This paper describes experiments with the National Centers for Environmental Prediction (NCEP) regional Eta Model, which runs operationally at the South African Weather Bureau. Experiments were designed to assess manual methods of improving the model representation of the surface topography at low resolutions where this might be feasible. The manual terrain modifications, applied in the then-operational 80-km Eta Model, involved either indirect alteration of the terrain through judicious rearrangement of the model coordinate surfaces, or the direct modification of the model terrain to accommodate meteorologically important features of the actual topography not well represented in the model. Also studied was the effect on the model terrain of changes in the horizontal grid configuration, with a potential for exploitation at higher model resolutions. Trials at higher resolutions also became available as part of a stepwise implementation of an upgrade to the 80-km system. This upgrade contained the important model enhancements installed at NCEP in January 1996 (new soil hydrology, land surface and viscous sublayer parameterization, and revised turbulence scheme) and February 1997 (notably changes to the radiation package and land surface model). A case of light nocturnal rainfall over Gauteng, linked to a topographical trigger and poorly forecast on the day, was the focus of a study to demonstrate the viability of the methods. Specifically, the link between changes in model topography and in precipitation forecasts from the model was investigated.
As expected, the model enhancement contained in the upgrade package provided the greatest impact on the precipitation predictions, and higher horizontal resolutions generally led to better representations of the terrain and more realistic forecasts. Nevertheless, the possibility exists of improving the high-resolution model topography yet further through modifications to the grid structure. If, however, higher horizontal resolutions are not affordable, it seems from this prototype study that slight grid shifts or manual modification of the model terrain could be useful and viable options. There is some potential in the judicious choice of the vertical coordinate structure, although what is good for one region might prove bad for another. There seems to be more potential in the manual correction of meteorologically significant deficiencies in the model terrain. These approaches are likely to be beneficial also in the presence of higher vertical resolution, which in itself may not be sufficient to improve the topography. Altering the terrain by hand, even at low resolutions, is a laborious task and only feasible at low resolutions but with the Eta Model being run on smaller machines, in centers with limited resources, this option seems to deserve consideration.
The National Centers for Environmental Prediction (NCEP) eta-coordinate regional forecast model (Eta Model; Mesinger et al. 1988) was adopted for regional operational modeling over the contiguous United States in June 1993 (Rogers et al. 1998). In November 1993 the Eta Model, together with a matching data assimilation system, became operational also at the South African Weather Bureau (SAWB). This version provided regional, short-range forecast guidance to the Central Forecasting Office of the SAWB until March 1996 when a major upgrade was implemented. The studies described here relate to this 1996 version and to the following upgrade at the SAWB, which became operational in August 1998. (A further upgrade ensuring year 2000 compliance took place in November 1999.)
The distinctive eta vertical coordinate has the desirable feature that coordinate surfaces are very nearly horizontal. In conjunction with the eta coordinate the model embodies a step-mountain formulation in which the physical lower boundary, that is, the model terrain, is constrained to lie along reference interfaces of the chosen eta-coordinate layers of the model, leading to a similar simplicity in the lower boundary condition as inherent in terrain-following sigma coordinate systems. Velocity components are set to zero along the vertical edges of the terrain. A preliminary form of the step-mountain terrain is derived from a global 10-min raw topography dataset. The elevation assigned to a grid box is either an averaged elevation, or the highest elevation (silhouette), within the box depending on the characteristics of the detailed topography (see Black 1994, and the discussion in Mesinger et al. 1988). To conform with the step-mountain formulation, the preliminary elevation at each grid point is adjusted upward or downward to the height of the closest eta reference interface. During the process great care is taken not to allow stagnant valleys completely surrounded by zero wind points. This process of adjustment might be regarded as a source of inaccuracy, but in fact it lends itself to experimentation with a view to improvement, being highly dependent on the choice of the model grid structure. Thus, changes in vertical as well as horizontal resolution will change the model topography, as will shifts in the vertical distribution of eta layers and the positioning of the horizontal grid. Determining the sensitivity to model mountain geometry (Mesinger et al. 1988) would assist in “tuning” the topography (Gollvik 1999) for the particular region where the model is to be used.
In August 1997 a case presented itself as a likely candidate for experimentation along these lines. There was an unexpected, light fall of rain (the first rain of the season) over Pretoria at about midnight on 30 August, but the then-operational 80-km Eta Model (and the forecasters) failed to predict the correct position of this convective precipitation over the Gauteng province (the shaded area in the map shown in Fig. 1). Moist air was moving southward over the Northern province and then southeast over the North West province toward Gauteng. It was considered that orographic lift provided the necessary trigger for rain to occur over the Highveld in an area where none was predicted by the model. The model had the required flow of moist air, as well as the instability and the correct upper air structure, but predicted the rain in the wrong place. Since no dynamical or thermodynamical error could be found in the Eta fields to explain this, the suggestion was thus put forward (R. Petersen 1997) that a slight change in the Eta topography might improve the positioning of the predicted precipitation.
