3.1. Atmospheric Water Budget (June 1995 to May 2002)
 Figure 1 presents the Mississippi River basin and its subbasins. Because river discharge cannot be measured reliably in the lower Mississippi below Vicksburg (32.3°S, 90.5°W), all area averages exclude this subbasin. In this form, area averages are consistent with the streamflow measured at Vicksburg. [However, for reference, figures in this article will depict the contour of the whole basin.]
 The 1995–2002 time series of the Mississippi basin area-averaged Eta model 12–36 h forecast precipitation and observed precipitation (based on the daily analyses of Higgins et al. ) are presented in Figure 2a, along with a sequence of numbers that represent the major changes in the model as described in section 2 and Table 1. The two curves depict close similarity in magnitudes and the year-to-year variability, but some changes are noticed over the years. Differences are larger before 1998, as the forecast precipitation (heavy line) tended to have higher month-to-month variability and discrepancies in magnitude. From mid-1998 onward, when atmospheric and land states (including soil moisture) began to be cycled without a dependence from NCEP's global model (indicated by 5 in Figure 2a), the observed and forecast precipitation tend to show a closer agreement. The last part of the record (marked with number 9) shows even closer correspondence. This was the time when the EDAS began to assimilate observed precipitation.
 The differences between the early and late periods are manifested in the scatterplots of monthly observed vs. forecast precipitation averaged for the Mississippi basin (Figure 3). During 1995–1997 (Figure 3a) the Eta model had a dry bias during both the cold and warm seasons, particularly in the range 1–3.25 mm day−1. On the other hand, during 1998–2002 (Figure 3b) a more even distribution of points along the symmetric axis is noticed. There appears to be a slight wet bias, particularly for large precipitation during winter, but the magnitude is smaller than the dry bias of previous years. The dry bias during the first years was observed in all subbasins except the Ohio basin (not shown), but it was removed beginning in mid-1998. The slight wet bias during the 1998–2002 winters affects the Missouri and Upper Mississippi basins.
Figure 3. Scatterplots of Mississippi basin area-averaged monthly observed precipitation versus Eta model 12–36 h forecast precipitation for (a) 1995–1997 and (b) 1998–2002. Warm season is defined as May-August, and cold season is defined as November-February.
Download figure to PowerPoint
 Evaporation was estimated as a residual of the water balance equation, Eres = Po + ∇ · qVdp + Qt where Po is the observed precipitation, ∇ · qVdp is the vertically integrated moisture flux convergence, and Qt is the local change in water content in the column. The last two terms are computed from the Eta model 12–36 h forecast. Qt, is significantly smaller than the other terms, and plays no role in a long time average as this one. In reality, evaporation estimated this way also includes other potential sources not accounted in the simple balance equation, but the results of Maurer et al.  show that this computation has less errors than other estimates. Moisture flux convergence, presented in Figure 2b, ranges between −2 mm day−1 and +2 mm day−1 most of the time. As with precipitation, the month-to-month variability seems to be smaller during the second half of the period. These changes may be due to the adjustment of the model to different initial conditions beginning in 2000. According to Figure 2c, the evaporation estimated as a residual of the water balance has larger amplitude, and larger interannual variability than the evaporation estimated with the VIC model, which has a rather uniform annual cycle. The residual evaporation also tends to peak about one month before the VIC estimate. During the winter 1996/1997 and less dramatically during other winters, negative values of evaporation are produced. This may be in part due to the underestimation of observed precipitation during the cold season, and to errors in the estimated moisture flux convergence. The parameterization of the Eta model evaporation, also presented in Figure 2c, was changed several times along the years, nevertheless it continued to have excessive values during spring. This will be discussed later in the article.
 Figure 4 presents the annual mean observed and forecast precipitation. The 12–36 h Eta model forecast precipitation (Figure 4a) bears most of the features of the high resolution rain gauge based analysis (Figure 4b). According to the difference field (Figure 4c), the forecast precipitation has a slight negative bias over the Mississippi basin, and positive over Florida and along the southeastern United States coastline (of about 1–2 mm day−1). However, this bias is notably smaller than that depicted in global reanalysis precipitation [Higgins et al., 1996]. Although the maxima and field structure toward the west seem similar in the model and observations (Figures 4a and 4b), the orographic effects are marked and produce large biases (Figure 4c) particularly over the Cascades and Sierra Nevada. The model forecasts over the southern coast of California and central Arizona are slightly drier than observations.
Figure 4. June 1995 to May 2002 annual mean fields of (a) Eta model 12–36 h forecast precipitation, (b) observed precipitation, and (c) their difference.
