To examine the impact from land model upgrades and different land initializations on the National Centers for Environmental Prediction (NCEP)’s Climate Forecast System (CFS), extensive T126 CFS experiments are carried out for 25 summers with 10 ensemble members using the old Oregon State University (OSU) land surface model (LSM) and the new Noah LSM. The CFS using the Noah LSM, initialized in turn with land states from the NCEP–Department of Energy Global Reanalysis 2 (GR-2), Global Land Data System (GLDAS), and GLDAS climatology, is compared to the CFS control run using the OSU LSM initialized with the GR-2 land states. Using anomaly correlation as a primary measure, the summer-season prediction skill of the CFS using different land models and different initial land states is assessed for SST, precipitation, and 2-m air temperature over the contiguous United States (CONUS) on an ensemble basis.
Results from these CFS experiments indicate that upgrading from the OSU LSM to the Noah LSM improves the overall CONUS June–August (JJA) precipitation prediction, especially during ENSO neutral years. Such an enhancement in CFS performance requires the execution of a GLDAS with the very same Noah LSM as utilized in the land component of the CFS, while improper initializations of the Noah LSM using the GR-2 land states lead to degraded CFS performance. In comparison with precipitation, the land upgrades have a relatively small impact on both of the SST and 2-m air temperature predictions.
Skillful short-term weather forecasts, which rely heavily on quality atmospheric initial conditions (ICs), have a fundamental limit of about two weeks (Lorenz 1963) owing to the chaotic nature of the atmosphere. Useful climate forecasts on seasonal time scales, on the other hand, require well-simulated large-scale atmospheric response to slowly varying lower-boundary forcings from both ocean and land surface. The critical importance of ocean memory has been well recognized (e.g., Shukla 1998; Wallace et al. 1998), whereby large-scale anomalies in the atmospheric general circulation on seasonally averaged time scales are forced first and foremost by large-scale anomalies in sea surface temperature (SST), especially over the El Niño–Southern Oscillation (ENSO) regions of the tropical Pacific Ocean. Because of the critical role of SST in coupled seasonal climate forecast models, ENSO SST indices have been widely applied in many statistical seasonal climate forecast applications for a long time, for example, in the 30- and 90-day seasonal outlooks over the United States issued by the Climate Prediction Center (CPC) at the National Centers for Environmental Predictions (NCEP). In contrast to SST anomalies, it has proven notably more difficult to demonstrate that land surface anomalies (soil moisture, snowpack) have meaningful positive impact on continental seasonal forecast skill in either coupled climate models or statistical seasonal forecast techniques. Past studies show that soil moisture anomalies can persist for months (Vinnikov et al. 1996) and soil moisture feedback can have notable effects on precipitation and modify other quantities such as 2-m air temperature through surface evaporation and surface energy processes (e.g., Shukla and Mintz 1982; Wu and Dickinson 2004). The feedback was also found to vary with climate regimes (e.g., Koster et al. 2000; Zhang et al. 2008). Because of the coupled nature of land–atmosphere interactions, these studies suggest that careful treatment of soil moisture and its associated anomaly in coupled climate models is important to improving skill of seasonal predictions, especially during summer times when the SST signal is relatively weak.
Because of a paucity of global soil moisture observations and the complex nature of land–atmosphere interactions, land-anomaly forcing is more difficult than ocean-anomaly forcing to separate from the natural chaotic variability of seasonal circulations as the land-anomaly impact has a smaller signal-to-noise ratio. Efforts to understand the linkage between land surface anomalies and the spawning of seasonal atmospheric circulation anomalies have to rely on developing more advanced land surface models. A better representation of land physics in climate models becomes the first step toward understanding how much the land contributes to climate variations. The efforts to understand the complex land–atmosphere interactions are also compounded by the fact that the prediction results from a given general circulation model (GCM) are sensitive to how the land component of the GCM is initialized and the starting dates used in the integrations (e.g., Dirmeyer 2001; Koster et al. 2000, 2006). Therefore, harnessing the impact of land surface anomalies for seasonal predictions is a promising challenge that requires not only a large number of members in the ensemble set of seasonal predictions (e.g., Tribbia and Baumhefner 1988; Brankovic et al. 1994) but also special care in the treatment of land surface initial conditions (e.g., initial soil moisture). The treatment becomes increasingly important at higher latitudes, such as in the contiguous United States (CONUS) where the soil moisture feedback was found to account for more variance of monthly precipitation anomalies (Zhang et al. 2008). The study by Koster et al. (2004a) also suggested that a proper global initialization of soil moisture may enhance precipitation prediction skill during the NH summer season. The recent results from the second phase of the Global Land–Atmosphere Coupled Experiment (GLACE-2) demonstrated again that the realistic land surface initialization can contribute to the skill of subseasonal precipitation forecast up to 45 days (Koster et al. 2010).