This case thus became the focus of a study in which either the distribution of eta levels would be adjusted to give a better resolution of the topography over the central plateau,which lies at about the 850-hPa level, or the Eta Model topography would be manually adjusted to correct obvious deficiencies that could have adverse meteorological consequences in the model. The resulting changes in the model topography were examined, first in relation to the true topography, but also in the light of their probable consequences in the prediction of precipitation. The boxed area in Fig. 1 (which contains the Gauteng province) indicates the area of interest for the study and will be referred to as the Gauteng area. The 1996-vintage 80-km Eta modeling system operational at the SAWB until August 1998 was used in the study.
At the time of this study an upgrade of the Eta modeling system was in progress. Steps in the upgrade procedure were designed to isolate the effects of increased horizontal and vertical resolution, as well as changes in model formulation and domain size. However, it was noticed that a different model terrain was generated in each of the six upgrade configurations, even where no change would have been expected. The reasons for the differences were investigated with a view to possible exploitation as methods of changing the model terrain. Once again, the resulting changes in the terrain were examined. Here effects such as increased horizontal resolution would naturally be dominant factors influencing the model forecasts. However, it was found possible to determine the influence of terrain changes on the prediction of the rainfall event through comparison of like configurations.
There have been several intensive precipitation evaluation studies involving the Eta Model. Some (e.g., Mesinger 1996; McDonald and Horel 1998; Colle et al. 1999) employ objective verification scores such as equitable threat scores; in others (e.g., Paccagnella et al. 1992; Gallus 1999), the evaluation is qualitative. Purely qualitative evaluation is used for the single case studied here, which serves as a prototype to show that changes to the model terrain, by manual modification or other means, are feasible and can have a sizeable effect on model predictions. Exploitation of such methods in order to find an optimal representation of the terrain could be a useful step toward the ultimate goal of improved operational forecasting, in particular of precipitation where improvement is crucially needed (see, e.g., Fritsch et al. 1998).
Section 2 provides a view of the Eta Model versions used in the study. Section 3 describes the design of the low-resolution pre-upgrade terrain experiments, and section 4 lists the upgrade configurations available. Some general characteristics of the various model terrains are discussed in section 5. Section 6 presents the actual weather situation on the case day. Case study results for the low-resolution pre-upgrade experiments are in section 7, with emphasis on the link between the topography representation and the predicted precipitation. Sections 8 and 9 focus on results from the upgrade configurations, and section 10 contains concluding remarks.
2. Model versions used in the study
In November 1993, the Eta Model within the NCEP Eta Data Assimilation System (EDAS) was installed at the SAWB (see Poolman et al. 1994). The data assimilation over a 12-h period had a 6-h cycle including a digital filter initialization step. The horizontal resolution was roughly 80 km, with 17 layers in the vertical, and model top at 50 hPa. Global guess and boundary fields here, and for the model versions described below, were 6-hourly, 12-h-old, coarse-resolution, pressure-level products of the NCEP global spectral model obtained from the Global Telecommunication System.
a. March 1996
A major upgrade, described by Black (1994) and Rogers et al. (1995, 1996), was implemented in March 1996 (see Riphagen et al. 1996) and was the operational regional model until August 1998. An earlier restriction disallowing single-gridpoint mountains (Mesinger et al. 1988) no longer applied, while the model top remained at 50 hPa, and “below surface” observations within 25 hPa of the Eta terrain pressure were permitted. The analysis now took place on both the horizontal and vertical grids of the Eta Model. In the forecast model the formulations of convective precipitation, surface layer processes, cloud cover, diffusion, and other processes were improved, and a cloud water model was included. At the SAWB the horizontal resolution remained 80 km but the number of layers was increased to 38, and the assimilation interval reduced to 3 h. Digital filter initialization was still included. In the following discussion runs with this model will be designated C8038.
b. August 1998
In August 1998 the next upgrade became operational at the SAWB (Riphagen et al. 1999). Important enhancements were contained in the package, comprising changes made at NCEP in January 1996 and February 1997. January 1996 saw the introduction (see Rogers et al. 1996; Betts et al. 1997) of new soil hydrology and land surface parameterizations (Chen et al. 1996), with a two-layer soil model and a seasonal vegetation cycle, to replace the previous bucket physics. Also implemented at NCEP at that time were a new viscous sublayer parameterization scheme (Janjić 1996a) and a revised version of the Mellor–Yamada level 2.5 turbulence scheme (Janjić 1996b). The NCEP February 1997 bundle of changes contained modifications to the radiation package and the land surface model designed (see Rogers et al. 1998; Black et al. 1997; Betts et al. 1997) to alleviate a problem noted of too large a diurnal cycle for model surface temperature, attributed to factors such as insufficient absorption of shortwave radiation, cloud underestimation, and a negative bias in surface evaporation.
