Abstract

A number of numerical experiments with a high-resolution mesoscale model were conducted to study the convective rainfall event that caused the 1996 Buffalo Creek, Colorado, flash flood. Different surface conditions and treatments of land surface physics were utilized to assess the sensitivity of this orographic moist convection to local and regional landscape forcing.

Given accurate large-scale synoptic conditions at the lateral boundaries, the mesoscale model with a convection-resolving grid shows reasonably good skill in simulating this convective event with a lead time of up to 12 h. Sensitivity experiments show that a primary reason for this success is the use of an advanced land surface model that provides time-varying soil-moisture fields. This land surface model plays an important role in capturing the complex interactions among the land surface, the PBL, cloud-modulated radiation, and precipitation. For the case simulated, such interactions contribute to the temporal and spatial distribution of surface heating at small scales, and the convective triggering and development.

Tests show that the landscape variability at small and large scales significantly affects the location and intensity of the moist convection. For example, on timescales of 6 to 12 h, differences in initial soil moisture associated with irrigation in the plains affect the evolution of the convection near the Continental Divide. Also, the surface modification by a wildfire burn influences the path of the major convective event that caused the flash flood.

A watershed-based quantitative-precipitation-forecast skill score is proposed and employed. The relative success with which this severe thunderstorm is simulated over complex terrain provides some hope that the careful treatment of land surface physics in convection-resolving models can perhaps provide some useful level of predictability.

1. Introduction

On 12 July 1996, several thunderstorms moved across the Buffalo and Spring Creek watersheds in the Front Range of the Rocky Mountains in Colorado, approximately 45 km to the southwest of Denver. During about a 1-h period, precipitation accumulations exceeded 80 mm over both watersheds, causing a flash flood that resulted in the loss of life and property. Numerous facztors undoubtedly contributed to the mesoscale modulation of the large-scale, unstable, upslope flow, such that the specific location and intensity of this event were determined. The purposes of this study are to 1) define the sensitivity of the intensity and position of the convection in this large-scale environment to local and regional landscape variability, and 2) assess the skill with which this convective event in mountainous terrain can be simulated using a convection-resolving model that employs high-resolution terrain and land surface properties.

The objectives are both practical as well as scientific. As operational weather prediction activities eventually employ convection-resolving storm-scale models, it will become necessary for us to have a better understanding of the sensitivity of the forecast convection to the accuracy of the surface-condition specification and the accuracy with which we model the response of the atmosphere to the surface variability. This study will address this question through the use of a convection-resolving model in sensitivity tests that define the degree to which different specifications of the surface landscape modulate the convection differently. Eventual operational storm-scale forecasts are most likely to be performed for only the most threatening convective events, at least initially when only modest-sized computational windows can be afforded. Thus, this study will employ a case in which intense thunderstorm precipitation caused a flash flood.

There is plentiful evidence that regional landscape variability and the adequacy with which its effects are represented in atmospheric models have a potentially strong impact on both local and distant simulated mesoscale weather and climate. For example, Segal et al. (1989) demonstrate that the irrigation in eastern Colorado produces local thermally forced circulations. Chang and Wetzel (1991) employ a mesoscale model to show that spatial variations in vegetation and soil moisture affect the evolution of the prestorm convective environment in the eastern High Plains. Chase et al. (1999) demonstrate in a modeling study that conversion of grasslands to dry land and irrigated farm land in northeastern Colorado affects the local atmospheric conditions as well as the rainfall in the mountains to the west. Numerous additional modeling and empirical studies document mesoscale impacts of surface variability associated with coastlines (Dalu and Pielke 1989), snowfield boundaries (Segal et al. 1991), surface moisture (Yan and Anthes 1988), and salt flats (Physick and Tapper 1990).

Other studies have evaluated the relationship between the scale of the landscape variability and the intensity of the atmospheric convective response. For example, in the modeling study of Yan and Anthes (1988), it was found that bands of moist land need to be 50–100 km or greater in width in order for rainstorms to be produced by the convergence of the “inland sea breeze” circulation. This supports the hypothesis of Anthes (1984) that planting bands of vegetation with widths of 50–100 km in semiarid areas could result in increases in convective rainfall. Avissar and Schmidt (1998), using a large-eddy simulation model, employ surface sensible heat flux waves of different wavelength, and look at the effects of the surface heterogeneity on the resulting thermal circulations and atmospheric heat fluxes in the convective boundary layer. The strength of the thermals increased with increasing wavelength of the surface variability up to 20 km. However, a further increase in the wavelength to 40 km resulted in a decrease in the strength of the thermal. In contrast to the previous studies, these were nonprecipitating thermals. In the study described here, the impact on the convective rainfall of landscape variability over a range of scales is evaluated.

These results also have implications regarding the predictability of convective rainfall over complex terrain. Numerical weather prediction models have been employed for four decades to produce research simulations and operational forecasts of precipitation. Even though the skill of the operational model predictions of large-scale circulations has improved markedly during the period, progress has been especially slow in improving summer-season quantitative precipitation forecasts (Fritsch et al. 1998; Georgakakos and Hudlow 1984). One of the impediments has been the lack of conventional in situ data on the thunderstorm scale. Thus, the development of techniques for the variational initialization of mesoscale models with WSR-88D radar reflectivity and Doppler winds holds promise of contributing significantly toward progress in this area (Droegemeier 1997; Xu 1996; Sun and Crook 1994, 1997). Unfortunately, for a variety of reasons, these radar data are often unreliable or nonexistent in areas of complex terrain. And, even when these radar data are available, they are most useful for the initialization of convective events that are in progress. Thus, the forecast lead time is marginal and perhaps insufficient for avoiding loss of life by flooding in rural communities without a good alert system. Therefore, a useful degree of predictability, if that is attainable, of heavy rainfall in mountainous terrain may have to rely on better representing local orographic and other landscape forcing in convection-resolving models. There have been some notably successful simulations that give us hope for possible eventual operational predictability in such situations. For example, Nair et al. (1997) describe a successful simulation of the flash flood producing the severe 1972 Black Hills convective storm in South Dakota, where no special mesoscale data were employed. This paper will provide some further encouragement that correctly defining the large-scale environment and local forcing can (but perhaps not routinely) lead to significant levels of predictability of orographic convective storms.

Section 2 summarizes the meteorological conditions during the simulation period, and a description of the modeling system is provided in section 3. The experimental design is described in section 4, the results of the control and various sensitivity experiments are described in section 5, and a summary and discussion are provided in section 6.

2. Meteorological conditions associated with the convection

During the day of 12 July 1996, before the main Buffalo Creek convective event triggered in the evening, a weak front moved from the northeast into Colorado, turning the winds to upslope easterlies in the eastern Plains as it advanced (Fig. 1). Surface dewpoint temperatures in advance of the front and upstream of the mountains were between 55° and 65°F. Upper-air flow was from the northwest, with a jet maximum approaching the Rockies from Canada. Ageostrophic circulations in the exit region of the jet could have been responsible for large-scale vertical motions over Colorado. A 500-hPa cold anomaly over the northern Rockies at 1200 UTC 12 July and 0000 UTC 13 July 1996 (not shown) might be related to upward vertical motion in this area. Figure 2 shows the Denver sounding for 0000 UTC 13 July. It is relatively unstable, and the boundary layer moisture has increased and deepened since the previous sounding. The 10–15 m s−1 northerly boundary-layer winds likely result from an outflow boundary. Thus, the combination of high surface dewpoints, weak frontal forcing, low-level upslope flow, and the possible upper-level support from the jet stream produced an environment that was conducive to convection over the Front Range of the Colorado Rockies.

Fig. 1.

Surface analysis for (a) 1800 UTC (1100 MST) 12 Jul and (b) 0000 UTC (1700 MST) 13 Jul 1996. The outline of the Buffalo Creek watershed (BCWS) is shown to the southwest of Denver

Fig. 1.

Surface analysis for (a) 1800 UTC (1100 MST) 12 Jul and (b) 0000 UTC (1700 MST) 13 Jul 1996. The outline of the Buffalo Creek watershed (BCWS) is shown to the southwest of Denver

Fig. 2.