In short, there are two pathways to better represent the land impact in seasonal climate prediction systems. One pathway is to advance the realism of the physical processes of the land surface component. A second pathway is to improve the specification of the initial conditions of the land states in the land surface model (LSM). To demonstrate the importance of each pathway in the operational coupled seasonal forecast model at NCEP, we proceed to improve the land representation in the global Climate Forecast System (CFS) along each of the two pathways.
First, we upgrade the Oregon State University (OSU) LSM (Pan and Mahrt 1987) used in the current operational CFS with the Noah LSM (Ek et al. 2003). The Noah LSM is equipped with more advanced land physics compared to its ancestor OSU LSM, including four soil layers, snowpack and frozen soil physics (Koren et al. 1999), and snow cover–weighted surface fluxes, among others (see Table 1 for the main characteristics of the two land models). The Noah LSM was implemented in NCEP’s operational medium-range Global Forecast System (GFS) in late May 2005 and thus is expected to be the land component of NCEP’s next generation CFS for seasonal predictions.
Second, we take special care in the treatment of initial land states. To provide the CFS with optimum land states, we embrace the view that the best approach for providing global- or continental-scale analyses of soil moisture and snowpack is to execute a continuously cycled, multiyear, global, uncoupled, temporal integration of a land surface model forced by global analyses of observed atmospheric land surface forcing, especially global analyses of observed precipitation and satellite-observed surface solar insolation (Dirmeyer et al. 1999; Mitchell et al. 2004; Rodell et al. 2004; Koster et al. 2004a). This approach is well illustrated in the Global Soil Wetness Project (GSWP) of the late 1990s (Dirmeyer et al. 1999). Such approaches have become known as Land Data Assimilation Systems (LDAS) that do not actually assimilate observations to directly update the land states but rather let the LSM-simulated land states physically evolve freely in response to the external analyses of near-surface atmospheric forcing, especially the precipitation analysis. Also, the LSM of any given global GCM has its own inherent annual-cycle climatology of soil moisture (and snowpack at high latitudes) over each region of the global landmass, and the inherent climatology in one model can be quite different from another (Dirmeyer et al. 2006; Koster et al. 2009). Therefore, the optimal land state initial conditions for the land component of a given GCM should be generated by long-term multiyear executions of the given LSM in the LDAS mode. Long-term executions are essential to allow for the multiyear spinup time scales of deep soil moisture. In addition to the treatment of initial land states, to achieve meaningful results, the CFS must be executed over the season of interest for multiple years (even multiple decades) to allow investigators to determine the model climatology and in an ensemble mode to better enhance the signal from the noise.
Following the above strategy, section 2 first describes the approach to generate optimal initial land states for the CFS and then describes the set of CFS summer-season forecast experiments to examine to what degree each component (LSM upgrades, initial land states) can contribute to improving the CFS seasonal predictions. Section 3 presents the results, and the main conclusions and brief discussions are given in section 4.
2. Experiment design
a. GLDAS experiments
As a prerequisite to the T126 resolution CFS experiments to be carried out (described in section 2b below), a Global Land Data Assimilation System (GLDAS) is constructed and executed over the 25-yr period of 1980–2004 using the Noah LSM as the land model and on exactly the same computational native grid as that of the T126 CFS. In so doing, we take care to use exactly the same terrain field and land mask as in the T126 CFS, as well as all the same specifications of land surface characteristics (soil class, vegetation class, etc.), land physical parameters, and same version of the Noah LSM (2.7.1) as employed in the CFS–Noah experiments. This T126 GLDAS–Noah suite provides both the instantaneous and daily climatological GLDAS–Noah land states (averaged over the 25 yr) used in the CFS experiments.
