Evaluation of land-surface interaction in ECMWF and NCEP/NCAR reanalysis models over grassland (FIFE) and forest (BOREAS) grassland

eScholarship provides open access, scholarly publishing services to the University of California and delivers a dynamic research platform to scholars worldwide. Abstract. The National Center for Environmental Prediction and National Center for Atmospheric Research (NCEP/NCAR) and European Centre for Medium-Range Weather Forecasts (ECMWF) reanalysis models are compared with First ISLSCP (International Satellite Land Surface Climatology Project) Field Experiment (FIFE) grassland data from Kansas in 1987 and Boreal Ecosystem-Atmosphere Study (BOREAS) data from an old black spruce site in 1996 near Thompson, Manitoba. Some aspects of the comparison are similar for the two ecosystems. Over grassland and after snowmelt in the boreal forest, both models represent the seasonal cycle of near-surface temperature well. The two models have quite different soil hydrology components. The ECMWF model includes soil water nudging based on low level humidity errors. While this works quite well for the FIFE grassland, it appears to give too high evaporation over the boreal forest. The NCEP/NCAR model constrains long-term drifts by nudging deep soil water toward climatology. Over the FIFE site, this seems to give too low evaporation in midsummer, while at the BOREAS site, evaporation in this model is high. Both models have some difficulty representing the surface diurnal cycle of humidity. In the NCEP/NCAR reanalysis this leads to errors primarily in June, when the surface boundary layer stays saturated and too much precipitation occurs. In the ECMWF reanalysis there is a morning peak of mixing ratio, which an earlier work showed resulted from too shallow a boundary layer in the (cid:127)norning. Over the northern boreal forest there are important physical processes, which are not represented in either reanalysis model. In particular very high model albedos in spring, when there is snow under the forest canopy, lead to a very low daytime net radiation. This in turn leads to a large underestimate of the daytime surface fluxes, the sensible and to daytime model surface temperatures that are as much as 15 K low. In addition, the models do not account for the reduction in evaporation associated with frozen soil, and they generally have too large evapotranspiration in June and July, probably because they do not model the tight stomatal control of the coniferous forest.


Introduction
The direct comparison between field experiment time-series data and model data froIn nearby gridpoints in data assimilation and forecast systems has proved very useful in identi•ing systematic errors in the model physical parameterizations [Betts et al., 1993[Betts et al., , 1996a[Betts et al., , b, 1997[Betts et al., , 1998a and in developing improved model parameterizations [I•iterbo and BeO•ars, 1995; Hong and Pan, 1996;Chen et al., 1996]. Primarily, these comparisons identify physical processes, which are represented poorly or not represented at all in the models. In this paper we evaluate the reanalysis models [Kalnay et al., 1996 Figure 1 compares daytime (a 1200-2400 UTC average) incoming solar radiation (SWDn) for the two reanalysis models, and an average of the FIFE "FLUX" station data (solid curve). We see that on sunny days (the upper envelope) the NCEP/NCAR reanalysis (dotted) has about 50 W m -2 higher incoming solar radiation, xvhile incoming solar radiation in the ECMWF reanalysis (dashed) shows little bias with respect to the data. In contrast, for the cloudy and rainy days (the lower envelope) the fall in SWDn is largest m the data and generally larger in the ECMWF reanalysis (at T-106), which has a prognostic cloud model, than in the NCEP/NCAR reanalysis, which has the lower resolution ofT-62 and a diagnostic cloud scheme.

Surface Evaporative Fraction and Precipitation
Figure 2 compares in the FIFE "FLUX" data (top curves on lefthand scale) with the reanalysis models daytime (1200-2400 UTC) evaporative fraction (EF) defined as schemes in these models are described by Viterbo and Beljaars [1995], for the ECMWF model, and Betts et al. [1996a] for the NCEP/NCAR reanalysis. Both models have only a single vegetation type in their transpiration formulation. The NCEP/NCAR model has a globally fixed vegetation fraction of 0.7, while the ECMWF model has a global distribution of vegetation fraction. Some care is needed in intercomparing forecast model data, representative of the 100-200 km scale for the global reanalyses, and field data representative of much smaller scales. Synoptic scale processes and advection are reasonably represented by the global reanalyses, which ingest sonde and satellite-derived data in their 6-hour analysis cycle, but the land-surface processes and subsurface state variables in the model are largely derived by the interactive model physical parameterizations for the soil, vegetation, surface and boundary layer, radiation and cloud fields. Consequently, comparing the model response on diurnal and seasonal scales and to intermediate timescale processes such as the dry-down of the soil after rain episodes, with field data gives insight into how well the observed physical processes are represented by the physical parameterizations in the model. This paper will show several examples.