The model top was now at 25 hPa. Time constraints dictated that the digital filter initialization step no longer be added at the SAWB. The stepwise implementation procedure prior to installation provided an opportunity for experimentation with different resolutions and domain sizes. In what follows, the initial U will be used to characterize variations of the upgrade modeling system, followed by numbers showing the horizontal and vertical resolution, and possibly a letter specifying the domain, for example, U4838S for the 48-km, 38-level upgrade model on the “small” domain (see Fig. 2). [The form of the model eventually chosen for operational implementation, as a good and affordable option, was in fact the U4838S system. This upgrade was accompanied by a sizeable decrease in rms errors (forecast vs own verifying analysis). For example, over comparable 6-month periods, the average rms error for 48-h forecasts of mean sea level pressure was reduced from 3.03 to 2.50 hPa, and for 500-hPa geopotential height from 3.41 to 2.50 dam.]
c. Special tools
The integration of the finite-difference equations making up the Eta forecast model takes place on a transformed Arakawa E grid (see, e.g., Mesinger et al. 1988; Black 1994) on the eta-coordinate surfaces of the model. At chosen output times, the values of model variables and derived quantities available at the E-grid points are interpolated (possibly with smoothing) to user-defined output grids defined (at the SAWB) in terms of the more familiar geodetic latitude and longitude coordinates, for use perhaps in graphical displays (which employ further interpolation and smoothing). During this process gradients may be smoothed out quite severely and thus tools were developed to view various quantities (such as terrain height and precipitation) on the E grid itself at a stage prior to the interpolation and contouring procedures. Such displays of the model terrain emphasize its discrete nature (as is the case in all finite-difference models), but also demonstrate that in the Eta Model case the heights of the terrain at the model grid points are drawn from a small set of values (around 20 in our applications), namely the heights of lower eta reference interfaces.
3. Design of the low-resolution pre-upgrade experiments
a. Experiment 1: C8038 (1) = C8038
The 80-km 38-level March 1996 Eta modeling system, operational at the SAWB when the tests were conducted, served as control for the low-resolution experimentation with the representation of model orography. A 24-h forecast was run from 0000 UTC 30 August 1997 to obtain predicted values at 0000 UTC 31 August when the rainfall was observed. (As for all the model runs described, the starting analysis was the end product of a 12-h data assimilation process involving the particular form of the model.) The model domain, as for all the 80-km configurations, was the “large domain” marked L in Fig. 2. The vertical structure is marked C38 in Fig. 3.
b. Experiment 2: C8038 (2)
Here the vertical structure of the model was manually rearranged to give more detail in a particular vertical region of the atmosphere (and to provide a larger pool of potential terrain elevations to draw from in creating the step mountains). For the basic C8038 system, in terms of height differences, the lowest layers were fairly thin with, for instance, depths of roughly 20, 40, 50, and 70 m, respectively, for the first four layers. However, by about the 850-hPa level, or 1500 m, the approximate height of the South African central plateau, the layer depths had increased to over 200 m, so that the topography of a major part of the country was not well represented. Thus for C8038 (2) the original C8038 eta levels were adjusted to allow increased resolution (depths of no more than 150 m) in this critical region, at the expense of lower regions, which for land areas may lie under the surface, and of regions higher up [see column C38(2) in Fig. 3]. The depths of the bottom four layers now became 25, 50, 80, and 110 m, respectively. Otherwise the experiment was identical to the control.
c. Experiment 3: C8038 (3)
It was suspected that aspects of the topography important for precipitation at Pretoria on the night of 30 August were poorly represented, or missing, in the C8038 run. Accordingly, while the preceding experiment sought a general improvement through changes in the vertical coordinate and hence indirectly in the model terrain, in this experiment a new model terrain was manufactured by directly altering the C8038 terrain heights by hand at certain critical grid points. In all, 18 blocks were altered through shifts to the elevations of higher or lower eta reference surfaces. Great care had to be taken to avoid the introduction of “windless valleys,” a laborious iterative task normally taken care of in the computer procedure to generate the model topography. The vertical structure was as for C8038 (see Fig. 3, C38).
4. Upgrade configurations available
A stepwise procedure, adopted in the process of reconfiguring the upgrade, provided the opportunity of conducting trials with various horizontal (80, 48, and 29 km) and vertical (38 and 50 level) resolutions, and on various domains, with the intention of isolating the various effects. For these runs, the model terrain was created using the standard NCEP code with no attempt at artificial manipulation. Figure 3 shows the various vertical structures, marked U38 or U50 as appropriate.
The basic upgrade (U8038) differed from C8038 in the NCEP enhancements (also of the orography generation) and the omission of the initialization stage, which had minimal effect on model forecasts to 24 h or so. U8050 differed from U8038 in the increase in vertical resolution, while U4838L differed from U8038 in the increase in horizontal resolution. The domain for all these configurations was the large domain marked L in Fig. 2.
At resolutions finer than 80 km, the large domain would have been too expensive in the operational environment, and thus a modelling system U4838S was set up, as U4838L, but on a smaller domain (marked S in Fig. 2). The small domain was inadvertently chosen with a different (geodetic) center than the large domain, resulting in a different placement of the grid points on the transformed grid, with important repercussions in the model terrain (see section 5a). Strictly for research and benchmark purposes at this stage because of computer limitations, a 29-km system U2938S was run on a domain virtually identical to that of U4838S, with the same center.
5. General terrain characteristics
a. Topography variations
The topographies of the various Eta Model configurations for the Gauteng area are shown in Figs. 4 and 5, with the true terrain in Fig. 6 for comparison. Uniquely, the Eta Model terrain depends directly on the vertical discretization and will vary with the number of eta levels, their placement, and the height of the model top. Even with the same vertical structure, features of the terrain may vary quite substantially with different block size in the horizontal, and with shifts in the position of the E-grid points, due to the adjustment of the terrain to the elevations of eta surfaces. For instance, the terrain may be adjusted upward at a location in one configuration and downward at a very slightly different position in another, depending on which eta surface is the closest at each of the points. Together with the valley-filling process this may result in the disappearance of entire mountain or valley systems.