The precipitation forcing for the GLDAS–Noah executions is from the CPC’s Merged Analysis of Precipitation (CMAP) pentad (5-day) analysis of observed precipitation (Xie and Arkin 1997), partitioned to 6-hourly amounts using temporal weights computed from the global precipitation fields of the NCEP–Department of Energy (DOE) Global Reanalysis 2 (GR-2) (Kanamitsu et al. 2002). The GR-2 is a coupled atmosphere–land data assimilation system, wherein the land component is driven by model-predicted precipitation, applies the OSU LSM with two soil layers, and nudges soil moisture based on the differences between model and the CMAP precipitation. Aside from the CMAP-anchored precipitation forcing, all remaining land surface forcings for the GLDAS–Noah executions are taken from the GR-2. These GR-2 surface forcings are adjusted for the terrain height differences between the T62 GR-2 and the T126 CFS using the adjustment algorithms that have been extensively and widely applied in the Environmental Modeling Center (EMC)’s longstanding North American Land Data Assimilation System (NLDAS) (Mitchell et al. 2004). Additional details and background of the NCEP–National Aeronautics and Space Administration (NASA) GLDAS are given in Rodell et al. (2004). Focusing on the Great Plains of the CONUS, Koster et al. (2004b) showed that the GLDAS land initialization does lead to a small but statistically significant improvement in summertime monthly precipitation forecast. However, its impact on seasonal predictions within a coupled ocean–land–atmosphere GCM has yet to be examined and is to be demonstrated here.
b. CFS experiments
The CFS used in our experiments is a modified version of the CFS described in Saha et al. (2006). The atmospheric component of the CFS is taken from a recent operational version of the NCEP medium-range GFS (August 2007). Key atmospheric physical parameterizations of this GFS include the simplified Arakawa–Schubert convection, a Rapid Radiative Transfer Model for longwave and the radiative transfer parameterization from Hou et al. (2002) for shortwave radiation, explicit cloud microphysics, nonlocal vertical diffusion, and gravity wave drag. The ocean model in the CFS is the Modular Ocean Model version 3 (MOM3) from the Geophysical Fluid Dynamics Laboratory (GFDL). As in all our CFS experiments here, the next generation of operational CFS at NCEP will execute at T126 horizontal (spectral) resolution and with 64 vertical sigma layers (L64) (higher than the currently operational CFS at T62/L28).
The four CFS experiments in this study differ only in the land component, while sharing the same resolution (T126, L64) and atmospheric and oceanic components described above. Three CFS experiments run with the Noah LSM (CFS–Noah) but initialized with land states in turn from 1) GR-2 (CFS–Noah–GR-2; hereafter referred to as case B), 2) GLDAS (CFS–Noah–GLDAS; hereafter referred to as case C), and 3) climatological GLDAS (CFS–Noah–GLDAS Climo; hereafter referred to as case D). In contrast to cases A and B, where the Noah LSM is initialized with instantaneous land states, soil moistures in case D are initialized with the GLDAS–Noah daily climatology but are allowed to evolve freely during the forecast. It sets out to examine the impact of soil moisture interannual variability. The fourth CFS experiment designated the control run uses the old OSU LSM (CFS–OSU) and the GR-2 initial land states (CFS–OSU–GR-2; hereafter referred to as case A) as in NCEP’s currently operational CFS. All together, these four experiments should demonstrate 1) whether the replacement of the OSU LSM with the Noah LSM yields meaningful improvement and 2) the impact of different Noah LSM land initializations on CFS summer-season predictions. The CFS seasonal forecast experiments are carried out for 25 summers (1980–2004), each with 10 ensemble members whose initial dates are from 19–23 April and 29 April–3 May with initial time of 0000 UTC. The time and dates of the ICs follow those used in NCEP’s existing hindcast database of the presently operational T62 CFS. In all four experiments, the atmospheric initial conditions come from the GR-2, and the oceanic initial conditions are taken from NCEP’s MOM3-based Global Ocean Data Assimilation System (GODAS) (Ji et al. 1995).