EF = LH/(LH +SH)
(1) where SH and LH are daytime averages of sensible and latent heat. The ECMWF reanalysis, which includes nudging of soil moisture based on 6-hour forecast errors in the low-level humidity [Betts et al, 1998a], follows the FIFE flux data more closely, although it has some significant high evaporative spikes when it rains in the model (see bottom curves). The ECMWF model tends to be low in EF in spring, and both models are high in fall, suggesting that a seasonal cycle of vegetation is needed. The NCEP/NCAR reanalysis has wider fluctuations and is generally low in midsummer [Betts et al., 1996a]. It has only two soil layers, and evaporation is high after rain, when the top 10-cm soil reservoir is replenished [see also Betts et al., 1996a]. However, this upper layer has only a few days storage, and as the model has a fixed 70% vegetation fraction, direct soil evaporation from the 30% "bare soil" helps dry the upper layer fairly quickly. By midsummer the deep (10-200 cm) layer has dried down, and it is not replenished by rain events, and after a few days without rain, EF is systematically low. The FIFE data was one data set used in off-line tests to develop the fourlayer ECMWF soil model [Viterbo and Beljaars, 1995], and the ECMWF reanalysis is able to approximately reproduce the seasonal cycle of soil moisture seen in the FIFE data in 1987 [Betts

Comparison Over FIFE Grassland
The FIFE data, discussed by Betts and Ball [1998], were an average of the surface automated meteorological stations (AMS), usually 10 in number, and the surface flux stations (variable in number from about 6 to as many as 20 during intensive field campaigns). Betts et al. [1996aBetts et al. [ , 1998a contain the full details of the data and the individual separate comparisons, but neither paper compares the two reanalyses, which we shall do here. There are some differences in our reanalysis products as well as in resolution discussed above. For the NCEP/NCAR reanalysis we have the near surface thermodynamic fields at the four synoptic analysis times of 0000, 0600, 1200, and 1800 UTC as well as the accumulated fluxes for the 6-hour short-term forecasts between the analysis times. For the ECMWF reanalysis, where fields and fluxes were archived every 3 hours (to represent better the diurnal cycle), we have fluxes accumulated every 3 hours, and the 3-hour and 6-hour forecast fields, as well as the analyses themselves. The bottom curves, with the right-hand scale, show the 24-hour total precipitation for each day that observed at the FIFE site and that for each model analysis cycle (four 6-hour short term forecasts from each analysis). Some differences are to be expected, as FIFE is a 15 X 15 km site and the two reanalyses represent short-term forecasts at nearby grid points at T-62 and T-106 resolution, but after June, the forecast models generally pick up the wet days associated with synoptic events. However, it can be seen that there is rain on more days in the NCEP/NCAR model in June (around day 165) than observed. Betts et al., [1996a] showed that this resulted from a spurious interaction among the land-surface, boundary layer (BL), radiation, and convection schemes, which we will summarize in section 2.4. Figure 3 shows the 24-hour average temperature (upper curves) and relative humidity (RH) for the FIFE automated meteorological station (AMS) data and the two reanalyses. Both reanalyses track the seasonal cycle of 24-hour mean temperature well. The most noticeable feature in the RH comparison is that while the models and data track quite well after the dry-down in late July (days 200-215), the NCEP/NCAR reanalysis has many days in June, when the model stays nearly saturated (100% RH) at the surface (see section 3, and Betts et al., [1996a]). These correspond to the periods of excess rain in Figure 2. In the corresponding temperature comparison the NCEP/NCAR model is biased a little cold in June, during the periods when it is biased wet.