The Denver radiosonde sounding for 0000 UTC 13 July 1996

Fig. 2.

The Denver radiosonde sounding for 0000 UTC 13 July 1996

The Buffalo and Spring Creek watersheds were especially vulnerable to flooding because, approximately 2 months earlier, a wildfire had burned 50 km2 of the watersheds, which drastically increased the ratio of rainfall runoff to infiltration. During the morning of 12 July 1996, the nearest convection to Colorado was associated with a dissipating line moving southward from eastern Oklahoma into Texas. However, by noon, isolated convection began to develop along a north–south line over the higher elevations of Colorado. As the afternoon progressed, the convection became more organized as it moved over the plains to the east. The last convective development over the mountains during the day initiated near 1800 mountain standard time (MST) (0100 UTC) in north central Colorado, and multiple convective storms associated with this outbreak caused the flash flood by producing rainfall in excess of 80 mm in 1 h directly over the burn area. There were at least five other convective rainfall events that occurred in the vicinity of the Buffalo Creek watershed during this outbreak.

Figure 3 shows the total rainfall in the vicinity of the watershed for the period 0000–0348 UTC 13 July (1700–2048 MST 12 July) 1996 based on measurements by the National Center for Atmospheric Research's (NCAR) S-Pol dual-polarization radar located to the east of Denver. This total spans the entire period of the convection that contributed to the flood, and includes the effects of multiple cells. The area of this figure corresponds to that of the highest resolution computational grid of the quadruply nested modeling system used here. See Warner et al. (2000) for a more complete discussion of the rainfall estimates. During this period, the discharge in Buffalo Creek increased from its typical July base flow of a few cubic meters per second to over 450 m3 s−1 in less than 30 min (R. Jarrett 2000, personal communication). About one-quarter of the northeast part of the watershed was burned by the wildfire (Fig. 6b). By midnight (MST), all significant convective activity had ceased for the day. Figure 4 displays the observed precipitation totals for each of the three coarser nested grids used in the simulation, for the 24-h period of the simulations from 1200 UTC (0500 MST) 12 July to 1200 UTC 13 July 1996. These analyses are based on the National Centers for Environmental Prediction (NCEP) 4-km multisensor precipitation analysis, that is produced using WSR-88D estimates, corrected by available gauge measurements. There are 2 h for which data are missing in the NCEP dataset. Also, there were blackout areas where no gauge and radar observations were available. These blackout areas sometimes account for as much as one-quarter of the area of grid 1. Thus, the 24-h amounts are somewhat underestimated.

Fig. 3.

The precipitation in the vicinity of the Buffalo Creek watershed for the period 0000–0348 UTC 13 Jul 1996, based on differential propagation phase measurements by the NCAR S-Pol dual-polarization radar located to the east of Denver. The Buffalo Creek watershed is outlined. This area corresponds to computational grid 4, the highest resolution grid used in the quadruply nested system. The grayscale legend indicates the rainfall amounts (mm), with the zero isohyett included

Fig. 3.

The precipitation in the vicinity of the Buffalo Creek watershed for the period 0000–0348 UTC 13 Jul 1996, based on differential propagation phase measurements by the NCAR S-Pol dual-polarization radar located to the east of Denver. The Buffalo Creek watershed is outlined. This area corresponds to computational grid 4, the highest resolution grid used in the quadruply nested system. The grayscale legend indicates the rainfall amounts (mm), with the zero isohyett included

Fig. 6.

Topography for grids (a) 3 and (b) 4. The isopleth interval is 250 m, with the 2000 and 3000 m contours drawn with heavier lines. Elevations of greater than 3000 m are stippled. The watershed boundary is outlined on grids 3 and 4. The area where the surface conditions were modified by the wildfire is shaded, as shown on grid 4

Fig. 6.

Topography for grids (a) 3 and (b) 4. The isopleth interval is 250 m, with the 2000 and 3000 m contours drawn with heavier lines. Elevations of greater than 3000 m are stippled. The watershed boundary is outlined on grids 3 and 4. The area where the surface conditions were modified by the wildfire is shaded, as shown on grid 4

Fig. 4.

Observed precipitation totals for the three coarsest-resolution model grids for the 24-h period of the simulations, from 1200 UTC (0500 MST) 12 Jul to 1200 UTC 13 Jul 1996. The analyses are based on the NCEP 4-km multisensor precipitation dataset, and are shown for (a) grid 1, (b) grid 2, and (c) grid 3. The grayscale legend indicates the rainfall amounts (mm)

Fig. 4.

Observed precipitation totals for the three coarsest-resolution model grids for the 24-h period of the simulations, from 1200 UTC (0500 MST) 12 Jul to 1200 UTC 13 Jul 1996. The analyses are based on the NCEP 4-km multisensor precipitation dataset, and are shown for (a) grid 1, (b) grid 2, and (c) grid 3. The grayscale legend indicates the rainfall amounts (mm)

3. Description of the model

The model used in this study is the Pennsylvania State University–NCAR nonhydrostatic mesoscale model, version five (MM5). For details about this modeling system, refer to Dudhia (1989, 1993), Grell et al. (1994), and Warner et al. (1992). The quadruply nested computational grids are depicted in Fig. 5. Grid 4 (the inner grid), grid 3, grid 2, and grid 1 (the outer grid) have mesh sizes of 52 × 58, 97 × 52, 76 × 76, and 73 × 84 points and grid increments of 1, 3, 9, and 27 km, respectively. The nested grids, each with 43 computational layers, interact during the simulation (i.e., two-way nesting). The highest vertical resolution is near the ground, and the lowest computation layer is approximately 15 m above the surface. The model top is located at 50 hPa.

Fig. 5.

The quadruply nested computational grids. Grid 4 (the inner grid), grid 3, grid 2, and grid 1 (the outer grid) have mesh sizes of 52 × 58, 97 × 52, 76 × 76, and 73 × 84 points and grid increments of 1, 3, 9, and 27 km, respectively

Fig. 5.

The quadruply nested computational grids. Grid 4 (the inner grid), grid 3, grid 2, and grid 1 (the outer grid) have mesh sizes of 52 × 58, 97 × 52, 76 × 76, and 73 × 84 points and grid increments of 1, 3, 9, and 27 km, respectively

The planetary boundary-layer parameterization, based on Hong and Pan (1996), is the “nonlocal” technique that is employed in the Medium-Range Forecast (MRF) model of NCEP. For the precipitation parameterization, the Grell technique (Grell 1993) is used on the outermost two grids, with no parameterization employed on the inner two grids. A simple explicit treatment of cloud microphysics is employed and is based on Dudhia (1989). Both ice and liquid phases are permitted for cloud and precipitation, but mixed phases are not permitted. The model uses a radiation scheme in which longwave and shortwave radiation interact with the clear atmosphere, cloud, precipitation, and the ground (Dudhia 1989).

The land surface model (LSM) used in the MM5 system is based on the diurnally dependent Penman potential evaporation approach of Mahrt and Ek (1984), the multilayer soil model of Mahrt and Pan (1984), and the primitive canopy model of Pan and Mahrt (1987). It has been extended by Chen et al. (1996) to include the canopy-resistance approach of Jacquemin and Noilhan (1990) and the surface-runoff scheme of Schaake et al. (1996). This LSM has been recently coupled to the MM5 modeling system (Chen and Dudhia 2001). It has one canopy layer, and the following prognostic variables: volumetric soil moisture and temperature in four soil layers, water stored on the canopy, and snow stored on the ground. The depths of the soil layers are 0.1, 0.3, 0.6, and 1.0 m, from the top layer to the bottom layer, respectively. The root zone is in the upper 1 m of soil, and the lower 1 m of soil acts like a reservoir with gravity drainage at the bottom.

An alternative that is simpler than the LSM for representing the surface energy and moisture budgets, and their response to landscape variability, is also tested. This simpler approach employs a five layer version (Dudhia 1996) of the “slab model” developed by Blackadar (1976, 1979) and tested further by Zhang and Anthes (1982). The bottom level extends to 32 cm below the ground surface, and there is no explicit representation of vegetation effects. The soil moisture remains constant during a simulation and is defined in terms of a moisture availability.