The results are presented in two main segments. First, we compare the GLDAS and GR-2 land states used in the CFS experiments. Second, we assess and compare the skills in CFS cases A–D in predicting CONUS June–August (JJA) anomalies over the entire 25-yr period. The predicted anomaly fields examined here are SST, precipitation, and 2-m air temperature, as they are the main products from CPC’s seasonal outlooks. In addition to the overall 25-yr evaluation, the CFS performance for precipitation and 2-m air temperature is also evaluated by contrasting the ENSO nonneutral and neutral years. The verification datasets are the NLDAS ⅛th degree (interpolated to T126 Gaussian grid) gauge-only analyses (Higgins et al. 2000; Mitchell et al. 2004) for precipitation, the global monthly surface analyses (Fan and van den Dool 2008) for continental 2-m air temperature, and the daily optimum interpolation (OI) analyses with 1° × 1° (Reynolds et al. 2002) for SST (interpolated to T126 Gaussian grid), respectively. The CFS skill is measured by the anomaly correlation (AC), which is defined as the correlation between the CFS predicted anomaly (defined with respect to the corresponding 25-yr CFS climatology) of the forecasted variable and the corresponding observed anomaly across the 25 forecasted summers. The CFS seasonal forecast AC skill is a key metric used to determine CFS forecast influence on official CPC seasonal outlooks. In addition to the main AC evaluation, two supplemental measures (defined in section 3b below) are used in the assessment of CFS performance for precipitation and 2-m air temperature.
a. Comparison of GLDAS and GR-2 soil moisture
We first compare the soil moisture of the 25-yr retrospective T126 GLDAS–Noah, used in CFS cases C and D, with that of the GR-2 used in CFS cases A and B. Figure 1 presents global maps of the 1 May climatology and 1 May 1999 anomaly of the 2-m soil moisture field from both GLDAS–Noah and GR-2–OSU. Figure 2 is as in Fig. 1, but zoomed in over the CONUS domain. Both the GLDAS and GR-2 simulate the soil moisture of a 2-m soil column, albeit with four soil layers in GLDAS–Noah and only two layers in GR-2–OSU. We choose to depict 1 May conditions, as that time is within the 19 April–3 May period from which the initial conditions are taken for the 10 members of the summer CFS hindcasts. Comparison between the two sources of soil moisture shows that the climatological values in early May of the GR-2–OSU soil moisture are consistently lower than those of GLDAS–Noah over the vast majority of both the global and CONUS domains (with some exceptions).
The anomaly fields of GLDAS–Noah and GR-2–OSU in Fig. 1 show some similarities in some regions of the globe—for example, South America, Africa, Australia, and Southeast Asia—but rather substantial differences over other regions, such as southwest Asia, Europe, and North America. Focusing on the latter substantial differences over CONUS, Fig. 2 clearly reveals the large disagreement over CONUS between GLDAS–Noah and GR-2–OSU soil moisture anomalies for the case of 1 May 1999. For instance, the GLDAS–Noah has much larger positive anomalies over Oklahoma, Kansas, Missouri, Arkansas, and Illinois, whereas the GR-2 shows a substantial negative anomaly over western portions of Montana, Wyoming, Colorado, New Mexico, and eastern Utah. To examine how well each of the two data assimilation systems performs, Fig. 3 shows the observed precipitation anomaly for the 90-day period ending on 1 May 1999 and its climatology (the same time of 1 May 1999 anomaly depicted in Figs. 1 and 2). The GLDAS–Noah soil moisture anomaly in Fig. 2 shows better spatial agreement with the observed precipitation anomaly than does GR-2–OSU. This is not surprising because the GLDAS–Noah, unlike the GR-2–OSU, uses an analysis of observed precipitation to force the land surface, whereas the GR-2–OSU uses the model precipitation from the background global model of the assimilation system. To compensate for use of model precipitation, over 5-day intervals, the GR-2–OSU computes the errors in the model precipitation compared to the observed precipitation analysis (the same precipitation analysis used directly by GLDAS–Noah) and then applies a soil moisture nudging scheme to try to offset the effects of the model’s precipitation errors; however, Fig. 2 suggests that the GR-2–OSU nudging scheme is not effective.