Download figure to PowerPoint
 Other components of the atmospheric water cycle are presented in Figure 5. The moisture flux convergence (Figure 5a) is positive over most of the Mississippi basin and particularly over the Ohio basin, where annual values achieve a maximum of about 2 mm day−1. Slightly negative values between −1 and 0 mm day−1 (divergence) are found over a narrow band of the Central U.S. in the western part of the Mississippi basin. A second region of divergence in the southeastern U.S. runs along the coastline, which is associated with the dominant moisture flux divergence over the oceans. To the west, patterns are more difficult to interpret and large values of either sign are found over the complex terrain of the Rockies, Sierra Nevada, and the Cascade Mountains. These kinds of patterns are typically produced in mesoscale model simulations where the moisture flux divergence can achieve larger magnitude and gradients than typical estimates from global reanalyses.
Figure 5. June 1995 to May 2002 annual mean fields of (a) vertically integrated moisture flux convergence, (b) evaporation as a residual of the water balance equation, (c) Eta model parameterized evaporation, and (d) VIC estimated evaporation.
Download figure to PowerPoint
 Evaporation computed as a residual of the atmospheric water cycle equation is largest in the eastern part of the country, toward the coasts of the Gulf of Mexico and the Atlantic Ocean (Figure 5b). Values progressively decrease toward the north, and mostly range between 1 and 4 mm day−1. The Eta model parameterized evaporation (from the 12–36 forecasts; Figure 5c) presents a much smoother field, and its magnitude over the eastern part of the basin is larger. Largest differences between the two evaporation estimates are found in the western part of the country, most prominently over California, and in particular the Central Valley, where the evaporation estimated as a residual from the Eta model exceeds 5 mm day−1. The large values in this region result from the moisture flux convergence term (Figure 3d) in the balance equation. Because of the nonexistence of evaporation observations, these fields are compared to VIC's evaporation, which responds to the surface water balance equation forced with observations or parameters derived from observations. Figure 5d shows that the VIC's evaporation field is also more uniform and smoother than the one estimated as a residual of the water balance. Values within the Mississippi basin range from somewhat less than 1 mm day−1 over the northwestern sector of the Missouri subbasin to near 3 mm day−1 over the lower Mississippi subbasin. VIC's evaporation along the Gulf of Mexico coast has lower values than the two Eta estimates. In all these cases the reader is reminded that VIC data sets are not available after July 2000, so that differences may also be due to the difference in the averaging period.
 The June 1995 to May 2002 mean annual cycle of the atmospheric water cycle averaged over the Mississippi basin is presented in Figure 6a. Observed precipitation achieves a maximum during May–June and smoothly decays to a minimum in December. Moisture flux convergence tends to remain constant at about 1 mm day−1 from October to May of the next year, when it begins to decrease and in fact converts to divergence from June to September, with largest divergence values near −1 mm day−1 during August. The divergence of moisture flux during summer is a typical feature of the Mississippi basin [see, e.g., Berbery and Rasmusson, 1999]. The local change of precipitable water is small at all times, and does not measurably contribute to the estimate of evaporation.
Figure 6. June 1995–May 2002 Mississippi basin area-averaged mean annual cycle of (a) moisture budget components, i.e., observed precipitation, Eta model 12–36 h forecast vertically integrated moisture flux convergence, local change of water vapor content, and evaporation (Eo(res)) estimated as a residual of the moisture balance; (b) the observed precipitation, and the Eta model 12–36 h forecast precipitation; and (c) VIC evaporation, evaporation as a residual of the water balance equation using observed precipitation, and the model parameterized evaporation.
Download figure to PowerPoint
 The evaporation estimated as a residual of the water balance equation is consistent with previous results, including the Eta-based 2-year estimate by Berbery and Rasmusson . During winter, values are small and positive although, as previously shown, individual years may achieve negative values. Evaporation increases rapidly during spring achieving a maximum of 3.5 mm day−1 in July. The decrease of evaporation during the second half of the summer may be associated with the drying of the soil moisture, in addition to the browning (senesence) of the vegetation. As will be shown later (Figure 8), by August the soil moisture dry-down has reached a sufficiently dry state as to contribute to vegetation transpiration stress—that is, decrease of evaporation through the plant canopy owing to the increasing soil moisture deficit of late summer (August–September).
