MiKlip first phase: PastLand

Optimum parameter and state estimation of the land and biosphere - initialization of a global coupled land model.

Predictions of the global climate depend on both the model's initial state and the anticipated change in aerosols and greenhouse gases; for decadal predictions anthropogenic climate change and natural variability are expected to be equally important. A close representation of the observed climate state in global coupled climate models is therefore crucial for (the initialization of) decadal predictions. However, predictability beyond two weeks is essentially influenced by time scales which are longer than typical scales of weather phenomena. These slow components which affect seasonal to decadal predictability of the Earth system are beside the oceans, glaciers and sea ice, the moisture content of the soil, snow cover and the terrestrial biosphere.

MiKlip PastLand aims at a comprehensive, combined state and parameter estimation of a climate model land surface scheme using simultaneously observations for different land surface variables. Focus is hereby given on a most realistic model parameter optimization at the model grid scale to improve the model predictive skills for seasonal to decadal predictions. A flexible observational framework will be built up that utilizes existing land surface observations obtained by remote sensing from satellites. Investigations will be based on the JSBACH land surface scheme, which is part of the MPI-M coupled ECHAM/MPI-OM/JSBACH Earth system model. Model predictive skills will be verified using coupled and uncoupled climate model simulations in hindcasting applications and independent observational data.
 
The variational state and parameter estimation for JSBACH will be carried out in offline mode, i.e. the model is driven with 'observed' atmospheric forcing. The project nevertheless addresses the initialization for the coupled model by analyzing and correcting potential biases in the exchange fluxes with the atmosphere that may be caused by the offline approach. The development and operation of a variational assimilation system for JSBACH in offline mode already constitutes a scientific and technological challenge and is an important first step towards a variational initialization of the entire MiKlip-model.

Goals

  • Identify regions where the memory of the land surface has an impact on the climate and on which time scales (from seasons to decades) this impact is noticeable.
  • Exploit the potential of new satellite observations (e.g. SMOS, <nobr>SENTINEL-2</nobr>) for Earth System research and an improved estimate of the land surface state.
  • Assessing the impact of observational datasets and initialization procedures on seasonal to decadal climate predictions.
  • Development and assessment of a combined optimum model and state estimation tool, which can be used for the initialization of seasonal to decadal climate predictions.
  • Evaluation of observation uncertainties on model initialization and prognostic skills.

 

PastLand - Results from DS1

PastLand has developed a five-layer scheme for the simulation of soil water fluxes for the land surface model JSBACH. This scheme uses layer depths of 0.065, 0.319, 1.232 und 4.134 m and ends at the bedrock or at 10 m. Diffusion and percolation are calculated using the Richards equation. This new scheme limits bare soil evaporation to the upper layer and the transpiration to the root zone. This means that lower layers can be considered as long term reservoirs. The evaluation of the scheme shows in a comparison to the standard ECHAM6 AMIP setup only small differences in the long term mean of precipitation, evaporation, run off and temperature. The soil moisture memory does, however, change significantly; it is increased in large parts of South America, Europe, East Asia and Central Africa and decreased in areas with little vegetation.

In parallel with the development of the new scheme, new field capacities for JSBACH were derived from a highly resolved data set (Kleidon, 2004), by converting it to the discrete Olsen classification. The new dataset was tested with offline and online simulations for both the 1- and the 5-layer soil schemes. Clear differences were found in the soil moisture and in the soil moisture memory, but the effect of the updated data set on the water fluxes and surface temperature was relatively small.

Two satellite derived datasets for soil moisture (ESA_ECV_SM and AMSR-E) were analysed in terms of their suitability for model validation. After spatial and temporal interpolation to cover grid cells with missing data, these datasets were compared to simulation output from JSBACH and to ERA-Interim Reanalysis. Even though the ESA ECV_SM dataset has some drawbacks, it is currently the most appropriate soil moisture data set to use for a validation of land surface models, because of its long time coverage (1979-2010) and its quasi-global spatial coverage. A variational data assimilation scheme to use with the offline version of JSBACH was developed and tested. This scheme allows the estimation of the initial state and process parameters (control variables). To this end, the scheme minimises a cost function that quantifies the difference between model output and observational data plus the deviation from priori values of control variables. This scheme was configured for 13 control variables.

Figure 1: Convergence of iterative minimisation. Cost function (black), norm of its gradient (red) and norm of difference of control vector to standard values (green) over iteration number.

The derivative code that was derived for the assimilation scheme was tested for an observational site in Botswana against finite differences of model simulations. Additionally, the reliability of the prototype was demonstrated in an identical twin experiment. Here, half-hourly values of water and carbon fluxes (pseudo observations) were generated from default values of the control variables. The interactive minimisation of the cost function was then started from perturbed values of the control variables. Within about 30 iterations the system recovers the standard parameter values, reduced the cost function values by about 13 orders of magnitude and its gradient by about 8 orders of magnitude.

More on PastLandNews-Icon

Contact

Max Planck Institut für Meteorologie (MPI-M)
Tobias Stacke
Stefan Hagemann
Alexander Löw
Christian Reick

FastOpt
Thomas Kaminski