2. Inter-Comparison

Table 2. Simulated 2 years (1987-1988) global mean land surface water and energy balance. Units are mm y^-1 for the moisture terms, W m^-2 for energy, dimensionless for the Bowen Ratio.

Group	Precip.	Evap.	Runoff	Net radiation	Sensible heat	Latent heat	Bowen ratio
JMA	740.3	496.14	242.66	79.7	38.97	39.59	0.98
NCEP	740.79	515.99	223.04	82.85	42.29	40.85	1.04
CSU	740.29	482.19	254.86	83.60	44.1	38.48	1.15
GSFC/Hydro	740.65	517.88	221.39	86.76	45.69	40.32	1.13
GSFC/Meso	740.13	608.74	142.07	83.44	35.2	47.43	0.74
GSFC/Clim	740.63	474.11	265.26	79.41	39.41	37.51	1.05
COLA	740.21	445.53	292.73	79.56	43.47	34.55	1.26
U.Arizona	740.31	493.49	244.19	67.84	28.39	39.38	0.72
Meteo France	740.3	535.85	202.18	81.70	39.73	42.51	0.93
Macquarie U.	740.31	556.91	200.45	73.70	30.15	44.16	0.68
CCSR/NIES	740.31	525.09	224.85	76.31	35.06	40.68	0.86

2.1 Comparison of LSPs

The ICC performed a number of comparisons of the model simulations. These include checks and comparisons of the water and energy balances on both the global scale and over selected large river basins. Table 2 shows the comparison of the two-year mean global balance of water and energy for each model. Precipitation and downward radiation were specified forcings that should be the same for each model, although there appear to be some small variations in these terms on the order 0.1% or less. The COLA version of SSiB produced the largest global Bowen ratio of all the models at 1.26; the BASE model from Macquarie University produced the smallest ratio: 0.68. The mean Bowen ratio of all the models is just below unity. This partitioning is reflected in the distribution of precipitation between runoff (both surface and baseflow) and evapotranspiration. The spread among models in their partitioning of the surface water balance between runoff and evapotranspiration in several large river basins is illustrated in Figure 2. In the majority of cases, the models under-forecast runoff. The observed evapotranspiration is derived from observed precipitation and streamflow, assuming no net change in storage during the two year period. Most of the models behave consistently between the basins, with the same models routinely appearing at either end of the distribution. The models as a whole particularly under-forecast runoff over the colder basins where snow is a major component of the precipitation.
 

Figure 3. The new snow formulation in the GSFC version of SSiB improves the timing of the simulated spring snowmelt.

One of the advantages of participation in GSWP, which many in the Production Group have discovered, is that once their LSPs have been set up with the proper framework to perform the two-year integration for the project, they have created an excellent testbed for model development. There are several prominent examples in this pilot phase.

Douville et al. (1999) found that they could improve the simulated basin-scale runoff in the ISBA LSP by implementing several changes. They found that by improving upon the treatment of convective precipitation as specified in the GSWP guidelines, and by employing the runoff scheme of Dümenil and Todini (1992), they were able to increase runoff, which was generally underestimated in the control simulation.

Sud and Mocko (1999) describe the development of an improved parameterization of snow that was inspired by the poor performance of their version of SSiB in the GSWP validation over the wheat belt of Russia and over Canada. The new snow scheme employs a separate snow layer with its own thermodynamic treatment and radiative properties. They found a significant improvement in the timing of the spring snowmelt (Figure 3), in the partitioning of the meltwater between infiltration and surface runoff, and more realistic shallow soil temperatures.

Boone and Wetzel (1999) find evidence that the use of aggregated soil hydraulic parameters at 1×1 resolution may lead to significant underestimation of runoff, and overestimation of evapotranspiration. The highly non-linear dependence of the fluxes of heat and moisture within the soil on conductivity and matric potential lead to considerable differences between explicitly heterogeneous and aggregated treatments of the soil properties. Their results suggest that heterogeneity can be difficult to account for in single-column treatments of the soil.

Pitman et al. (1999) found, during sensitivity experiments to examine the role that the specification of leaf area index (LAI) may have on simulation of the surface energy and water balance, that the difference between the use of the 1×1 monthly data set of LAI from ISLSCP Initiative I and a simpler specification where LAI is simply a function of vegetation type and month was well within the range of uncertainty for the global LAI data set. Thus, they conclude that the simpler approach, historically used by many land surface schemes but gradually being abandoned, is still justifiable. The implication in this case is that an element of model development which many are now undertaking, namely the inclusion of LAI information derived from high-resolution satellite data, may not currently be beneficial.

Douville et al. (1999) proposed a land data assimilation system, based on an optimal interpolation technique using errors in the predicted near-surface temperature and relative humidity (Mahfouf 1991), to produce a soil moisture analysis.  This technique has been recently implemented in the operational weather forecast system at Météo-France (Giard and Bazile 1998) and validated against the First ISLSCP Field Experiment (FIFE; Sellers et al. 1992) data set in the ECMWF single column model (Douville et al. 1998).  In order to implement the method in the framework of GSWP, it is necessary to compare the predicted 2 m parameters with 2 m observations, and therefore to replace the ISLSCP screen-level meteorological forcing by the forcing at the lowest level of the ECMWF model from either the reanalysis (Gibson et al. 1994; experiments JU0, JW2, JF2, JU4 and JU5) or the operational analysis (JU6 and JU7).  Under ideal conditions,  i.e. using pseudo-observations produced at each model grid point by a reference 31-day (July 1987) ISBA simulation, reintegration of the GSWP experiment with perturbed initial soil wetness (wilting point for JW2, field capacity for JF2) shows the land data assimilation technique to converge to the soil wetness of the control case within a few weeks, as illustrated for selected grid points close to the sites of the FIFE and ABRACOS (Anglo-Brazilian Amazonian Climate Observation Study; Gash and Nobre 1997) field campaigns (Figure 3). Using real SYNOP observations and unperturbed initial soil wetness, the analyzed soil moisture is different from the reference forecast (bottom of Fig 3) and can be quite sensitive to the choice of the meteorological forcing (ECMWF reanalysis or analysis).  Using a high pass filter of the 2 m temperature errors in experiments JU5 and JU7, the spurious effect of systematic biases in the meteorological forcing is avoided, and therefore sensitivity to the choice of the forcing is reduced. This is true for the FIFE grid point, but the analysis remains very dependent on the ECMWF forcing for the ABRACOS grid point.  This result indicates the need for a more accurate atmospheric forcing, or for implementing the data assimilation with a free boundary layer in order to give more robustness to the results and to be more confident in the resulting soil moisture analysis.
 

Figure 4. Evolution of soil wetness index (SWI) from forecast (JU0) and assimilation modes for grid points corresponding to FIFE (left) and ABRACOS (right) field sites.  Top: SWI initialized at wilting point and field capacity, and assimilating pseudo-observations of temperature and humidity from the JU0 simulation.  Bottom: Assimilating actual observations with ECMWF analysis meteorological forcing (JU4, JU5) and with ECMWF operational analysis (JU6, JU7) — JU5 and JU7 use filtered increments.