Surface landscape characteristics are defined here based on a number of alternative datasets. There are two datasets employed for defining basic landscape variability, with one having high resolution and the other low. For high-resolution conditions, the U.S. Geological Survey (USGS) Earth Resources Observing System (EROS) 1-km dataset (Loveland et al. 1995) is used to define the vegetation type and the State Soil Geographic (STATSGO) 1-km database is used for soil type (Miller and White 1998). The EROS dataset needed significant correction based on field reconnaissance. When these high-resolution surface data were used, conditions on coarser resolution grids in the nest were defined based on the vegetation and soil type that represented the largest area of the grid cell. The low-resolution dataset for vegetation and soil texture is based on 1° datasets that are currently used in the NCEP Eta Model.

The soil moisture first guess is based on an NCEP–NCAR 2.5° resolution reanalysis (Kalnay et al. 1996). This is then modified based on the NCEP 4-km analysis of precipitation for the 1-week prior to the model simulation. The effect of antecedent precipitation decreases as a function of time, where the dry down rate is dependent on the water-retention capacity of the particular soil types defined by the STATSGO database. Utilizing the 4-km NCEP rainfall analysis to modify the smooth soil-moisture fields obtained from the NCEP–NCAR reanalysis not only mitigates the biases in the model precipitation used in the reanalysis system, but also preserves surface heterogeneity at small scales. Surface conditions in all experiments but two are altered to account for changes in the area burned by the wildfire. Also, one experiment includes the soil-moisture effects of center-pivot irrigation in the plains of northeastern Colorado. Further details on the specification of the landscape characteristics can be found in section 4 that discusses the experimental design.

The model initial conditions are defined by analyzing radiosonde and surface data to the model grids using a successive-correction objective-analysis procedure (Benjamin and Seaman 1985), where the first guess field is the NCEP 40-km analysis from the Eta Data Assimilation System (EDAS). For all but one of the experiments, lateral-boundary conditions for the outer grid, grid 1, are defined using linear temporal interpolation between 6-hourly NCEP EDAS analyses. Because the primary purpose of the modeling system used in this study is to help identify sensitivities of convection to surface conditions rather than to produce a forecast, data assimilation by Newtonian relaxation (Stauffer and Seaman 1990; Stauffer et al. 1991) is employed on grid 1, where the model solution is “nudged” toward analyses of upper-air and surface observations, available at intervals of 12 and 3 h, respectively. The simulations are permitted to evolve freely on grids 2, 3, and 4, but are constrained by the large-scale information passing from grid 1 to grid 2 and by the land surface conditions. In the experiment whose purpose is to demonstrate the operational predictability of this convective event, lateral boundary conditions applicable after the initial time are based on NCEP Eta Model forecasts. Also, naturally, no data assimilation is performed on grid 1. Figure 6 shows the topography for grids 3 and 4. The duration of the simulations is 24 h, spanning the period from 1200 UTC (0500 MST) 12 July to 1200 UTC 13 July 1996.

It should be noted that Warner and Hsu (2000) demonstrate a significant sensitivity of the resolved convection on the inner grids of a nested system to the specific convective parameterization used on the coarser outer grids. Such sensitivity has also been observed with this case in terms of modest shifts in the locations of the precipitation maxima. However, because all experiments described here employ the same convective parameterization on the outer two grids, the differences among experiments will only result from the documented differences in the land surface conditions or land surface physics.

4. Experimental design

A primary goal of this study is to define the sensitivity of the rainfall location, timing, and accumulation to the surface landscape forcing, for this convective storm. To accomplish this, a number of sensitivity tests were performed, wherein the surface conditions and the land surface physics were varied. An additional experiment helped define the predictability of this convective event in an operational setting. Table 1 summarizes the experiments. The experimental conditions that were employed in each of these sensitivity tests are described below.

Table 1.

Summary of the experiments performed

Summary of the experiments performed
Summary of the experiments performed

a. The control experiment (experiment CON)

In this experiment, the best estimates of the surface landscape properties were used. For vegetation and soil type, the USGS EROS and STATSGO datasets, respectively, were employed. Figure 7 shows the high-resolution land surface vegetation conditions for grid 3. Coniferous forest is the dominant vegetation in the mountains, while grass and shrubs are dominant on the Plains. In the area of the wildfire, the normal vegetation type is coniferous forest.

Fig. 7.

High-resolution land surface vegetation conditions for grid 3. In the area of the wildfire, the normal vegetation type is coniferous forest. The outlined box is the grid 4. The category numbers are defined as follows: 2) broadleaf-deciduous trees, 3) broadleaf and needleleaf trees (mixed forest), 4) needleleaf-evergreen trees, 6) broadleaf trees with ground cover (savanna), 7) ground cover only (perennial), 8) broadleaf shrubs with perennial ground cover, 9) broadleaf shrubs with bare soil, 10) dwarf trees and shrubs with ground cover (tundra), 11) bare soil, and 12) irrigated cropland

Fig. 7.

High-resolution land surface vegetation conditions for grid 3. In the area of the wildfire, the normal vegetation type is coniferous forest. The outlined box is the grid 4. The category numbers are defined as follows: 2) broadleaf-deciduous trees, 3) broadleaf and needleleaf trees (mixed forest), 4) needleleaf-evergreen trees, 6) broadleaf trees with ground cover (savanna), 7) ground cover only (perennial), 8) broadleaf shrubs with perennial ground cover, 9) broadleaf shrubs with bare soil, 10) dwarf trees and shrubs with ground cover (tundra), 11) bare soil, and 12) irrigated cropland

The physical effects of the wildfire on the surface properties were accounted for in this experiment. Relative to the normal physical properties 1) the albedo was reduced from 15% to 5% (assuming that previous rainfall had washed away the ash, and what remained was charred grass and wood), 2) the soil moisture content in all layers was reduced from 0.3 to the wilting point, and 3) the roughness length was reduced from 50 to 30 cm. The postburn value of soil moisture is arguable, however the surface remains quite hydrophobic after a burn, and thus a very high percentage of rainfall runs off. Clearly there is little to no transpiration.

Outside of the burn area, the soil moisture was defined based on the NCEP–NCAR reanalysis as modified by antecedent precipitation. Figure 8 depicts the initial root-zone soil-moisture analysis for grid 3, and shows some detail in the soil moisture variability that results from the NCEP 4-km precipitation analysis. The effects of Plains irrigation, and irrigation in other areas, is not accounted for in this control experiment because of uncertainty in the specific irrigation protocols used. Rather, the surface moisture is defined using the normal procedure, as though the area is nonirrigated agriculture. Lateral boundary conditions for grid 1 are based on analyses of data, and data assimilation is used on grid 1 (see section 3 for details).

Fig. 8.

The initial root-zone soil-moisture 9 (m3 m−3) analysis for grid 3 in the control simulation. The outlined box is the grid 4

Fig. 8.

The initial root-zone soil-moisture 9 (m3 m−3) analysis for grid 3 in the control simulation. The outlined box is the grid 4

b. Emulation of a forecast environment (experiment FORE)

This experiment is identical to CON, except that a few modifications have been made in order to emulate an operational forecast setting. No data are assimilated on grid 1, and the grid 1 lateral boundary conditions are defined based on the Eta Model forecast initialized at 1200 UTC 12 July 1996, rather than on analyses of data.

c. The sensitivity to plains irrigation (experiment IRRIG)

Figure 7 shows (green) the irrigated agriculture area, as defined by the USGS EROS landuse dataset, in the eastern part of grid 3. In this experiment, the soil moisture in this area was increased from its normal value in CON to the field capacity, which is determined by the soil texture. The volumetric soil moisture for the irrigated cropland was originally between 0.24 and 0.32. It was augmented to the field capacity based on soil texture, with the new values ranging between 0.26 and 0.35. After this adjustment, most of the irrigated cropland had a volumetric soil moisture of about 0.32. Even though this area is modest in size, Chase et al. (1999) demonstrate that plains irrigation can have impacts in the mountains through modifying the horizontal temperature gradients, and therefore pressure gradients, which influence the intensity of upslope flow. Also, greater evaporation increases the upslope moisture flux. This experiment will determine the degree to which this soil-moisture perturbation in the plains modulates the relatively distant heavy convective events to the southwest near the Continental Divide. The practical question, for which insight will be provided by this sensitivity experiment, relates to how carefully we must specify finescale surface properties, both natural and anthropogenic, in order to not negatively impact forecasts of convection.