Domains chosen with a different (geodetic) center (e.g., U4838L and U4838S) will naturally have noncoincident grid points seeing that the transformation formulas depend directly on the definition of this center. However, even with the same center, U2938S and U2938VS had diametrically staggered grid points in both directions, due to their odd and even choice, respectively, of the total number of zonal E-grid points. On identical horizontal grids, the unadjusted terrain (created from the fine-mesh topography) for C8038 differed from that of U8038 and U8050, due to technical changes in this creation procedure embodied in the NCEP upgrade package. Thus the terrain and land–sea mask were different for all nine Eta modeling systems considered. For instance, note the position of Durban (approximately 30°S and 31°E, denoted in Figs. 4 and 5 by a “station dot” on or near the southwest coast) relative to the terrain for the various options.
b. Adjustments to eta reference surfaces
Of concern are the magnitudes of the adjustments of the original topography to the heights of the reference levels of the model, and of the further adjustments needed to fill windless valleys. This is of special importance in the analysis procedure where surface observations located well below or above the model terrain may be rejected although of good quality. If accepted, risky extrapolation of the background field to below-surface locations may be necessary. [Operationally, with both the 1996 and 1998 model versions, roughly 100 out of 300 surface observations over South Africa were rejected for this reason at each main synoptic time. See Riphagen et al. (1999).] Problems also arise with verifications at such stations (see Nutter and Manobianco 1999; Yucel et al. 1998).
From the differences on the E-grid between the final terrain heights and the unadjusted heights (not shown) for all the configurations, it was found that changes were mostly reasonably small, but with some notable exceptions. Adjustments of more than 200 m over the Gauteng area are shown in Table 2, together with the corresponding unadjusted and adjusted heights, and the reference heights skipped in the valley-filling process.
Except in the case of C8038 (3) (see below), it can be seen that none of these larger differences were negative; that is, these were all adjustments upward. In all cases, instead of moving to the nearest reference surface, the terrain height was raised at least one level; that is, these were valley-filling cases. Note the differences in the entries for the two 48-km systems and the two 29-km systems. For instance, U4838L had four adjustments greater than 200 m, but U4838S none at all. Over the Gauteng area the maximum number of levels skipped was two, in contrast to four over a more mountainous western region (28°–36°S, 15°–27°E) also investigated, but not shown here. Similarly, over Gauteng the adjustments tended to be a smaller proportion of the original heights, compared to the western region where at several locations of lower lying terrain the adjustment was of the same order as the unadjusted height. The worst case seen was for U4838S, in the western region, where an unadjusted height of 491 m was raised through four levels to 1038 m, a change of 547 m.
c. Deficiencies in the unadjusted terrain
Four of the five large adjustments (shown in italics in Table 2) for the C8038 (3) experiment, in which the adjusted C8038 (1) terrain was manually modified, were in fact artificial modifications since the corresponding adjustments for C8038 (1) were small. This indicates that it was the preliminary unadjusted terrain that was considered deficient at these points. Due to the averaging process within a (large) grid block used to generate the unadjusted terrain from the given fine-mesh topography, actual, meteorologically significant mountains may well be smoothed out.
d. Effects of increased model resolution
The magnitudes of the large adjustments did not necessarily decrease with increased horizontal resolution, presumably because the more detailed the representation of the terrain, the greater the potential for deep valleys. It was noteworthy in the western area, though, that in a small area (32°–33°S, 19°–20°E) where U2938S, U4838L, and U4838S experienced the most severe adjustments of the order of 400–500 m, the U2938VS terrain only needed small adjustments with no valley filling. Thus a small shift of the grid through about 15 km made a crucial difference.
The increased vertical resolution for U8050 in the critical lower layers is evident in Fig. 3. The spacing improved from around 250 to 150 m or less at the elevations of the central plateau. Views of the layer distribution in terms of pressure rather than height (not shown) are similar to an NCEP 50-level configuration (Rogers et al. 1998, Fig. 2b). However, the availability of a greater number of more closely spaced reference surfaces, to which the unadjusted heights could potentially be adjusted, did not necessarily lead to smaller adjustments. Note, for example, in Table 2, the large adjustment of 267 m (from 1139 to 1402 m) at 29.2°S and 24.4°E. This is a valley-filling case where two nearer reference heights, 1117 and 1256 m, were bypassed. Something similar happened at the same point for C8038, but not for C8038 (2) or (3), suggesting that manual alteration as in experiments 2 or 3 might have been beneficial for U8050 as well to alleviate the effects of low horizontal resolution.
