Abstract

The use of dense pressure observations is investigated for creating mesoscale ensemble analyses and improving short-term mesoscale forecasts. By exploiting additional observation platforms, the number of pressure observations over the Pacific Northwest region is increased by an order of magnitude over standard airport observations. Quality control and bias correction methods for these observations are discussed, including the use of pressure tendency as an alternative observation type with fewer bias concerns. The enhanced station density provided by these observations contributes to localized adjustments for a variety of mesoscale phenomena. These adjusted analyses yield improved forecasts, including more accurate forecasts of frontal passages and convective bands. Assimilating dense 3-h pressure tendency observations also reduces the error in some forecast surface fields similarly to raw pressure observations, suggesting further investigation into pressure tendency as a mesoscale observation type.

1. Introduction

Short-term numerical weather forecasts continue to suffer from poor definition and prediction of mesoscale weather features (e.g., Roebber et al. 2004). This is particularly true for small-scale, but potentially high-impact features such as the timing and structure of frontal passages (Colle et al. 2001) or convective development and evolution (Melhauser and Zhang 2012; Hanley et al. 2013). Not only can these features present significant hazards to public safety, but numerous industries, from aircraft operations to wind energy, could benefit from more accurate prediction of mesoscale phenomena. As noted by Eckel and Mass (2005), the ability to forecast small-scale variability is highly dependent on both model resolution and the scales of variability with which the model is initialized. This paper investigates the utility of dense surface pressure observations for initializing a high-resolution model with the scales of variability necessary to forecast mesoscale, short-term, high-impact weather events more accurately.

Surface pressure observations are chosen because they exhibit desirable properties for data assimilation and model initialization. Representativeness error can be significant for temperature and wind observations (e.g., Benjamin et al. 1999; Janjic and Cohn 2006; Ancell et al. 2011) due to their high sensitivity to the immediate surroundings. Pressure observations should not suffer as significantly from representativeness error since pressure tends be more spatially homogeneous and pressure measurement is not as dependent upon instrument exposure. The potential for a systematic bias does exist with pressure observations, but this issue may be addressed effectively in a variety of ways as will be discussed in this paper.

Pressure observations can also provide information about the three-dimensional structure of the atmosphere through, for example, the covariance between surface pressure and geopotential height (through hydrostatic balance), which extends the influence of surface pressure observations beyond the boundary layer. These properties formed the basis for the twentieth-century reanalysis project (Whitaker et al. 2004; Compo et al. 2011), which uses primarily surface pressure observations. In idealized perfect-model experiments, Anderson et al. (2005) show that errors in the tropospheric mass field of a general circulation model are reduced by increasing the density and frequency of assimilated surface pressure observations. Similarly, Dirren et al. (2007) show that regional assimilation of pressure observations from the automated surface observing system (ASOS) network is able to capture the synoptic-scale upper-air pattern over western North America and the eastern Pacific Ocean.

While these studies have shown the ability of surface pressure observations to constrain synoptic-scale patterns due to the long covariance length scales of surface pressure, less attention has been paid to the role of pressure observations in constraining mesoscale phenomena. Many mesoscale weather features are known to exhibit distinct pressure signatures. For example, Mass and Ferber (1990) find pressure perturbations of up to 10 hPa in association with flow around an isolated topographical barrier. Localized pressure perturbations on the order of 2–5 hPa have been reported in association with strong frontal passages and in convective cold pools (Goff 1976; Engerer et al. 2008). Wheatley and Stensrud (2010) show that assimilating surface pressure observations and, to some extent, 1-h surface pressure tendency observations, increases the accuracy of modeled convective cold pools.

Adequately resolving pressure signals associated with localized mesoscale weather phenomena requires a suitably dense network of observations. Fortunately, surface pressure or altimeter setting observations are reported from most weather observing platforms. Many regional-scale operational data assimilation systems regularly assimilate pressure observations from ASOS and automated weather observing system (AWOS) stations, collectively referred to here as aviation routine weather reports (METARs). However, these networks have observations at an average spatial separation of around 100 km (Wheatley and Stensrud 2010), which may be insufficient to capture small mesoscale features. To further increase the observation density, this research explores observations from other readily available, but underutilized sources, particularly from privately owned weather stations in networks such as the Citizen Weather Observer Program (CWOP; http://www.wxqa.com) and the Weather Underground (http://www.wunderground.com). The addition of these observation networks greatly enhances the density of available observations to a level where it is possible to resolve the pressure signal of many localized phenomena. This study builds on the results of Wheatley and Stensrud (2010) and Dirren et al. (2007) in examining the ability of pressure observations to describe mesoscale features (with scales of 20–200 km) by introducing high-density pressure observations to the assimilation process.