d. The sensitivity to landscape changes caused by the wildfire (experiment NOBURN)

As noted earlier, in CON, and most other experiments, the surface characteristics of the 50 km2 area burned by the wildfire were modified to reflect postfire conditions. In this experiment, the landscape properties of the area of the burn are defined as their natural values. As with IRRIG, comparison of these results with those of the control experiment will define the degree to which small-scale differences in the landscape can affect convective-scale precipitation. However, the specific nature of the effect of the burn area on convection is important here. If the modification of the surface fluxes by the burn area is such that there is a tendency for convection to focus over this flood-vulnerable area, the surface-hydrologic implications would be enormous. Even though conclusions based on one case must be viewed cautiously, a comparison of the track and intensity of convective cells in the vicinity of the burn area, for the experiments with and without the burn area, will be revealing.

e. The sensitivity to landscape-variability resolution (experiment LOWRESLS)

The previous two experiments address the sensitivity of the convection to small and isolated changes in the surface landscape. In contrast, in this experiment the surface conditions over all computational grids are defined based on the relatively coarse 1° vegetation and soil datasets that are used in the NCEP Eta Model. The soil moisture is defined in the same way as in the control, based on the NCEP–NCAR reanalysis with modifications for antecedent precipitation.

f. The sensitivity to soil-moisture-variability resolution (experiment LOWRESSM)

This experiment is the same as in CON, except that the initial soil-moisture field is based on the 2.5° NCEP–NCAR reanalysis and is not modified through the use of antecedent precipitation.

g. The sensitivity to land surface model sophistication (experiment SLAB)

The landscape characteristics and initial soil moisture are identical in this experiment to those in CON. However, the simple slab model (see section 3) is used for calculating surface sensible and latent-heat fluxes instead of the more complex LSM that employs explicit vegetation effects and time-varying soil moisture fields. The slab soil model has five layers (extending to 32 cm below the surface) for computing the soil temperature, but its soil moisture remains constant throughout the simulation. The fact that the slab model was designed to be applied at large scales, including bare-soil and vegetated areas, implies that vegetation is included implicitly. Thus, it was decided to use the root-zone soil moisture in CON to define the temporally constant soil-moisture availability parameter of the slab model. In the LSM, the vegetation roots are specified to extend from the first soil layer to the third soil layer, and the root-zone, hence, consists of the upper three soil layers, with a total depth of 1 m. Conceptually, the moisture availability (W) in the slab model represents the degree to which soil water is available for evaporation. From a hydrologic standpoint, zero moisture availability (i.e., no evaporation in slab) is equivalent to a volumetric soil moisture at the wilting point (Θw), and a moisture availability of unity (i.e., free evaporation without soil moisture stress) is equivalent to a volumetric soil moisture at the field capacity (Θref). Parameters Θref and Θw depend on the soil texture (see Chen and Dudhia 2001 for details). We can, therefore, convert the root-zone volumetric soil moisture (Θ) in the LSM to the moisture availability (W) in the slab soil model: W = (Θ − Θw)/(Θref − Θw).

h. The sensitivity to the use of climatological moisture availability with the simple slab model (experiment CLIMSM)

This experiment is analogous to the previous experiment SLAB, except that a climatological soil-moisture availability field, which is defined based on landscape-category type, is employed with the slab model rather than the one based on the NCEP–NCAR reanalysis modified by antecedent precipitation. This is the configuration of the MM5 model that has generally been employed in applications over the last 5 years.

5. Results of the experiments

For brevity, the results described here will emphasize the precipitation simulations. Because of the wide range of scales simulated over the four grids, there is an opportunity to evaluate the precipitation forecasts from the synoptic scale to the mesogamma scale. Because the grids are two-way interacting (i.e., simulations are not separately produced for each grid), it is not possible to evaluate the forecasts over the same area. The use of data assimilation on grid 1 ensures that large-scale conditions are simulated reasonably accurately, but not necessarily precipitation. On the finer grids, which are primarily over mountainous terrain, the paucity of conventional data and the representativeness of those data make it difficult to utilize conventional mesoscale analyses for model verification. Because the precipitation simulation reflects the multiscale response of the atmosphere to many dynamic processes, its verification against high-resolution radar data should reflect the overall veracity of the simulations.

a. Implications for the predictability of orographically forced summer convection

An important practical numerical forecasting issue is the following. Given a reasonable forecast of the large-scale meteorological conditions by coarser resolution operational models, to what extent can embedded high-resolution state-of-the-science mesoscale models with an accurate land surface specification predict flash-flood-scale convective rainfall events in complex terrain with a lead time of at least 6–12 h? If the convective cells occur in response to the modulation of the large-scale upslope flow by particular topographic or other surface features that are represented in the model, it is reasonable to anticipate that a simulation might have deterministic skill in predicting the locations of specific events if the larger scales are simulated reasonably. If the local forcing that triggers convection within the large-scale flow is not well represented in the model, the exact locations and timing of convective events will be relatively unpredictable. Thus, at worst, with a good simulation of the upslope flow, stability, and other synoptic-scale forcing, the overall intensity, timing, and area-coverage of the convection should be predictable. At best, with some skill at capturing the important local surface forcing, the convective event might be deterministically simulated.

It is worth noting that, even if mesoscale data were available and had been used in the initialization of this simulation, it is highly unlikely that this would have contributed to the deterministic predictability of the important convective events that did not initiate until 8 h after the model initialization.

In CON, the synoptic-scale atmospheric conditions on grid 1 are based on the NCEP EDAS analysis and MM5 data assimilation, and provide large-scale boundary conditions for grid 2. Thus, the interpretation of this experiment could be that grid 2's lateral boundary conditions are forced by an accurate large-scale forecast. A comparison of the observed rainfall in Figs. 3 and 4 with that simulated by CON in Fig. 9 shows that CON performed reasonably well in capturing the overall pattern and some of the details in the 24-h precipitation. Recall that subgrid precipitation on grids 1 and 2 is parameterized using the Grell (1993) approach. The grid 1 simulation reflects the observed southward sweeping arc-shaped area of rainfall between eastern Colorado and western Nevada, and the rainfall over Texas (Figs. 4a and 9a). The model placement of the larger accumulations in eastern Colorado, New Mexico, and Texas is also reasonable, even though there are some isolated areas in the simulated rainfall for which the totals are likely too high. However, the model overpredicts the extent of the area of light precipitation (note the zero isohyett). Grid 2, on which precipitation is also parameterized, has an area coverage that primarily reflects conditions over Colorado. The CON simulation accurately defines the overall rainfall boundary, except that it probably erroneously places accumulations in southwest Colorado. It also misses the rainfall in the Nebraska panhandle. As on grid 1, the model produces maxima that are larger in magnitude than those reflected in the lower-resolution NCEP radar estimates.

Fig. 9.

The 24-h simulated precipitation totals for grids (a) 1, (b) 2, and (c) 3 from CON. (d) The 4-h precipitation (0000–0400 UTC 13 Jul 1996) totals for grid 4 from CON, where the period spans the evening rainfall event. The grayscale legend indicates the rainfall amounts (mm)

Fig. 9.

The 24-h simulated precipitation totals for grids (a) 1, (b) 2, and (c) 3 from CON. (d) The 4-h precipitation (0000–0400 UTC 13 Jul 1996) totals for grid 4 from CON, where the period spans the evening rainfall event. The grayscale legend indicates the rainfall amounts (mm)

For grid 3 over the Front Range and plains of Colorado, Fig. 4c shows that most of the observed precipitation is limited to the steeper eastern slopes, even though it is to be expected that radar-beam blockage has caused an underestimate in the rainfall amounts and coverage over the higher elevations. However, satellite images (not shown) confirm a lack of significant convective cloud to the west of where the NCEP analysis shows accumulations. The model simulation, which explicitly represents convection on this grid, reasonably captures the western boundary of the precipitation, including the westward extension of the rainfall in the southern part of the grid (Fig. 9c). The three observed maxima with totals in excess of 25 mm in the southern half of the grid have approximate counterparts in the simulation (including the Buffalo Creek maximum). However, the model erroneously also places heavy precipitation farther north, over the plains near the eastern boundary of the grid.