 The difference between the 12–36 h forecast Eta precipitation over the Mississippi basin and observations is not uniform along the year (Figure 6b). During the warm season, the model tends to produce less precipitation than suggested by observations. During winter and early spring, the model produces slightly more precipitation (at most 0.5 mm day−1), and this does not necessarily mean an erroneous forecast, since others have shown the winter underestimation of solid and liquid precipitation [e.g., Groisman and Legates, 1994]. Similar biases were found by Adam and Lettenmaier  not only at regional scales but even on the mean annual global terrestrial precipitation. This emphasizes the need to have more studies in this area given the implications it has for model performance evaluation.
 Figure 6c presents the two estimates of evaporation, as a residual of the water balance equation and the model parameterized evaporation. The first aspect to be noticed is that although their magnitude is close to the VIC's estimate, they are larger during spring and smaller during autumn, causing a one month shift with respect to VIC's, which achieves a maximum during July–August. The excessive model evaporation during spring causes a shift in other variables as well: as a result of the Eta land surface parameterizations producing too much evaporation in spring, by August the Eta soil moisture is sufficiently more depleted than VIC (not shown), and consequently the Eta surface evaporation in August becomes more stressed than that of VIC. The bare soil evaporation in the Eta scheme was improved in the 24 July 2001 implementation (see Table 1) to greatly reduce the Eta bias of surface evaporation over relatively moist “bare” soils (meaning sparse green vegetation; this situation is most pervasive in spring over the Mississippi Basin). Hence, until the 24 July 2001 implementation, the Eta had a low-level near-surface air temperature cool bias that arose as a result of the land surface evaporation being too high and the Bowen Ratio being too low over moist bare soils over the eastern two thirds of the U. S. during spring (before the emergence of crops and before significant early-summer green-up of natural vegetation) [Mitchell et al., 2002].
 Table 2 presents the 7-year means of the water cycle components, and the observed streamflow at Vicksburg is added for comparison. Two estimates of observed precipitation are presented, the first one from Higgins et al.  and the second one that was used as a forcing for VIC. The latter has a correction due to orography effects, but over the Mississippi basin their difference is about 0.1 mm day−1. The difference between the long-term average of the Mississippi basin-averaged observed and forecast precipitation is about 2%. Differences are larger between evaporation estimates; the residual evaporation is about 5% lower than VIC's, but the parameterized evaporation is about 17% larger. The implied magnification of the residual evaporation with respect to the difference in precipitation is due to the nature of the computation, since the moisture flux convergence is about an order of magnitude less than precipitation.
Table 2. June 1995 to May 2002 Annual Mean Basin-Averaged Precipitation (P), Streamflow (S), Moisture Flux Convergence (MFC), and Evaporation (E)a
| ||Value in mm day−1||Value in m3 s−1|
| Pm||1.98|| |
|PVIC − EVIC||0.52||17840|
| Eres = Po − MFC||1.48|| |
| Em||1.94|| |
 Because of the surface water balance equation, in a long-term average the model moisture flux convergence should equal the observed streamflow. Observed streamflow for the period 1995–2000 (observations were not available afterward) is on average 17128 m3 s−1, which is equivalent to 0.50 mm day−1 if the basin area is taken into account, while the Eta model forecast moisture flux convergence is 0.54 mm day−1. The difference is less than 10% and represents a notable achievement by the Eta model. The estimate of streamflow as the difference between VIC's precipitation and evaporation is 0.52 mm day−1, despite the record length being about two years shorter.
 These estimates do not consider the effect of upstream diversions and reservoir evaporation, which is not accounted for in the Eta model. Roads et al.  discusses and applies adjustments and corrections for these two effects. According to Lettenmaier (personal communication, 2003) the “naturalized flows” should be of the order of 8% larger than the measured flows.
3.2. Land Surface Water Budget
 Figure 7 presents the annual mean fields related to the surface water as produced by the Eta model parameterizations. In this case, the averages are performed for the period June 1998–May 2002, to avoid the earlier period when surface parameterizations had important changes; we still included several months after the July 2001 upgrade (Table 1) to have four complete years. Note that the Eta model estimates reported in Roads et al.  are somewhat different to those reported here. The reason is that those estimates were based on the period 1996–1999, thus the first half of the period was not favorable given the model changes explained in section 2.
Figure 7. June 1998 to May 2002 annual mean (a) water equivalent of accumulated snow depth, (b) soil moisture for the 0–200 cm layer, and (c) runoff. All fields are computed from the Eta model 12–36 h forecasts.