While Figs. 1, 2, and 3 are for a particular time of year, Fig. 4 shows the multiyear time series and climatological annual cycle of monthly soil moisture from both the GLDAS–Noah and GR-2–OSU for the 2-m soil moisture spatially averaged over Illinois with comparison to those derived from the Illinois network of soil moisture observations (Robock et al. 2000). Clearly in Fig. 4, the soil moisture climatology of GR-2–OSU remains consistently lower than that of GLDAS–Noah on the month-to-month and year-to-year basis. Moreover in Fig. 4, the higher GLDAS–Noah soil moisture is in closer agreement with the observations than that of GR-2–OSU, though still exhibiting some low bias, but substantially less so than GR-2–OSU. One key implication to be drawn from Fig. 4 and Figs. 1 and 2 (climatology for 01 May from two sources of soil moisture) is that initializing the Noah LSM component of CFS from the notably different land states of the GR-2–OSU dataset (e.g., soil moisture anomaly and climatology) will result in the incorrect interpretation of GR-2 soil moisture in the Noah model, a situation that is likely a significant contribution to the poor performance of CFS–Noah–GR-2 case (case B) presented later.
b. Results of CFS experiments
SST serves as a slowly varying state predicted by the coupled CFS. The CFS predicted change in SST has an important impact on the CFS seasonal atmospheric predictions because of SST’s large thermal inertia. SST anomalies over in the tropical Pacific Ocean, particularly over the Niño-3.4 region (defined as 5°S–5°N, 120°–170°W), have traditionally been recognized to have substantial influence on the U.S. summer-season predictions. The SST has been one of the most skillful forecast fields in the current operational CFS. Thus, we want to illustrate if the land-related changes in CFS impact this critical CFS forecast entity by starting with the examination of CFS skill in predicting SST anomalies. Figure 5 presents the 25-yr time series of CFS predicted mean JJA SST anomalies (with respective to each CFS case’s own climatology) over the Niño-3.4 region with comparison to the observations. As might be expected, because of the large body of water and its long thermal persistence, Fig. 5 shows that the differences in CFS predicted SST anomalies over the Niño-3.4 region (and other tropical regions; not shown) over each individual year are very small (as are differences in predicted Niño-3.4 mean SST climatology—less than 0.05°C; not shown). Most years are in reasonably good agreement with the observations (with some exceptions). With these similarities in predicted SST among the four CFS cases, the divergence in CFS performance over land shown in the next section is likely due to the different overland conditions and land fluxes arising from using different land models and different initial land states.
The current operational CFS (using an old atmospheric component but with the same OSU LSM and ocean model used in this study) manifests considerably useful skill in predicting the tropical Pacific SST (and maintained here in the four CFS experiments) but showed virtually no skill for predicting anomalies of summer-season precipitation over CONUS. To examine if the land upgrades in the CFS achieve meaningful improvements in precipitation skill, we focus on evaluation of the CONUS-averaged AC score (hereafter referred to as AAC, a single value obtained from averaging the AC score on each grid point over entire CONUS domain), which is used to assess the overall performance from the CONUS point of view. We also present CONUS map of precipitation AC score (a map of AC score is hereafter referred to as “skill mask”) to illustrate geographical patterns of the precipitation AC score (range −1 to +1) and the percentage count of CONUS grid points (out of a total of 989 CONUS grid points on the T126 Gaussian grid) with positive AC scores (range from 0 to 100; hereafter referred to as PAC). Essentially, the AAC score measure treats the CONUS as one grid point, and the skill map shows its subgrid variability. The purpose of the PAC measure is to do a check on the AAC score to make sure that it indeed represents a majority of grid points, as a few significantly large positive (negative) AC grid points could contribute to higher (lower) AAC scores. Both the AAC and PAC measures are derived from the AC skill mask and used in both the precipitation and 2-m air temperature evaluations below.