For the highest resolution grid, we focus on the 4-h (0000–0400 UTC 13 July 1996) accumulated precipitation that reflects the effects of the thunderstorm that caused the flash-flood event. Our primary interest is in the timing and location of this convection. The S-Pol-based analysis in Fig. 3 shows a rainfall maximum of about 80 mm centered over the northeast, lower-elevation, part of the Buffalo Creek watershed. The CON simulation (Fig. 9d) produces an area of precipitation that is modestly larger than that reflected in the observations, but the maximum (90 mm) is in approximately the correct location. The model also simulates a secondary maximum to the northeast of the main convective event, but it is located too close to the watershed compared to the secondary maximum observed by radar (Fig. 3). As with the S-Pol radar total in Fig. 3, the model-based total for the period results from contributions by a number of individual cells that pass over the watershed in a brief period. Figure 10 shows the paths of individual simulated convective cells that contribute to the total in Fig. 9d. The earlier tracks occurred to the northeast of the watershed, and as the evening progressed, the path centers moved more to the south over the watershed.

Fig. 10.

The paths of individual simulated convective cells that contribute to the total in Fig. 9d. The cell tracks are identified with the initiation time (MST) and location, and the subsequent simulated locations of the maxima of the 5-min accumulations (dots), plotted at 5-min intervals

Fig. 10.

The paths of individual simulated convective cells that contribute to the total in Fig. 9d. The cell tracks are identified with the initiation time (MST) and location, and the subsequent simulated locations of the maxima of the 5-min accumulations (dots), plotted at 5-min intervals

Because the human impact of this convective event was related to the flash flood that was produced, our objective verification of CON and the various sensitivity experiments will be keyed to the aspect of the simulations that has relevance to the flood; that is, how much of the simulated rainfall fell over the Buffalo Creek watershed where the flood occurred, and especially over the burn area where the runoff-infiltration ratio was very high. This hydrologically oriented metric of model rainfall-simulation error (shown in Table 2) is, to our knowledge, relatively unique, and is arguably more appropriate than conventional measures, such as threat scores, for applications such as this. Because most of the runoff that caused the flash flood in Buffalo Creek likely occurred over the burn area's intersection with the watershed, this bias is perhaps the most relevant. For CON, this number is very close to unity.

Table 2.

Bias (simulated/observed) of rainfall for a period encompassing the flood-producing convective event (0000–0348 UTC, as in Fig. 3), relative to the S-Pol radar-estimated rainfall of Fig. 3, over the Buffalo Creek watershed (BCWS), over the burn area (BURN), and over the overlap of BCWS and BURN

Bias (simulated/observed) of rainfall for a period encompassing the flood-producing convective event (0000–0348 UTC, as in Fig. 3), relative to the S-Pol radar-estimated rainfall of Fig. 3, over the Buffalo Creek watershed (BCWS), over the burn area (BURN), and over the overlap of BCWS and BURN
Bias (simulated/observed) of rainfall for a period encompassing the flood-producing convective event (0000–0348 UTC, as in Fig. 3), relative to the S-Pol radar-estimated rainfall of Fig. 3, over the Buffalo Creek watershed (BCWS), over the burn area (BURN), and over the overlap of BCWS and BURN

Figure 11 depicts the temporal distribution of the observed and simulated grid-total rainfall for each of the four grids, where rainfall in embedded inner grids is excluded. As mentioned earlier, CON captures well the spatial pattern of 24-h rainfall in grid 1, but it does not produce a distinctively diurnal cycle with a maximum rainfall at 2100 UTC (1400 MST) as observed. In CON, and in the other simulations, there is unrealistically intense nocturnal rainfall in the last half of the period. Figures 4a and 9a show this excess to be in northern New Mexico and northeastern Texas. Nevertheless, the simulation compares fairly well with observations in terms of the diurnal cycle for grids 2 and 3. For grid 4, CON correctly simulates the two major rainfall events (1800–2100 UTC and 0000–0400 UTC, simulation times 6–9 h and 12–16 h), but it produces an apparently erroneous peak at about 2200 UTC (simulation time 10 h).

Fig. 11.

Diurnal variation of grid-total rainfall for each model grid, based on observations and all simulations. The ordinate scale is grid-total rain mass, in kg (30 min)−1 multiplied by the factor at the top of each column. The number inset in the upper-right corner of each plot is the 24-h, grid-total rain mass in kilograms

Fig. 11.

Diurnal variation of grid-total rainfall for each model grid, based on observations and all simulations. The ordinate scale is grid-total rain mass, in kg (30 min)−1 multiplied by the factor at the top of each column. The number inset in the upper-right corner of each plot is the 24-h, grid-total rain mass in kilograms

Experiment FORE was designed to evaluate the model's ability to predict this flash-flood-producing convective event in an actual forecast environment. Thus, no data assimilation was used on grid 1, and the Eta Model forecast (initialized at the same time as MM5) was used for the lateral boundary conditions of grid 1. The spatial patterns of the 24-h total rainfall for grids 1, 2, and 3 (not shown) for FORE are quite similar to those of CON, but FORE produces less rainfall in northern New Mexico and Texas. Weak simulated convection in northeast Minnesota, which is absent in CON, agrees better with observations. In general, FORE tends to produce weaker but more widespread convection than does CON. On grid 4 (Fig. 12), the simulated FORE rainfall amounts are less than from CON and the observations, but heavy rainfall is correctly simulated over the northeastern part of the Buffalo Creak watershed. Table 2 shows that there was a considerable low bias to the rainfall simulated over the watershed and the burn area. In Fig. 11, the diurnal variation of the grid-total rainfalls in FORE are similar to those for CON. The temporal distribution from FORE on grid 4 actually agrees better with observations, but the rainfall corresponding to the flash-flood event (the second maximum) is weaker.

Fig. 12.

The 4-h (0000–0400 UTC 13 Jul 1996) rainfall totals for grid 4 for all experiments. The grayscale legend indicates the rainfall amounts, same as in Fig. 9 

Fig. 12.

The 4-h (0000–0400 UTC 13 Jul 1996) rainfall totals for grid 4 for all experiments. The grayscale legend indicates the rainfall amounts, same as in Fig. 9 

The sensitivity simulations to be described in the next sections show that interactions between local landscape properties and the large-scale meteorological conditions seem to play an important role in determining the location and amount of the evening heavy rainfall. As a first step to better understanding the nature of the interactions and how they focus convection, the evolution will be discussed of some important surface fields in CON and how they might provide insight into the causes of the observed rainfall distribution. This has implications about the predictability of the event. Figures 13 and 14 show the simulated upper-soil-level moisture and temperature, respectively, on grid 4 at three times (1100, 1400, 1700 MST) during the morning and afternoon prior to the evening convective event in CON. Figure 15 displays the 3-h rainfall totals from CON for the three periods that precede the times for which soil moisture and temperature are displayed in Figs. 13 and 14. Due to the initial dry soil conditions, the soil moisture remains low before the onset of significant simulated rainfall (Fig. 13a, 1100 MST). Note that simulated light morning rainfall (see the 0.1-mm isohyett in Fig. 15a) wetted the vegetation canopy to the west of the watershed, but not the soil significantly. At this time, the distribution of the soil temperature primarily reflects elevation and surface-property variability (Fig. 14a). In the early afternoon, light convective rain is simulated in a number of areas (Fig. 15b). The cloud shading and the soil moistening (Fig. 13b, 1400 MST) lead to the surface temperature pattern that is reflected in Fig. 14b for 1400 MST. In the burn area, the lower albedo and drier, unvegetated surface produce surface temperatures that are a few degrees Celsius higher than the surrounding ambient values. By late afternoon, (1700 MST), the soil temperature has been further influenced by the evaporation and cloud-shading effects of the afternoon rainfall (Fig. 14c). Dominant features include a warm surface to the northwest of the watershed that is flanked by cooler areas to its northeast and southwest, where the cool areas are associated with antecedent afternoon rainfall events. Between the warm and cool areas, the thermal contrast is as high as 6°C over a few kilometers. Also, the surface temperature anomaly associated with the burn area is now in excess of 5°C. This surface high-temperature anomaly can potentially enhance the development of moist convection in an unstable environment through the generation of upward motion and lower static stability. A comparison of the northwest–southeast path of the main convective cell in CON, as seen in the rainfall totals of Fig. 9d, and the simulated surface soil temperature (Fig. 14c) and moisture (Fig. 13c) distributions in the late afternoon prior to the development of this cell, reveals a possible relationship between the path of the convective event, and the antecedent rainfall pattern. The simulated flood-producing convective event triggers to the northwest of the watershed over the warm area, and then follows the axis of the warm anomaly (that includes the burn area) to the southeast. Thus, even though the track of the convection was determined by the prevailing steering flow, the fact that the path was aligned with the warm-surface anomaly could have contributed to the severity of the convection.