6. Case: Rainfall event at about 0000 UTC 31 August 1997
a. Synoptic conditions
The synoptic analysis for 0000 UTC on 30 and 31 August obtained from the SAWB Central Forecasting Office is shown in Fig. 7. A prominent frontal system, well represented in the upper air, approached the country from the west on 30 August with a shallow coastal low evident in the lower levels in the vicinity of Port Elizabeth (PE) on the south coast. On 31 August the surface low pressure system, associated with the cold front, passed south of the country, while the coastal low slipped eastward along the coast. A surface trough developed over the central parts of the subcontinent in conjunction with the frontal system and coastal low. This surface trough usually plays an important role in the development of rainfall over the interior of southern Africa, since it enhances the southward flow of moisture over the eastern parts of the country. This is clearly illustrated in the increase in relative humidity values at 850 hPa (boxes in Figs. 7b and 7e) from 30 to 31 August.
b. Rainfall event and actual terrain
Light convective rain fell on the night of 30 August 1997 over the eastern parts of the country (see Fig. 6). Most stations reported rainfall amounts below 5 mm, with falls of 5–10 mm recorded at some stations. Pretoria reported only 4 mm. Some falls greater than 20 mm did occur, isolated in two separate maxima. The first (northern maximum) was situated at 27°S and 25.5°E, where falls of up to 30 mm were observed. The second (southern maximum) was at 30°S and 27°E, with falls of approximately 22 mm.
The rainfall patterns correlated well with the higher topographical regions of the study area (Fig. 6). The rainfall band loosely followed the contours of the Drakensberg Mountains and the Witwatersrand Ridge (WR), while the 5-mm area closely followed the contours of these mountains. Furthermore, the southern maximum coincided with the area of elevated topography just southwest of the Lesotho Highlands, while the northern maximum fell within the topographical gap created by the Kuruman Heuwels and the Witwatersrand Ridge. Thus it seems likely that the topography played an important role in enhancing the rainfall mechanisms.
7. Low-resolution pre-upgrade results
It was a feature of a full analysis of this case (Bruyère et al. 2000) that all the model configurations of interest here handled the dynamical situation well and in a similar fashion (with the variation that might be expected in any ensemble of forecasts). Thus here only a broad outline is given together with views of the analyzed and predicted dynamical fields in Fig. 8. Total precipitation fields are shown in Fig. 9, but the convective precipitation fields were virtually identical (and similarly for Fig. 10). The large initials in Fig. 9 (and Fig. 10) indicate the locations of the topographical features shown in Fig. 6, namely, Kuruman Heuwels (KH), Bloemhof Dam (BD), Witwatersrand Ridge, Waterberg/Strydompoort (W/S), and Tugela Valley (TV).
The model of the day available to the forecasters was C8038. The 0-h analysis for 0000 UTC 30 August evaluated well and was able to capture the main features seen on the observational charts. The 24-h forecast drifted slightly and moved the coastal low slightly too quickly eastward to just south of Durban (DN). The surface trough over the interior, with the resultant increase in moisture to the east of it, was well captured. As expected, the small changes in the lower layers embodied in the topography experiments did not significantly affect the synoptic forecast ability of the model, and the C8038 (2) and (3) forecasts were very similar to those of C8038.
b. Prediction of the rainfall event and model topography
In experiment C8038 the rainband was placed too far west (Fig. 9a). C8038 predicted light rainfall, mostly below 5 mm as observed, but failed to predict the two separate areas of maximum rainfall. Instead it predicted a single maximum of above 20 mm between the observed maxima, with decreasing rainfall outward from this central point. The topographical contours were not followed. The C8038 topography (Fig. 4a) exhibited various deficiencies hypothesized to have contributed to the incorrect placement of the rainfall band. While C8038 correctly represented the Lesotho Highlands as well as the Drakensberg Mountains, the Witwatersrand Ridge, and the Waterberg/Strydpoort Mountains were totally absent. The Tugela valley was visible, but not clearly defined, while the Kuruman Heuwels and the valley to the east of it, within which lies the Bloemhof Dam (hereafter referred to as the Bloemhof valley), were incorrectly represented. The Bloemhof valley was represented as a valley within higher-lying ground, but in reality the Kuruman Heuwels is an outcrop within a lower-lying area.
2) C8038 (2)
In this experiment, the basic C8038 vertical structure was replaced with a configuration designed to increase the resolution around 850 hPa, roughly the height of the South African plateau. In Fig. 3 note the finer layers for C8038 (2) between levels 10 and 14, where the layer depths have been reduced from well over 200 m to about 150 m. From Fig. 4b it is clear that a better representation of the plateau area was achieved. There was significant improvement in the depiction of the Witwatersrand Ridge, Kuruman Heuwels, and, to a certain extent, the Tugela valley. Although placed too far south by about 1°, the Bloemhof valley was also better represented. The loss of some resolution in the lower-lying regions, to make up for the increased number of layers around 850 hPa, had a clear effect in the northwestern quadrant where the topography was now markedly lower. A small but noteworthy difference between the C8038 and C8038 (2) predicted precipitation fields (Fig. 9b) was the southward displacement of the maximum, in concert with the southward topographical shift of the Bloemhof valley for C8038 (2), both through approximately 1°. A second difference was the way in which the eastern edge of the rainfall band now followed the outline of the Witwatersrand Ridge more closely.