Comparison Over Boreal Forest
In this section we will compare 1996 data from the BOREAS old black spruce site, 40 l•n west of Thompson (deciduous aspen, dry conifers, including jack pine, mixed stands, as well as lakes and fens), we do not yet have a landscape surface flux average, so we will use this black spruce timeseries for illustration and as a reference. Evapotranspiration from both the black spruce and the other coniferous species is low in the boreal forest, considerably lower than the total evaporation in the global forecast models. In spring and early summer the lakes are cooler than the air, and evaporation from them is also small. The boreal landscape has frozen soil and is covered by snow about half the year. Neither the albedo with snow under the trees nor the effect of frozen soil is represented properly in the global reanalyses. Both models have consequently large errors in spring, when incoming solar radiation is high, but the ground is still frozen and there is snow under the canopy. The NCEP/NCAR data (from May to October, 1996) are from that reanalysis, but the ECMWF data are from the closest grid point in the 1996 T-213 operational model, which has the same land surface model as their reanalysis model (since the ECMWF reanalysis is not available for 1996). Not only is the spacial resolution higher, but we have for each day an hourly time series from a 24-hour short-term forecast from 1200 UTC, instead of 3 hourly values. However, as we shall compare chiefly daytime means, this higher temporal resolution is not significant here. We shall contrast in some figures some additional data from the March 1997 ECMWF model to show the impact of changes in the snow albedo introduced in late 1996 in the operational model.  Figure 6 extends till the end of July. On sunny days the SH flux at this spruce site remains generally higher than in the ECMWF model and much higher than in the NCEP reanalysis. a T-62 model, so some differences can be expected. We see that on rainy days, R,,,t falls more in the ECMWF model (to values similar to those observed), suggesting that their prognostic cloud scheme is handling the cloud field reasonably well. We also see that the NCEP/NCAR reanalysis has sozne spurious rainfall peaks, as over the FIFE site. The reason appears to be the same: namely, the znodel BL stays nearly saturated all day (see Figure l0 later) with a high t?•, and rain persists. One consequence is that the SH flux stays low during these rainfall events in the NCEP/NCAR reanalysis, and evaporation stays high. In June and July both models generally have a larger LH flux and a smaller SH flux than the spruce site (Figures 6 and 7). Figure 9 compares daytime (1200-2400 UTC) EF (upper curves) and precipitation (lower curves) for the period after snowmelt to the end of July. Observed daytime average EF remains quite low at the black spruce site all summer. (The spikes generally correspond to days of low fluxes). Typically, the ECMWF model has a higher EF and the NCEP/NCAR reanalysis is higher still, particularly in June when the surface stays near saturation. For the BOREAS northern study area, deciduous species and fens, which have a much higher EF than the black spruce, cover less than 20% of the area. Thus even without a proper landscape average for comparison, it seems likely that evaporation in the global models is too high in early summer. The tight stomatal control over evaporation by the forest is not well represented in the models. Note that the agreement between the ECMWF model and the data over the boreal forest is not nearly so good as over the FIFE grassland (Figure 2). In Figure 9, from days 142 to 151, we see the steady fall of EF in the ECMWF model, as the soil moisture falls in the absence of rain; much like the behavior at the FIFE site in late July. However, the boreal spruce forest (where EF is primarily biophysically controlled) does not behave in this way at all. The data (solid line) show a small fall after the rain, probably as the moss layer dries out [Betts et al., 1998c], but EF then recovers and increases slowly, as the soil has just melted, and water is presumably more available to the trees. Clearly, further developments are needed to represent well the physiological controls of the boreal forest on evaporation over the season. The lower curves on the right-hand scale of Figure 10 compare RH. We only show the ECMWF model data from May 1, because its temperature error is so large before then. After snowmelt the ECMWF model follows the observed RH quite well, despite the biases in EF seen in Figure 9. In the NCEP/NCAR reanalysis the high RH values near 100% in May are before the snowmelt. However, RH remains high in this model in June (days 153-182) after the snowmelt, as seen over the FIFE site in June (discussed earlier in section 2.3). Again, there is excess rain (seen in Figures  8 and 9) and it is likely that the cause is the same: that the BL does not deepen sufficiently during the day and stays saturated with a high Oe and persistent rain [Betts et al., 1996a;Hong and Pan, 1996]. Later in the sununer both models and the spruce site track much better. When the RH is biased high h• the NCEP/NCAR reanalysis, not unexpectedly, temperature tends to be low. Figure 11 compares the mean diurnal cycle of precipitation for the months May-September 1996. The data show an afternoon peak at 1930 (an 1800-2100 UTC average), while the ECMWF model peak is broader and has significant rain in the 1500-1800 UTC morning period, when the observations show a minimum. Although we could not conclude this from the single site comparison, this is in fact a characteristic error of the ECMWF reanalysis model [Betts et al., 1998b]; the model precipitation maximum is close to local noon. The NCEP/NCAR reanalysis has only a 6-hour time resolution. It has an 1800-2400 UTC maximum, but the values are too high, because of the many extra days when there is rain in this model (see Figures 8 and 9).