Fig. 13.

The volumetric upper-soil-layer moisture (m3 m−3) on grid 4 from CON at (a) 1100, (b) 1400, and (c) 1700 MST 12 Jul 1996. The isopleth interval is 0.01 m3 m−3

Fig. 13.

The volumetric upper-soil-layer moisture (m3 m−3) on grid 4 from CON at (a) 1100, (b) 1400, and (c) 1700 MST 12 Jul 1996. The isopleth interval is 0.01 m3 m−3

Fig. 15.

The 3-h rainfall from CON for the periods ending at (a) 1100, (b) 1400, and (c) 1700 MST 12 Jul 1996. The grayscale legend indicates the rainfall amounts (mm)

Fig. 15.

The 3-h rainfall from CON for the periods ending at (a) 1100, (b) 1400, and (c) 1700 MST 12 Jul 1996. The grayscale legend indicates the rainfall amounts (mm)

Fig. 14.

The upper-soil-layer temperature (K) on grid 4 from CON at (a) 1100, (b) 1400, and (c) 1700 MST 12 Jul 1996. The isotherm interval is 2 K

Fig. 14.

The upper-soil-layer temperature (K) on grid 4 from CON at (a) 1100, (b) 1400, and (c) 1700 MST 12 Jul 1996. The isotherm interval is 2 K

If this speculation about the relationship between the evening convection and the distribution of antecedent rainfall is correct, the predictability of the major evening event is thus related to the predictability of the precursor rainfall events during the day. In fact, the WSR-88D radar-based rainfall analysis for grid 4 for the morning and afternoon compare favorably with the model-simulated precursor rainfall during the day, shown in Fig. 15. That is, there were two rainfall areas observed upstream of the watershed, in locations similar to those simulated and shown in Fig. 15b. Thus some combination of topography and other landscape properties perhaps determined the similar locations of the precursor rainfall in both the model and atmosphere.

b. Impact of the land surface models on the simulated convection

Two numerical experiments, SLAB and CLIMSM, were conducted to investigate the impact of different physical treatments of land surface processes on this simulation of rainfall. Both simulations utilize the simple slab model. Experiment SLAB employs a surface-moisture availability field that is based on the root-zone soil moisture used in CON. Thus, the only difference between CON and SLAB is the method of parameterization of land surface processes. In CLIMSM, a climatological surface-moisture availability is used, as was standard in earlier releases of MM5. Note that, unlike in CON that used the LSM, the soil-moisture availability remains constant in these two simulations.

Both SLAB and CLIMSM produce large-scale patterns of 24-h rainfall for grids, 1, 2, and 3 (not shown) that are quite similar to those from CON, but SLAB has larger precipitation amounts (Fig. 11). Figure 12 shows the rainfall totals on grid 4 for the evening convective event for SLAB and CLIMSM. Experiment SLAB produces a mere 3 mm of rainfall in the northern part of the Buffalo Creek watershed, and hence totally misses the event. The CLIMSM experiment produces somewhat more rainfall than does SLAB, but the convection is too weak and is displaced to the north and west of where it was observed.

To illustrate the importance of the land surface physics parameterization to the correct simulation of this convective event, the comparison of CON and SLAB will be the focus because differences can be largely attributed to the different land surface models (recall that the constant SLAB soil-moisture availability was derived from the initial LSM soil moisture in CON). On grid 4, the SLAB simulation produces larger early afternoon rainfall amounts than does CON, while the latter correctly has larger rainfall amounts in the late afternoon and evening (Fig. 11). The latent heat flux in SLAB is much higher than that in CON in late morning and early afternoon (not shown), and in some areas it is 200 W m−2 higher. This higher latent-heat flux in SLAB resulted in a cooler surface, a lower sensible-heat flux, and less development of the PBL. This results in a more-moist and shallow PBL with larger moist static energy, which is conducive to convective activity in the early afternoon. If, in fact, the evening convection in CON is modulated by the effects of the afternoon convection, as speculated earlier, there is clearly no way that SLAB could correctly capture the evening event because there is no connection between simulated rainfall and soil moisture. The accuracies for these two simulations of the rainfall over the watershed and the burn area are very low (Table 2).

It is interesting that CLIMSM produces a more realistic rainfall distribution than does SLAB that uses the actual estimated soil moisture. This could be related to the fact that the SLAB model and the soil-moisture climatology maps have been used together for many years, and the maps may have undergone some adjustment to produce better simulations when used with the SLAB model.

c. Sensitivity of orographic summer convection to local and regional landscape variability

1) Sensitivity to landscape conditions at small scales

Two experiments involve small-scale changes to the landscape characteristics of CON. In NOBURN, the surface conditions in the area of the wildfire (Fig. 6b) are defined in terms of their preburn values. In IRRIG, the soil moisture in a modest sized area of the plains (see Fig. 7 for the irrigated crop area) is set at field capacity to reflect the existence of irrigation.

The temporal distributions of rainfall in Fig. 11 for the inner two model grids show some differences between NOBURN and CON. The existence of the 50 km2 burn area has a 5%–10% influence on the grids 3 and 4 total rainfall. On grid 4, the rainfall maximum to the northeast of the watershed in NOBURN is too large in magnitude (Fig. 12), and the CON maximum is more correctly located closer to the northern part of the watershed and the burn area. Over the northern part of the watershed, rainfall from the flood-producing convective event is over 25 mm greater with the burn area than without it. The spatial distributions of the ground temperature at 0000 UTC (1700 MST) 13 July in NOBURN (not shown) and CON (Fig. 14c) show that the existence of the burn caused simulated late afternoon surface temperatures there to be about 5°C warmer. Thus, the modification of the landscape by the wildfire had the net consequence of moving the simulated convective rainfall maximum closer to the burned area that exhibited maxima in the sensible heat flux (not shown) and near-surface temperature. Note that the rainfall biases in Table 2 are closer to unity when the burn area is accounted for.

In IRRIG, the relatively small-scale change in initial soil moisture over the plains on grid 3 has a significant influence on the rainfall amount and temporal distribution on grids 3 and 4 (compare CON and IRRIG in Figs. 11 and 12). For example, the heavy rainfall area to the northeast of the watershed is significantly greater in extent in IRRIG, but the rainfall over the watershed is less. This sensitivity is consistent with Chase et al. (1999), who describe the use of a regional model for sensitivity studies in which they replaced grassland with irrigated and dry farmland on the northern Colorado plains in multiday simulations. They found that these landscape changes on scales similar to those in IRRIG could have significant impacts on mesoscale weather and climate, in areas distant from the direct forcing. Our study further indicates that, even on timescales as short as 12 h, changes in initial soil moisture can affect the development of convection in distant areas. This is likely to be through differences in local PBL heating, whose effects can be transmitted rapidly through mass-field adjustments by gravity waves, rather than through advection.

2) Sensitivity to landscape conditions at large scales

Results are presented here for two numerical experiments in which the land surface conditions were changed over the entire computational area. In one experiment (LOWRESLS), the 1° resolution vegetation and soil datasets that have a spatial resolution similar to those used in most operational forecasting models are utilized instead of the 1-km datasets. The other (LOWRESSM) is otherwise identical to CON except that the 2.5° soil-moisture fields from the NCEP–NCAR reanalysis are used as initial conditions without being adjusted by the antecedent precipitation.