Figure 6a presents the CONUS AC skill mask for JJA ensemble mean total precipitation of the 25-yr reforecasts from the four CFS configurations. In Fig. 6a, both cases C and D with the Noah–GLDAS land states appear to have a somewhat larger area of high AC scores (above 0.4) compared to the other two, and case B is obviously far inferior to the other three. Also, each case yields different preferred regions for better performance, with cases C and D showing a tendency toward high scores in a majority of Pacific Northwest states and northern Great Plains, while case A appears to yield somewhat better AC scores in the southwest monsoon region and the states that border the Gulf of Mexico. The PACs are 64.9 in case C, 58.7 in case D, 58.4 in case A, and 44.3 in case B. Case C has the highest count among the three Noah configurations. Although the large differences in spatial distribution of the preferred regions among the four CFS configurations can be linked to different land models and different land initializations, it should be pointed out that the sampling (uses 10 ensemble members and is confined to the 25 years) may also contribute to the disparities as nature only gives one realization each year for comparison. It becomes even more important for the contrasted ENSO and ENSO-neutral years (discussed later) where the covering periods are much shorter.
Using a different measure, the bar chart of AAC in Fig. 6b clearly shows that case C with both the Noah model upgrade and inclusion of the GLDAS–Noah yields the highest value, indicating that upgrading of the OSU LSM–GR-2 combination to the Noah LSM–GLDAS combination does improve the overall CONUS precipitation prediction. Strikingly, case B, where the CFS–Noah model configuration uses the GR-2 initial land states, yields the lowest AAC value by far (essentially zero), indicating that the CFS performance can be quite different even with the same new land model but initialized with incompatible land states generated from a different land model. As shown in Figs. 1–3, the differences in both the climatology and spatial anomalies of soil moisture are quite large between the GR-2–OSU and GLDAS–Noah land states, with GLDAS–Noah having higher soil moisture in the warm season over most nonarid regions worldwide, than that of GR-2 with the OSU LSM.
We emphasize that even though case C yields the highest AAC score in Fig. 6b, its value is relatively low (of order 0.084). As also shown in Fig. 6a, over the regions of positive AC, the values usually do not exceed 0.4. Hence, the seasonal prediction skill of summer precipitation over CONUS is still a major challenge, needing much more effort (possibly in GFS physics unrelated to the land surface).
To provide insight into the CFS performance with the four configurations in different ENSO SST regimes, the 25 years are stratified into ENSO nonneutral and neutral years using the observed May–July (MJJ) Niño-3.4 SST anomaly magnitude of 0.7°C as the threshold for nonneutral years (slightly larger than the commonly used 0.5°C threshold, to better capture stronger ENSO events). As a result, the 25 years are split into 10 nonneutral and 15 neutral years. The nonneutral years are 1982, 1983, 1987, 1988, 1991, 1992, 1993, 1997, 1999, and 2002 (among which, only 1988 and 1999 are cold ENSO years, and the rest are warm ENSO years), and the neutral years are 1980, 1981, 1984, 1985, 1986, 1989, 1990, 1994, 1995, 1996, 1998, 2000, 2001, 2003, and 2004, respectively. The CFS prediction skill from the land upgrades is next separately assessed for the nonneutral and neutral years.
