddy covariance data correctly it is necessary to develop a measurement operator. This ensures a correct weighting of model predictions on one hand and measurements on the other hand in data assimilation procedures. Eddy covariance data have to be corrected both for systematic and random measurement errors. Eddy covariance data are strongly affected by systematic measurement errors due to the energy balance deficit. A first study conducted a multi-site analysis of the energy balance problem for about 25 European FLUXNET sites. The relation between the energy balance deficit and different variables (atmospheric stability, friction wind velocity, time of the day, season) was investigated in a multivariate framework. Clear relations could be found, but without a clear indication on how to reduce the energy balance gap. Neglecting storage terms only played a minor role in the gap, in spite of the fact that most sites were forested.
- A second study focused on random errors of eddy covariance data. We developed a more robust method to estimate random errors of eddy covariance data. This method can be used if measurements from two eddy covariance towers are available, which are typically 500 m – 1 km separated. In addition to current approaches, also a correction for the energy balance deficit is applied, and it is considered that the remaining differences in fluxes between the two neighboring towers are not only related to random errors, but also have a systematic component (e.g., differences in local soil and vegetation properties around the tower, very local precipitation differences).
- The assimilation of eddy covariance data was tested for simple 1D HYDRUS models which simulate variably saturated flow in soils including transpiration and evaporation. Test runs indicated that eddy covariance data alone only marginally improve model predictions and the characterization of soil hydraulic parameters. Soil moisture measurements at three levels in a soil profile are more helpful. However, the combination of both eddy covariance data and soil moisture measurements clearly gave the best results and made it possible to identify all unknown soil hydraulic parameters.
The data assimilation procedure will be further extended for a TERENO site. A 1D or small 2D CLM model will be built for that site and the assimilation of eddy covariance data will be tested. The aim is to estimate both soil and vegetation parameters.
Weblink to project:
Other relevant link:
W. Kessomkiat, H.J. Hendricks Franssen, A. Graf and H. Vereecken. 2012. Estimating Random Errors of Eddy Covariance Data: A More Robust Two-Tower Approach. Agricultural and Forest Meteorology. In preparation.
Hendricks Franssen, H.J., R. Stöckli, I. Lehner, E. Rotenberg and S.I. Seneviratne. 2010. Energy balance closure of eddy covariance data: a multi-site analysis for European FLUXNET stations. Agricultural and Forest Meteorology 150, 1553-1567.
Forschungszentrum Jülich GmbH
Tel. +49 02461 1795
Alternative contact person:
Prof. Dr. Harrie-Jan Hendricks Franssen
Forschungszentrum Jülich GmbH
Tel. +49 02461 4462