The temporal evolution of the grid total rainfall on both grids 1 and 2 for LOWRESLS (Fig. 11) is close to that of CON. On grid 3, LOWRESLS has significantly more rainfall and a distinct peak near 0000 UTC, which does not have a counterpart in CON and the observations. On grid 4, even with these low-resolution vegetation and soil datasets, the rainfall maximum over the northern part of the watershed is captured well (Fig. 12), but the simulation tends to overestimate the maximum rainfall and it misses entirely the secondary maximum farther to the northeast.

In LOWRESSM, utilizing the coarse-resolution initial soil moisture also produces significant changes in the grid-total rainfall on grids 3 and 4 (Fig. 11). On grid 4, the most significant difference from CON is the displacement of the main rainfall maximum to the southeast (Fig. 12).

In both of these simulations, the evolution of the surface temperature on grid 4 is similar to that in CON (Fig. 14c). Specifically, both simulate the warm area to the northwest of the watershed that is embedded between the relatively cooler areas. However, the thermal gradient between the warm area and surrounding cool areas is weaker in these two simulations than in CON. In contrast to the poor rainfall simulations in SLAB and CLIMSM, the utilization of a more advanced land surface model in these simulations, together with time varying soil moisture, produced reasonable simulations in spite of the coarse resolution of the landscape and initial soil moisture.

6. Summary and discussion

Two different land surface–process parameterizations and a variety of different specifications of the large- and small-scale structure of the landscape variability were tested in convection-resolving model simulations of the Buffalo Creek, Colorado, flash flood–producing convection. In all the simulations, the model produced reasonably good estimates of the large-scale rainfall. On the convection-resolving scales, the spatial and temporal distributions of the simulated rainfall amounts were strongly dependent on the model landscape specification and land surface physics. However, the use of the more sophisticated and realistic treatment of the land surface physics was more critical than the use of high-resolution landscape variability in terms of producing heavy rainfall to the northeast of the Buffalo Creek watershed. Indeed, all the simulations with the more complex land surface physics produced heavy rainfall that correctly intersected the northeastern end of the watershed. However, the exact location, intensity, and spatial extent of the evening heavy rainfall on the highest resolution grid were found to be very sensitive to the details of the landscape variability and initial soil moisture. Importantly, the modification to the landscape properties by the wildfire at the northeastern end of the watershed, and the resulting higher daytime surface temperature, have the apparent effect of influencing the track of the convection such that it is closer to the burned area, rather than more to the northeast as seen in a simulation without the burned landscape.

Even though the primary motive of this study is to help better understand the sensitivity of heavy convection to landscape variability in complex terrain, there are equally important and related questions raised about the predictability of such severe convective weather. We speculated earlier that perhaps the best 12–24 h convection-resolving model predictions that might be hoped for in such upslope-flow situations would be the correct timing, intensity, and general area of the rainfall. That is, unless the convection is initiated by some major land surface property, or combination of properties, that we can accurately account for in the model, we will not be able to predict heavy rainfall that far in advance with sufficient specificity to locate it in particular mountain watersheds. The success with which this Buffalo Creek thunderstorm event was simulated by a convection-resolving model with good land surface physics certainly does not refute this statement. Rather, perhaps the point should be that the local forcing of convection in complex terrain (and elsewhere) can be through rather subtle mechanisms, and that occasional forecast successes can be expected so long as we properly treat the surface physics and capture the landscape variability at finescales.

There are clearly many characteristics of the modeling system, in addition to the treatment of the land surface forcing, that likely had a quantitative impact on the simulated rainfall. The sensitivities to, for example, the microphysics, the PBL parameterization, and the convective parameterization on the coarse grids, were not evaluated. And, it is likely that particularly poor treatments of any of these processes may have negated the validity of the sensitivity studies—for example, in the extreme, if no rainfall had been produced in CON. In any case, the fact that the “default MM5 physics” produced a reasonably good rainfall simulation in CON can give us some confidence that the model configuration did not have a qualitative effect on the results of the sensitivity experiments.

A related question that was addressed is the degree to which the success of the control simulation here depends on its retrospective nature. To answer that, the lateral boundary conditions of the outer grid were forced with the operational Eta Model forecast for the 24-h period. This did cause some deterioration in the forecast quality, compared to CON, but a major event was still simulated near the burn area, which would have raised appropriate concern for life and property if the information had been available operationally.

Finally, an important outcome is that the representation of the burn area in the model caused an enhancement of the convection over that area. Even though it is uncertain how to properly model the land surface physics there, if this response is correct the practical impact is great. Yates et al. (2000) use a runoff and discharge model to estimate that the discharge of Buffalo Creek at the exit of the watershed was a factor of four greater because of the burn. Additionally, Jarrett (2000, personal communication) states that the convective activity near the watershed has been enhanced since the time of the wildfire. This small area of temperature excess is almost certainly incapable of initiating convection through the development of a significant local circulation. However, it is perhaps not unreasonable that the greater instability in that region could focus existing convection, or that an atmosphere that is on the threshold of triggering convection might first initiate convection there.

Fig. 11.

(Continued)

Fig. 11.

(Continued)

Acknowledgments

This research was supported through special funds from the National Science Foundation that have been designated for the U.S. Weather Research Program activities at NCAR. The authors benefited from many useful discussion of the case with Robert Jarrett of the U.S. Geological Survey.