Conclusions
We have compared the surface fluxes and surface temperature and humidity for the NCEP/NCAR and ECMWF reanalysis models over the FIFE grassland site, and one of the BOREAS study areas, using a black spruce data set for illustration. The separate FIFE comparisons are shown h• more detail, but not compared, by Betts et al. [1996aBetts et al. [ , 1998a. We have found some similarities in the model land-surface interaction over the two sites as well as some additional differences in the northern boreal region. Over the FIFE grassland both reanalyses track daytime temperature quite well, although the ECMWF model is too cold at sunrise. The models have quite different soil hydrology componentsß The ECMWF model, which has a four-layer soil model and includes nudging of soil moisture in the analysis cycle based on near-surface humidity errors, tracks daytime evaporative fraction over the season well for the FIFE grassland site, although it has high evaporation on rainy days, and needs an improved seasonal cycle. The NCEP model with a two-layer soil model has a wet bias h• June at the FIFE site, with excessive rainfall from BL errors interacting with the radiation and convection schemes, and a dry bias in midsummer, after its deep soil reservoir has dried. Both models have errors in the diurnal cycle of humidity, agah• associated with errors in the diurnal cycle of the BL growth.
Over the boreal forest we do not have landscape average fluxes, so our conclusions are more tentative and confined to the spring and early summer seasons when the model errors are largeß The reanalysis models generally do not handle the boreal forest as well as the FIFE grassland. The biggest error in both models is in the spring, when too high forest albedos with snow under the canopy 0.20 lead to large errors in surface radiation, SH and LH fluxes, and temperature. These are systematic errors of global scale at northern high latitudes. Both models do not account for the snow being under the canopy and the soil being frozen, which limit evaporation. While the ECMWF soil hydrology/vegetation model with soil moisture nudging works quite well for the FIFE grassland, it appears to give evaporation that is biased high over the boreal forest. The NCEP/NCAR model constrains long-term drifts by nudging deep soil water toward a climatology with a 60day timescale. Over the FIFE site, this seems to give too low evaporation in midsummer, while at the BOREAS site, evaporation in this model is much too high. As a result, in Jtme and July both models have much more evaporation than the black spruce site (the dominant landscape vegetation). One physical interpretation ofthis is that the models do not account for the tight stomatal control of the conifers, which limits evapotranspiration over the boreal forest. As more detailed studies of the surface energy balance in BOREAS are completed and, in particular, landscape averages are derived, we will be able to refine our conclusions further and extend them into the fall. However, we believe our conclusions are qualitatively correct as the fraction of deciduous forest (which has a higher EF) is relatively small near this site. We also show the diurnal cycle of precipitation for this site. The ECMWF model has a near-noon precipitation maximum, which has been seen in other studies, and is probably linked to the late morning peak in mixing ratio. The NCEP/NCAR summer precipitation is higher than observed. This research is part of an ongoing effort to evaluate and improve the land-surface parameterizations in operational forecast models at NCEP and ECMWF. As mentioned in the text, some of the errors identified have already been corrected in the operational forecast models, while others will be dealt with in future model versions.