Download figure to PowerPoint
 The water equivalent of accumulated snow depth presented in Figure 7a shows a smooth latitudinal gradient within the Mississippi basin, with values of about 10–20 mm toward the north. Large values of snow accumulation exceeding 50 mm are found over mountainous regions, particularly affecting the western part of the Missouri basin, although the largest values fall outside the basin. The structure of the field is similar to that from VIC (not shown here, but see the WEBS document online at http://ecpc.ucsd.edu/gcip/gcipwebs.html). The total content of soil moisture in the layer 0–2 m is presented in Figure 7b. The western subbasins (Arkansas/Red and Missouri) are the driest, with values between 400 and 500 mm, while toward the east, the Ohio subbasin and the Mississippi delta are the ones with highest soil moisture (600–700 mm).
 The Eta model forecast runoff within the basin (Figure 7c) is largest on the eastern part of the basin and the lower part of the Mississippi, and is similar to the estimates from the VIC model (again, see http://ecpc.ucsd.edu/gcip/gcipwebs.html). Runoff achieves large values near the Rockies, the Central Valley in California and the coast of Texas, but no observations are available to verify these features.
 The Mississippi basin-averaged mean annual cycle of surface water variables is presented in Figure 8. For two variables, snow and runoff, the VIC estimates are included for comparison (VIC is available only until mid-2000). Observed streamflow at Vicksburg is also included, and its units have been converted to mm day−1 taking into account the basin's surface area. Compared to VIC, the Eta model snow water equivalent (Figure 8f) has a positive bias during winter, and decays faster during spring. Because of the limited overlap between the Eta and VIC time series, no clear conclusions can be inferred for the runoff (Figure 8i) and the observed streamflow (Figure 8j), but the magnitudes and year-to-year changes are similar.
Figure 8. June 1998 to May 2002 Mississippi basin area-averaged mean annual cycle and time series of water equivalent of accumulated snow depth (a and f), total soil moisture (b and g), evaporation (c and h), runoff (d and i), and observed streamflow (e and j) converted to mm day−1. Dotted lines in Figures 8f and 8i are estimated from VIC. Dotted line in Figure 8e is the 1962–2000 average.
Download figure to PowerPoint
 The mean annual cycle of all variables seems to have a consistent evolution. The Eta model's water equivalent of accumulated snow (Figure 8a) has nonzero values starting in November, achieves a maximum of about 25 mm in January and later decays (snowmelt) until April-May. Soil moisture (Figure 8b) achieves a maximum in spring, about 3 months after the maximum in snow. Then it decays monotonically until October, due to the increasing evaporation (see Figure 8c), and reduced precipitation during summer. Recall that during the first half of the year, the model evaporation is larger (by about 1–2 mm day−1) than the estimates as a residual from the water balance as shown in Figure 6c. The reasons for the excessive evaporation during spring and subsequent reformulation of the bare soil evaporation are discussed by Mitchell et al.  and Ek et al. (submitted manuscript, 2003). Model runoff (Figure 8d) also achieves a maximum during late winter and spring, while snow is melting, after which it also decays until the following winter. The annual cycle of Eta model runoff is like the longer term (1962–2000) mean annual cycle of the observed streamflow (dotted line in Figure 8e). However, the observed streamflow mean annual cycle for the period 1998–2000 is irregular and almost constant from February to July. (Further work with routing and management models is needed to transfer model runoff to an accurate simulation of streamflow.)
 Although model upgrades prevent a reliable analysis of the interannual variability, there are changes that occur consistently among all variables and observations. The winter of 1999/2000 had a minimum in water equivalent of accumulated snow (Figure 8f) which was followed by low values of soil moisture the next spring (Figure 8g), and somewhat smaller evaporation and runoff also during spring (Figures 8h and 8i). Note that observed streamflow (Figure 8j) also was smaller during the year 2000, which is a confirmation that the independently computed values in the model are consistent. Similarly, the peak in snow in January 1999 was followed by larger values of soil moisture and runoff in March 1999, and evaporation in June 1999. A similar sequence can be noticed after the peak in snow in December 2000.
 The annual means of the surface variables are presented in Table 3. Differences in precipitation are larger (5%) than for the 7-year period, despite the improvements in model estimates, but at the same time a shorter period is covered. VIC's evaporation falls in between the two Eta evaporation estimates: the Eta model parameterized evaporation (Em) and the estimate as a residual of the balance equation differ from VIC's by about +17% and −20% respectively. Measured streamflow at Vicksburg was converted to mm day−1 to compare to runoff estimates. The difference between VIC's runoff and streamflow is of about 4%, while the Eta model parameterized runoff differs from them by about 15–20%.