REFERENCES

REFERENCES
Anthes
,
R. A.
,
1984
:
Enhancement of convective precipitation by mesoscale variations in vegetative covering in semiarid regions.
J. Climate Appl. Meteor
,
23
,
541
554
.
Avissar
,
R.
, and
T.
Schmidt
,
1998
:
An evaluation of the scale at which ground-surface heat flux patchiness affects the convective boundary layer using large-eddy simulations.
J. Atmos. Sci
,
55
,
2666
2689
.
Benjamin
,
S. G.
, and
N. L.
Seaman
,
1985
:
A simple scheme for objective analysis in curved flow.
Mon. Wea. Rev
,
113
,
1184
1198
.
Blackadar
,
A. K.
,
1976
:
Modeling the nocturnal boundary layer.
Preprints, Third Symp. on Atmospheric Turbulence, Diffusion and Air Quality, Raleigh, NC, Amer. Meteor. Soc., 46–49
.
Blackadar
,
A. K.
,
1979
:
High resolution models of the planetary boundary layer.
Advances in Environmental Science and Engineering, J. Pfafflin and E. Ziegler, Eds., Vol. 1, Gordon and Breach, 50–85
.
Chang
,
J-T.
, and
P. J.
Wetzel
,
1991
:
Effects of spatial variations of soil moisture and vegetation on the evolution of a prestorm environment: A numerical case study.
Mon. Wea. Rev
,
119
,
1368
1390
.
Chase
,
T. N.
,
R. A.
Pielke Sr.
,
T. G. F.
Kittel
,
J. S.
Baron
, and
T. S.
Stohlgren
,
1999
:
Potential impacts on Colorado Rocky Mountain weather due to land use changes on the adjacent Great Plains.
J. Geophys. Res
,
104
,
16673
16690
.
Chen
,
F.
, and
J.
Dudhia
,
2001
:
Coupling an advanced land-surface/hydrology model with the Penn State/NCAR MM5 modeling system. Part I: Model implementation and sensitivity.
Mon. Wea. Rev
,
129
,
569
585
.
Chen
,
F.
, and
and Coauthors
,
1996
:
Modeling of land-surface evaporation by four schemes and comparison with FIFE observations.
J. Geophys. Res
,
101
,
7251
7268
.
Dalu
,
G. A.
, and
R. A.
Pielke
,
1989
:
Analytical study of sea breeze.
J. Atmos. Sci
,
46
,
1815
1825
.
Droegemeier
,
K.
,
1997
:
The numerical prediction of thunderstorms: Challenges, potential benefits, and results from real-time operational tests.
WMO Bull
,
46
,
324
336
.
Dudhia
,
J.
,
1989
:
Numerical study of convection observed during the winter monsoon experiment using a mesoscale two-dimensional model.
J. Atmos. Sci
,
46
,
3077
3107
.
Dudhia
,
J.
,
1993
:
A nonhydrostatic version of the Penn State/NCAR mesoscale model: Validation tests and the simulation of an Atlantic cyclone and cold front.
Mon. Wea. Rev
,
121
,
1493
1513
.
Dudhia
,
J.
,
1996
:
A multi-layer soil-temperature model for MM5.
Preprint, Sixth PSU/NCAR Mesoscale Model Users' Workshop, Boulder, CO, 49–50
.
Fritsch
,
J. M.
, and
and Coauthors
,
1998
:
Quantitative precipitation forecasting: Report of the Eighth Prospectus Development Team, U.S. Weather Research Program.
Bull. Amer. Meteor. Soc
,
79
,
285
299
.
Georgakakos
,
K. P.
, and
M. D.
Hudlow
,
1984
:
Quantitative precipitation forecast techniques for use in hydrologic forecasting.
Bull. Amer. Meteor. Soc
,
65
,
1186
1200
.
Grell
,
G. A.
,
1993
:
Prognostic evaluation of assumptions used by cumulus parameterizations.
Mon. Wea. Rev
,
121
,
764
787
.
Grell
,
G. A.
,
J.
Dudhia
, and
D. R.
Stauffer
,
1994
:
A description of the fifth generation Penn State/NCAR mesoscale model (MM5).
NCAR Tech. Note NCAR/TN 398+STR, 138 pp
.
Hong
,
S-Y.
, and
H-L.
Pan
,
1996
:
Nonlocal boundary layer vertical diffusion in a medium-range forecast model.
Mon. Wea. Rev
,
124
,
2322
2339
.
Jacquemin
,
B.
, and
J.
Noilhan
,
1990
:
Sensitivity study and validation of a land surface parameterization using the HAPEX-MOBILHY data set.
Bound.-Layer Meteor
,
52
,
93
134
.
Kalnay
,
E.
, and
and Coauthors
,
1996
:
The NCEP/NCAR 40-Year Reanalysis Project.
Bull. Amer. Meteor. Soc
,
77
,
437
471
.
Loveland
,
T. R.
,
J. W.
Merchant
,
J. F.
Brown
,
D. O.
Ohlen
,
B. C.
Reed
,
P.
Olson
, and
J.
Hutchinson
,
1995
:
Seasonal land-cover regions of the United States.
Ann. Assoc. Amer. Geogr
,
85
,
339
355
.
Mahrt
,
L.
, and
K.
Ek
,
1984
:
The influence of atmospheric stability on potential evaporation.
J. Climate Appl. Meteor
,
23
,
222
234
.
Mahrt
,
L.
, and
H-L.
Pan
,
1984
:
A two-layer model of soil hydrology.
Bound.-Layer Meteor
,
29
,
1
20
.
Miller
,
D. A.
, and
R. A.
White
,
1998
:
A conterminous United States multi-layer soil characteristics data set for regional climate and hydrology modeling.
Earth Interactions, 2. [Available online at http://EarthInteractions.org.]
.
Nair
,
U. S.
,
H. R.
Mark
, and
R. A.
Pielke Sr.
,
1997
:
Numerical simulation of the 9–10 June 1972 Black Hills storm using CSU RAMS.
Mon. Wea. Rev
,
125
,
1753
1766
.
Pan
,
H-L.
, and
L.
Mahrt
,
1987
:
Interaction between soil hydrology and boundary-layer development.
Bound.-Layer Meteor
,
38
,
185
202
.
Physick
,
W. L.
, and
N. J.
Tapper
,
1990
:
A numerical study of circulations induced by a dry salt lake.
Mon. Wea. Rev
,
118
,
1029
1042
.
Schaake
,
J. C.
,
V. I.
Koren
,
Q. Y.
Duan
,
K.
Mitchell
, and
F.
Chen
,
1996
:
A simple water balance model (SWB) for estimating runoff at different spatial and temporal scales.
J. Geophys. Res
,
101
,
7461
7475
.
Segal
,
M.
,
W. E.
Schreiber
,
G.
Kallos
,
J. R.
Garratt
,
A.
Rodi
,
J.
Weaver
, and
R. A.
Pielke
,
1989
:
The impact of crop areas in northeast Colorado on midsummer mesoscale thermal circulations.
Mon. Wea. Rev
,
117
,
809
825
.
Segal
,
M.
,
J. H.
Cramer
,
R. A.
Pielke
,
J. R.
Garratt
, and
P.
Hildebrand
,
1991
:
Observational evaluation of the snow breeze.
Mon. Wea. Rev
,
119
,
412
424
.
Stauffer
,
D. R.
, and
N. L.
Seaman
,
1990
:
Use of four-dimensional data assimilation in a limited-area mesoscale model. Part I: Experiments with synoptic data.
Mon. Wea. Rev
,
118
,
1250
1277
.
Stauffer
,
D. R.
,
N. L.
Seaman
, and
F. S.
Binkowski
,
1991
:
Use of four-dimensional data assimilation in a limited-area mesoscale model. Part II: Effects of data assimilation within the planetary boundary layer.
Mon. Wea. Rev
,
119
,
734
754
.
Sun
,
J.
, and
N. A.
Crook
,
1994
:
Wind and thermodynamic retrieval from single-Doppler measurements of a gust front observed during Phoenix II.
Mon. Wea. Rev
,
122
,
1075
1091
.
Sun
,
J.
, and
N. A.
Crook
,
1997
:
Dynamical and microphysical retrieval from Doppler radar observations using a cloud model and its adjoint: Part I: Model development and simulated data experiments.
J. Atmos. Sci
,
54
,
1642
1661
.
Warner
,
T. T.
, and
H. M.
Hsu
,
2000
:
Nested-model simulation of moist convection: The impact of coarse-grid parameterized convection on fine-grid resolved convection through lateral-boundary-condition effects.
Mon. Wea. Rev
,
128
,
2211
2231
.
Warner
,
T. T.
,
Y-H.
Kuo
,
J. D.
Doyle
,
J.
Dudhia
,
D. R.
Stauffer
, and
N. L.
Seaman
,
1992
:
Nonhydrostatic, mesobeta-scale real-data simulations with the Penn State University/National Center for Atmospheric Research mesoscale model.
Meteor. Atmos. Phys
,
49
,
209
227
.
Warner
,
T. T.
,
E. A.
Brandes
,
J.
Sun
,
D. N.
Yates
, and
C. K.
Mueller
,
2000:
:
Prediction of a flash flood in complex terrain. Part I: A comparison of rainfall estimates from radar, and very short range rainfall simulations from a dynamic model and an automated algorithmic system.
J. Appl. Meteor
,
39
,
797
814
.
Xu
,
O.
,
1996
:
Generalized adjoint for physical processes with parameterized discontinuities. Part I: Basic issues and heuristic examples.
J. Atmos. Sci
,
53
,
1123
1142
.
Yan
,
H.
, and
R. A.
Anthes
,
1988
:
The effect of variations in surface moisture on mesoscale circulations.
Mon. Wea. Rev
,
116
,
192
208
.
Yates
,
D. N.
,
T. T.
Warner
,
E. A.
Brandes
,
J.
Sun
,
C. K.
Mueller
, and
G. H.
Leavesley
,
2000
:
Prediction of a flash flood in complex terrain: A comparison of flood discharge simulations using rainfall input from radar, a dynamic model, and an automated-algorithmic system.
J. Hydrol. Eng
,
6
,
265
274
.
Zhang
,
D-L.
, and
R. A.
Anthes
,
1982
:
A high-resolution model of the planetary boundary layer—Sensitivity tests and comparisons with SESAME-79 data.
J. Appl. Meteor
,
21
,
1594
1609
.

Footnotes

Corresponding author address: Fei Chen, National Center for Atmospheric Research/RAP, P.O. Box 3000, Boulder, CO 80307–3000. Email: feichen@ucar.edu

*

The National Center for Atmospheric Research is sponsored by the National Science Foundation.