3) C8038 (3)
The manual adjustment here of the model terrain (Fig. 4c) attempted to rectify certain specific problems regarded as meteorologically important. These shortcomings were solved to the extent allowed by the coarse 80-km grid resolution and the restriction of no totally enclosed valleys. The Witwatersrand and Waterberg/Strydpoort Mountains were added, taking into consideration that there is naturally a valley between the Witwatersrand Ridge and the Waterberg/Strydpoort Mountains. An outcrop was added in the northernmost part of the Drakensberg, and the definition of the Tugela valley was enhanced. Last, the Kuruman Heuwels/Bloemhof valley area was corrected, taking into account the position of the Bloemhof valley, as well as the gap between the Kuruman Heuwels and the Witwatersrand Ridge. As for the preceding experiment, there was a small improvement in the realism of the predicted precipitation (Fig. 9c). The area of maximum predicted precipitation broadened, increasing the amounts at the locations of both the southern and northern observed maxima. The presence in this experiment of the Waterberg/Strydpoort Mountains had a beneficial effect on the rainfall field. This now bulged eastward into the valley created between the Witwatersrand and the Waterberg/Strydpoort Mountains, while just farther north it was prevented from extending eastward by the barrier created by these mountains.
8. Major effects among the upgrade runs
A summary of the impact tests as originally designed for the stepwise upgrade process is given in Table 3. Immediately obvious consequences of the major factors here for the terrain and precipitation fields are highlighted below. The effects of topography changes were not considered in the design of these experiments. However, valid information might be available through comparisons of like (or nearly like) configurations and each test is also examined for its suitability to identify terrain effects.
a. Comparison 1
The NCEP enhancements were expected to have a large effect, and indeed did so on the case day as can be seen at a glance in comparing the precipitation forecasts in Figs. 9a and 10a. Positive aspects here were the eastward shifting of the rainfall band and the development of two distinct maxima, although the precipitation amounts were much too small. The NCEP package also contained enhancements in the creation of the terrain, but their effects cannot be isolated from those of the total package. In any case, the new code did not greatly change the terrain (see Figs. 4a and 5a), although there was an improved depiction of the Kuruman Heuwels/Bloemhof valley area.
b. Comparisons 3 and 5
As expected, the increase in horizontal resolution also had large effects. At 48 km the increase in detail in the terrain was immediately apparent (see Figs. 5a, 5c, and 5d). The Waterberg/Strydpoort Mountains were now defined, although not yet separated from the Drakensberg Mountains. The Witwatersrand Ridge, Kuruman Heuwels, and Tugela valley were all better depicted. At 29 km the detail improved dramatically and most of the important features were now present (Figs. 5e and 5f). The Waterberg/Strydpoort Mountains were clearly defined and there was a major improvement in the definition of the Tugela valley. Similarly, the detail on the precipitation fields was markedly more realistic for the 48-km configurations (Figs. 10a, 10c, and 10d), and the 29-km predicted precipitation fields (Figs. 10e and 10f) looked very similar to the recorded precipitation field (Fig. 6). With the increase in resolution from 80, to 48, to 29 km the maximum rainfall amounts began to approach the observed, increasing from 10, to 12, to 20 mm. Although the precipitation fields followed the topography well, the increased terrain resolution is so interwoven with the general resolution increase that terrain effects cannot clearly be inferred here.
c. Comparison 2
The increase in vertical resolution might have been expected to be the dominant effect here and to result in improvement in the model performance. On the day of the study this turned out to be the case (Bruyère et al. 2000) in the upper air, but the lower-level synoptics for U8050 were slightly degraded, and the precipitation was poor (see Figs. 10a and 10b, and section 9). It seems likely that this degradation was related to the quality of the U8050 topography, which was also disappointing (see section 5d, Figs. 5a and 5b, and section 9). It thus seemed reasonable to examine the terrain–precipitation connection here, especially as these two factors were interlinked in the weather situation studied.
d. Comparisons 4 and 6
During the stepwise upgrade it came to light, chiefly through use of the graphical tools developed for the low-resolution topography experiments, that tests for the effect of domain size had been contaminated through unexpected differences in the topographies for the pairs of models involved. As explained in section 5a above, seemingly minor choices in configuring the horizontal grid led to quite considerable differences in the terrain representations for the 48-km and 29-km pairs of runs (Figs. 5c and 5d, 5e and 5f). There were several reasons for suspecting that the terrain effect was the more important here. It was observed that the only changes in the precipitation fields (Figs. 10c and 10d, 10e and 10f) occurred at locations where the terrain changed (see section 9). The smaller-domain configurations would not in general have been expected to produce improvement but, in both sets of trials here, it turned out that the smaller-domain runs gave more realistic precipitation forecasts, corresponding to more realistic terrain representations also. Synoptically, there was not a clear domain size signal: both good and poor aspects in the model fields were associated with both the larger and the smaller domains. The very small domain of U2938VS played a role only in systems very close to the domain boundary outside of the Gauteng area (Bruyère et al. 2000). Thus, the area of interest was deemed far enough away from the lateral boundaries (see Fig. 2), and the forecast period short enough, for there not to be significant effects on the particular rainfall event considered here, where there was a clear orographic signal. Such aspects are considered by Warner et al. (1997) to be less sensitive to lateral boundary conditions. The special measures taken in the Eta Model near the lateral boundaries (see Black 1994) designed to prevent proliferation of boundary errors are also believed to be very effective (Staniforth 1997; F. Mesinger, 1997, personal communication).