Figure 7a presents the CONUS AC skill masks of JJA mean precipitation computed from the 10 nonneutral years. As might be anticipated, the CFS performs much better compared to the corresponding 25-yr averages owing to strong ENSO signals. As shown in Fig. 7a, the four CFS configurations appear to yield roughly similar geographical patterns, with positive scores extending from the Midwest region all the way to the Pacific Northwest states (with some exceptions). As demonstrated in past studies (see references in the introduction), large-scale anomalies in the atmospheric general circulation are strongly driven by SST anomalies. For the 10 nonneutral years, all four AAC scores in Fig. 7b have increased by similar amounts of about 0.07–0.09 compared to their respective AAC scores in Fig. 6b. Given the similarities in predicted SST among the four cases (Fig. 5), it is not surprising to see the consistently better AC scores across all four cases in Figs. 7a and 7b, with the 2 Noah–GLDAS configurations having slightly higher counts (68.6 in case C and 69.6 in case D) than the other two (66.9 in case A and 62.3 in case B). Interestingly, case D where the Noah LSM is initialized with the climatological GLDAS land states has the best performance, indicating that during the 10 ENSO nonneutral years the instantaneous GLDAS land states do not have a distinct advantage as long as the land states used are land model compatible and self consistent. Despite the skill gain in precipitation prediction from the land upgrades, the Student’s t test indicates that none of the differences in either the 25-yr averages (Fig. 6b) or the 10 nonneutral years (Fig. 7b) is significant at the minimum 90% confidence level.
While Figs. 7a and 7b are for the 10 nonneutral years, Figs. 8a and 8b are as in Figs. 7a and 7b but for the 15 neutral years. Compared to the nonneutral years in Fig. 7a, the CONUS skill masks of precipitation of Fig. 8a illustrate that the skill of all four CFS configurations decreases dramatically with the weak ENSO signals and the degradations mainly occur over most of the relatively drier Midwest region and the Pacific Northwest states. Nevertheless, among the four cases in neutral years, case C (with Noah LSM and GLDAS) exhibits the best. The PAC values in the case of neutral years are 55.9 in case C, 43.9 in case A, 42.1 in case D, and 39.0 in case B. Hence, case C has the largest area of positive AC scores, and the Noah–GR-2 CFS has the smallest area. The AAC values in Fig. 8b further show that case C is the only configuration that has a positive AAC value. Moreover, compared to case D where the climatological GLDAS land states are used case C shows clear advantages, suggesting that the land-anomaly initial state’s forcing contributes more to seasonal prediction skill in neutral years. Therefore, providing the CFS with initial land state anomalies consistent and exactly matching the external forcings appears most advantageous when the ENSO signal is weak. Admittedly, the positive AAC value even in case C is low, yet the Student’s t test (one-tailed) performed indicates that the differences between the third bar (in Fig. 8b; Noah–GLDAS) and the other three bars are statistically significant at the 90% confidence level.
3) 2-m temperature
Similar to the current operational CFS, the experimental CFS in all four configurations has better performance in 2-m air temperature prediction than precipitation prediction. Figures 9a and 9b are as in Figs. 6a and 6b, but for JJA averaged 2-m air temperature. The CONUS skill masks of Fig. 9a show that the four configurations have much better skill (AC values above 0.2 over a majority of the CONUS) than that of precipitation and are in closer agreement in their spatial patterns, although case D appears to have a slightly larger area of high AC scores over the northern Great Plains among the three Noah configurations, and case A performs somewhat better over the Gulf states than the other three. As indicated from their PACs (98.1 in case A; 96.7 in case D; 93.4 for case C; and 90.1 in case B), case A, where the OSU LSM was initialized with its compatible GR-2 land states, has the highest count, and the differences among the three Noah configurations are small. With these indications from both skill mask and PAC measures, it is not surprising to see their closer AAC values shown in Fig. 9b, where case A is slightly higher than both cases C and D, indicating that, on average, the land upgrades and the CFS with different land initializations with the Noah model have no clear advantages for the 2-m air temperature prediction.
We also separately examine the 2-m air temperature skill during the ENSO nonneutral and neutral years. Figures 10a and 10b present the skill masks and their respective AAC values for the 10 nonneutral years. Figure 10a illustrates that the 2-m temperature AC scores are much higher over most of the CONUS (more than 0.6) than their 25-yr averages when the ENSO signals are strong. All four CFS configurations have gains over most of the eastern CONUS while maintaining good performance over the western United States (with some exceptions) and have closer spatial agreements (compared to those in Fig. 9a), where all four cases show good performance over the central and north-central United States and somewhat degraded performance over the south-central, southeastern, and northern Midwest regions. The differences are only in the degree of gain or degradation. The performance with the three Noah configurations is very close (albeit case B is again the worst), and although case D appears to have a larger area of high AC scores over the north-central regions than the other two, this gain is almost entirely countered by worse performance over the southern Great Plains areas as reflected in Fig. 10b, where case D only shows a slightly higher AAC value. The PACs are 94.5 with OSU–GR-2, 92.7 with Noah–GLDAS, 89.2 with Noah–GR-2, and 87.1 with the climatological GLDAS land states, respectively. Consistent with their 25-yr averages, case A has the best performance, indicating that the benefits from land upgrades in the CFS may not be easily realized when the ocean signal is strong.