Table 3. June 1998 to May 2002 Annual Mean Basin-Averaged Observed Precipitation (Po), Eta Model Precipitation (Pm), Evaporation From VIC (EVIC), From the Eta Model (Em), and From the Water Balance Equation (Eres), Eta Model Runoff (Rm), and VIC Runoff (RVIC)
| ||Value, mm day−1|
| Eres = Po − MFC||1.33|
| RVIC (6/1998–5/2000)||0.50|
3.3. Land Surface Energy Balance
 The Eta model is known to have a positive bias in downward shortwave radiation at the surface [see, e.g., Berbery et al. 1999], and the results here verify it. The other surface radiation terms compensate such bias, and consequently the surface energy balance is closer to observations and other models [Berbery et al., 1999]. Figure 9 compares the downward shortwave radiation of the Eta model (Figure 9a) with that estimated from GOES satellite products derived by Pinker et al.  (Figure 9b). Differences in the southwestern United States are of the order of 10%, but increase toward the northeast, achieving a bias of 30–40% near the Great Lakes and eastward (Figure 9c). This bias is present throughout the year, but is largest during summer (Figure 9d).
Figure 9. The 1998–2001 summer (MJJA) mean fields of downward shortwave radiation estimated from (a) the Eta model 12–36 h forecasts and (b) from GOES satellite products, (c) their ratio, and (d) their mean annual cycle.
Download figure to PowerPoint
 The terms of the summer (defined as May through August) mean surface energy balance are presented in Figure 10. Model constraints mandate that the net radiation flux at the surface be mostly compensated by sensible, latent and ground heat fluxes (plus other minor terms). Net radiation (Figure 10a) shows largest gain (∼200 W m−2) over the eastern part of the country, with a weak gradient and minimum values (∼120/150 W m−2) toward the western U.S. Loss of energy at the surface is mostly partitioned between sensible heat and latent heat fluxes, with a minor contribution of the ground heat flux. The loss of energy by sensible heat (Figure 10b) is largest in the western semiarid regions and other areas with reduced clouds. Minimum values between −40 and −20 W m−2 are noted south of the Great Lakes, over the Upper Mississippi and Ohio basins. In general, within the Mississippi basin, values do not exceed −80 to −100 W m−2. The latent heat flux (Figure 10c) has largest values toward the east, where there is more moisture availability, and smallest toward the west. As a result of the opposing gradients, the Bowen ratio (Figure 10e), is less than one over most of the Mississippi, reflecting the dominance of the latent heat, and increases toward the semiarid regions of the southwestern United States and northern Mexico. Over desert regions it exceeds 10, highlighting the different climate regimes. The ground heat flux (Figure 10d) is typically one order of magnitude less than the other terms, and therefore is a small part of the surface energy balance.
Figure 10. The 1998–2001 summer (MJJA) mean fields of (a) net radiation flux, (b) sensible heat flux, (c) latent heat flux, (d) ground heat flux, and (e) Bowen ratio. The time series of the energy terms is presented in Figure 10c.
Download figure to PowerPoint
 Figure 11a presents the mean annual cycle of the surface energy terms area-averaged for the Mississippi basin. Latent heat exceeds the magnitude of the sensible heat at all times, and during summer is about double. Therefore June-July values of net radiation, which are about 170 W m2, are balanced two thirds by the latent heat (∼100 W m−2), somewhat less than one third by the sensible heat (∼60 W m−2) and less than 10 W m−2 by the ground heat flux. The Bowen ratio (Figure 11b) increases during January–March due to the increase of sensible heat, and then remains nearly constant during spring and summer. However, with the progress of the warm season and the drying of the vegetation, there is a reduction of the latent heat (evapotranspiration) that results in a peak of the Bowen ratio in September. The ratio's negative values during winter imply that the sensible heat is toward the surface (as seen in Figure 11a). The time series of the same variables (Figure 11c) do not depart significantly from the mean annual cycle, and are consistent with the time series presented in Figures 2 and 8. The summer of 1999 had the largest precipitation (Figure 2) and accordingly the latent heat was largest and sensible heat small. Conversely, precipitation during the summer of 2000 was smaller, and consequently sensible heat is larger and latent heat smaller.
Figure 11. June 1998 to May 2002 mean annual cycle of the Mississippi basin area-averaged (a) surface energy balance terms (b) and the Bowen ratio. (c) The time series of the energy terms. NRF stands for net radiation flux, SHF stands for sensible heat flux, LHF stands for latent heat flux, and GHF stands for the ground heat flux.
Download figure to PowerPoint