9. Terrain effects in the upgrade runs
Comparisons 1, 3, and 5 from Table 3 are disqualified in the search for terrain effects. The remaining comparisons can be recategorized as in Table 4 from the point of view of studying the impact of terrain changes on model precipitation.
a. Terrain-effect comparison 1
Relative to U8038, the U8050 representation of the topography around the height of the plateau improved, and the Drakensberg Mountains, the Witwatersrand Ridge, and the Tugela valley were now more clearly defined (Figs. 5a and 5b). As noted in section 5d, however, it happened that for U8050 higher vertical resolution and the availability of more potential terrain elevations did not lead to smaller adjustments or a reduction in valley-filling problems. Thus the topography had several significant defects, notably a loss of definition at the lowest layers. For U8050 no attempt was made to correct these defects by improving the eta-layer distribution as for C8038 (2), or by modifying the 80-km terrain manually as for C8038 (3). As for C8038, the Kuruman Heuwels/Bloemhof valley area was again represented as a valley within higher-lying ground, where in reality the Kuruman Heuwels is an outcrop within a valley. A pronounced difference could be seen between the predicted precipitation fields of U8038 and U8050 (Figs. 10a and 10b). U8050 had steeper topography than U8038 in the Witwatersrand Ridge area that seemingly created a barrier effect, preventing the moisture from moving farther south and resulting in the northern rainfall maximum being placed farther north than in U8038. Thus here degradation in the terrain seems clearly linked to degradation in the precipitation forecast.
b. Terrain-effect comparison 2
The most noticeable difference between the precipitation fields of the two 48-km runs was in the placement of the northern maximum (Figs. 10c and 10d). This maximum was predicted north of the Witwatersrand Ridge by U4838L, but south of the ridge by U4838S, which was closer to the observed. The major differences between the topographies of the two runs were the width of the Witwatersrand Ridge and the gap between this ridge and the Kuruman Heuwels, with U4838S again the more realistic. Thus both the topography and the predicted precipitation in the vicinity of the gap were closer to reality for U4838S than for U4838L, suggesting a direct relationship between the two factors.
c. Terrain-effect comparison 3
The topography was very well represented in the two 29-km runs (Figs. 5e and 5f) but there were two main differences. First, the gap between the Kuruman Heuwels and the Witwatersrand Ridge, while close to reality for both configurations, was larger for U2938S than for U2938VS. Second, while the Waterberg/Strydpoort Mountains were evident in both cases, for U2938VS they were separated from the Drakensberg Mountains for the first time, apart from C8038 (3). U2938VS also saw the outcrop in the northernmost parts of the Drakensberg Mountains. Differences seen in the precipitation fields (Figs. 10e and 10f) can be attributed to the two main differences in the topographies. The wider gap for U2938S allowed more flow to the south and therefore more rain was forecast south of the Witwatersrand Ridge than for U2938VS. For U2938VS the majority of the rainfall was to the north of the Witwatersrand Ridge extending to the east. U2938VS and C8038 (3) were the only two models to extend the spread of rain to the east. This seems directly attributable to the separation by these two models of the Waterberg/Strydpoort Mountains from the Witwatersrand Ridge.
In general, the representation of the terrain in the configurations of the Eta Model studied was found to be highly dependent on small shifts in the placement of the horizontal grid points, and also on changes in the vertical distribution of eta-coordinate surfaces (and thus the pool of discrete heights available to define the terrain). At low horizontal resolution, an increase in vertical resolution might not necessarily result in a better representation of the topography, due possibly to “valley filling” problems, or lack of care in selecting the vertical grid. Although a great improvement may be achieved through refinement of the horizontal grid, large deviations in the model terrain (related to the step-mountain and valley-filling formulations) may remain, with possible adverse effects in both the handling of surface observations and the verification of model values at affected stations.
In the low-resolution topography experiments with the March 1996 version of the model, C8038 (2) changed the model topography indirectly through a rearrangement of the vertical structure of the model, and in C8038 (3) the model topography was altered to better represent meteorologically significant features. In the weather situation studied a strong link existed between the actual topography and the observed precipitation. In these experiments a similarly clear link was established between the changes in the model terrain and changes in predicted precipitation. It was notable that differences in predicted precipitation occurred only in areas where the topography was altered. Although the differences between the precipitation forecasts were often small, improvements were seen for both C8038 (2) and (3), where the precipitation started to respond to the creation of the Witwatersrand Ridge by following the contours created by this barrier. Another quite considerable improvement was seen in C8038 (3), where the valley created between the Witwatersrand Ridge and the Waterberg/Strydpoort Mountains resulted in the eastward bulging out of the rainfall field in that area. Thus, even with a resolution as coarse as 80 km, improvements may be possible if efforts are made to ensure that the topography is represented correctly in the model.