Figures 11a and 11b are as in Figs. 10a and 10b, but for the 15 ENSO-neutral years. Similar to what is seen in the precipitation prediction, these scores are much lower than either of the 25-yr averages or those obtained from the nonneutral years, and the degradation is also area dependent. Figure 11a illustrates that case C has the largest area of positive AC scores over states surrounding the Great Lakes (which both cases A and D do not have), case D appears to have higher negative AC scores over the southern Midwest and northeastern states, whereas case B has a relatively stable performance over the entire CONUS. Among the three CFS–Noah configurations, case C has the highest PAC value of the four cases (72.2 for Noah–GLDAS, 67.8 for OSU–GR-2, 66.5 for Noah–GR-2, and 52.1 for Noah–GLDAS climatology), although its AAC just shows a marginal improvement compared to case A in Fig. 11b. Even with the skill gains with case C (the highest PAC and AAC), the Student’s t test performed on the data indicates that all the differences in 2-m air temperature AAC score (in Figs. 9b, 10b, and 11b) are not statistically significant at least at the 90% confidence level.
4. Conclusions and discussions
Because of the coupled nature of land–atmosphere–ocean interactions, it is very difficult to separate contributions from each component to seasonal prediction skills. Focusing on the CONUS domain, the present study uses a controlled approach to examine the impact from the land model upgrade and the use of different initial land states on summer-season predictions in the next generation of NCEP CFS. Our results show the positive impact on summer-season precipitation prediction from the combined use of the Noah LSM and its self-consistent GLDAS initial land states. Initializing the Noah land model with the relatively drier GR-2 soil moisture negatively impacts the CFS performance, although the Noah LSM has better land parameterizations. This is particularly important as the GR-2 land states are being used in initializing the current operational CFS for seasonal predictions. With implementation of the Noah LSM in the next generation of CFS, this study demonstrates that the companion Noah–GLDAS is necessary to realize the benefits from the land model upgrade. The small but statistically significant improvement on precipitation prediction skill in case C over case D during the ENSO-neutral years indicates that the realistic initialization of soil moisture in case C contributes more to seasonal predictability.
The small differences in the CFS 2-m temperature prediction skill among the four configurations suggest that the land–atmosphere coupling is weak in the CFS. Similar results were also found in Lu et al. (2005) where they found that the NCEP GFS is one of the models that have weak land–atmosphere interactions among the 12 participating GCMs in the first phase of the GLACE project (Koster et al. 2006), and the causes are likely due to the treatment of boundary layer mixing and moisture convection in the GFS.
For operational purposes, the variables examined here are limited to the equatorial Pacific SST, precipitation, and 2-m air temperature over the CONUS. More analyses of the differences among the four configurations in atmospheric general circulation anomalies, surface water and energy components, other ocean model indexes, and possibly over different geographical regions would be desirable to further our understanding of land impact on seasonal predictions and could serve as a basis for future Noah land model upgrades.
We thank Drs. Suranjana Saha, Cathy Thiaw, and Hua-Lu Pan for providing the CFS script and help with its initial setups, Drs. Huug van den Dool and Kingtse Mo for their constructive suggestions on analyzing the results for ENSO nonneutral and neutral years separately, and Drs. Helin Wei and George Gayno for providing utility programs used to generate the initial land states for the CFS runs. The careful reading and critical comments by Drs. Randal Koster and Dennis Lettenmaier helped improve the quality of this article and are greatly appreciated. This study was supported by the NCEP Core Project component of the NOAA Climate Program Office–Climate Prediction Program for the Americas (CPPA).