Major improvements in the ability to place the precipitation correctly were achieved only once the wide-ranging NCEP model enhancements contained in the August 1998 upgrade were introduced (U8038). (However, at the same time, the maximum rainfall amounts for U8038 were surprisingly degraded.) The terrain change from C8038 to U8038 was an integral part of the upgrade, and was relatively small. The increases in horizontal resolution to 48 and 29 km led to an improvement in the accuracy of the model terrain, and the realism of forecasts of precipitation (both position and amount). Although it was not possible to separate the effect of refinement of the topography from the general refinement, it was noticeable that as the resolution improved and more of the smaller topographical detail could be resolved, the predicted precipitation patterns followed the topography more and more closely, approaching the observed. At yet higher resolutions, such as the 22-, 11-, and 5.5-km models described by Gollvik (1999), a stage might be reached where refinement of the orography is no longer accompanied by better forecasts of precipitation, but for the 48- and 29-km Eta Model configurations described here the refinement seemed definitely beneficial.
In contrast to the low-resolution pre-upgrade experiments that were designed with the aim of improving the model topography, three of the upgrade configurations could be regarded as inadvertent experiments on the effect of terrain changes. As for C8038 (2), the vertical structure of the model was rearranged for U8050, also around the elevation of the central plateau. However, in the case of U8050, no special care was taken in choosing the distribution of the eta layers, and no attempt was made to alleviate any shortcomings of the 80-km resolution terrain as was done for C8038 (3). These circumstances, together with possible problems with valley filling and large adjustments in the process of creating the step-mountain terrain, probably led to the disappointing model terrain. Particular aspects of the degradation in the U8050 terrain could be linked directly to corresponding degradation in the precipitation prediction. The general advantage expected from the increase in vertical resolution could thus be overshadowed by the effects of poor choices in the vertical discretization, or incorrectly placed topographical features. This configuration could therefore no doubt have benefited from measures such as taken for C8038 (2) and (3).
The two 48-km configurations, and the two 29-km configurations, were involved in the other two inadvertent experiments. Here the differences in the model topographies came about through the different choice of the domain center (48 km) and odd–even choices in the number of zonal grid points (29 km). These seemingly insignificant elements of setting up a model configuration had quite a significant impact on the model topographies, once again through the adjustment to eta surfaces and the valley-filling process in creating the step mountains. The effect of the terrain differences was considered far more important here than domain-size effects, as was borne out in the close correspondence between changes in the terrain and changes in the precipitation. It turned out that in each case it was the small-domain topography and the small-domain precipitation prediction that gave the best results. In the operational environment, at higher horizontal resolutions, it might be worthwhile at implementation time to carry out a few experiments with small grid shifts to check that important features of the topography are handled as well as possible.
Here we demonstrated the effect of changes in the Eta Model topography (effected through direct or indirect means) only on predicted precipitation, through a case where terrain and precipitation were linked. Although the effects were generally small, improvements in the terrain representation were accompanied by improvements in the realism of forecasts of precipitation, and vice versa. However, the model as a whole would surely benefit from having the best possible representation of the terrain. Often the generation of the model topography is regarded as a black box within an operational modeling system package. In light of this study, however, it would seem to be instructive and beneficial to look beyond the black-box solution, possibly developed for another region and other conditions, and seek the locally optimal terrain representation. Even if modifications are not attempted, for proper interpretation of forecasts from the Eta Model in particular, it would seem extremely important to have a knowledge not only of the actual topography of the forecast area, but also of the limitations of the model topography.
To sum up, as could be expected, model enhancements provided the greatest impact on the precipitation predictions, and higher horizontal resolutions generally led to better representations of the terrain and more realistic forecasts. The possibility exists here of improving the model topography yet further through small modifications to the grid structure at the time of setting up the model. However, it also seems even from this limited study that if higher horizontal resolutions are not affordable for some reason, slight grid shifts or manual modifications that improve the representation of the model terrain could be useful and viable methods of improving model products. There is some potential in the judicious choice of the vertical coordinate structure, although what is good for one region might prove bad for another. There seems to be more potential in the manual correction of meteorologically significant deficiencies in the model terrain. In particular, possibly excessive smoothing of meteorologically important mountains through averaging over large grid blocks could be counteracted. These approaches are likely to be beneficial also in the presence of higher vertical resolution, which in itself may not be sufficient to improve the topography. As mentioned above, altering the terrain by hand, even at low resolutions, is a laborious task. Much juggling is needed to avoid the introduction of windless valleys, an iterative task more suited for computers. At higher resolutions this process would be all but impossible by hand, but then the terrain would naturally be better represented in any case. At lower resolutions, such as the 80 km used here or even coarser, the alteration of a few terrain blocks in strategic places might be feasible and make all the difference. With the Eta Model being run on smaller machines, in centers with limited resources, this option seems to deserve some consideration. Thus, the manual alteration only becomes feasible at low resolutions, and it is only at low resolutions that one is motivated to attempt it.
The generosity of NCEP in supplying the Eta Data Assimilation System code is gratefully acknowledged. Cray Research are thanked for their kindness in allowing us to run the U2938S model on their computers. The stimulating suggestions of the anonymous reviewers have greatly enhanced the readability of the article.
Corresponding author address: H. A. Riphagen, South African Weather Service, Room 5089, Forum Bldg., 159 Struben St., Pretoria 0002, South Africa. Email: firstname.lastname@example.org