The impact of assimilating high-density altimeter observations is examined by focusing on the Pacific Northwest region of the United States. The mesoscale weather phenomena and complex geography of this region provide unique forecasting challenges and a rich suite of phenomena for evaluating a regional mesoscale prediction system (Ancell et al. 2011). Furthermore, regional ensemble prediction and data assimilation over this region is currently being performed through the University of Washington real-time ensemble Kalman filter system (UW-rtEnKF; Torn and Hakim 2008), which provides a preexisting ensemble framework that is used throughout this study.

Ensemble Kalman filter data assimilation is the chosen method to combine observations with ensemble estimates of the atmospheric state. This data assimilation technique has been widely used for operational and research weather data assimilation across all scales, including regional and mesoscale applications (e.g., Anderson et al. 2005; Dirren et al. 2007; Ancell et al. 2011). In particular, the flow-dependent covariances from ensemble estimates can increase the utility of pressure observations over using a fixed covariance. This method allows rigorous testing of the hypothesis that ensemble-based data assimilation with a dense network of surface pressure observations, coupled with innovative observation quality control and bias removal techniques, has the potential to yield substantially improved mesoscale analyses and forecasts.

2. Methods

a. Pressure observation sources

Most regional data assimilation systems limit surface observations to METAR observations, which include only data from the ASOS and AWOS networks (e.g., Dirren et al. 2007; Wheatley and Stensrud 2010). However, these networks may not have sufficient spatial density to adequately resolve smaller mesoscale phenomena. To increase the number of observations, several additional observation sources are sought. Observations from publicly available networks within the Pacific Northwest region (Fig. 1) yield an additional 1300–1700 hourly pressure observations beyond the METAR observations. The sources of these observations are summarized in Table 1. Figure 1 shows the spatial density of these additional observations as compared with the METAR observations. The most dramatic increases in density are apparent in urban areas, but there is noticeable improvement in spatial density throughout much of the region.

Fig. 1.

Locations of ASOS/METAR, buoy, and ship altimeter observations (red) and all other altimeter locations for the period of study (blue).

Fig. 1.

Locations of ASOS/METAR, buoy, and ship altimeter observations (red) and all other altimeter locations for the period of study (blue).

Table 1.

Sources of observations used.

Sources of observations used.
Sources of observations used.

Table 1 shows that over three-quarters of the available pressure observations come from loosely regulated networks (e.g., Weather Underground, Citizen Weather Observer Program) and as such there is little guarantee of observation quality. Before using these observations for data assimilation, observation preprocessing measures, including quality control and bias correction, were developed to reduce the impact of erroneous or poorly represented observations in the assimilation process.

The vast majority of pressure observations report altimeter setting, which is the station pressure adjusted for the elevation of the station above sea level using the U.S. Standard Atmosphere, 1976 (COESA 1976) sea level temperature and lapse rate. For the purposes of this research, the terms “pressure observation” and “altimeter observation” are used interchangeably; however, the actual quantity used for data assimilation is the altimeter setting. All observations reporting station pressure but not altimeter setting are converted to altimeter setting using the reported observation elevation.

Observations beyond a range of acceptable altimeter values (800–1100 hPa) are discarded. As noted in Torn and Hakim (2008), model terrain fields may not resolve the localized orographic features that can surround and influence an observation. An elevation comparison check between the reported observation elevation and the model-derived terrain elevation at that location and surrounding grid points is performed (Torn and Hakim 2008; Ancell et al. 2011). Observations where any of the model terrain elevation values within a 3 × 3 gridpoint box surrounding the observation location differed by more than 200 m from the reported observation elevation are discarded. Out of all potential observations locations, 260 (approximately 15%) are rejected because of this constraint.

To estimate the time-invarying bias in these observations, a high-resolution analysis of the surface pressure field is used for verification. The Rapid Update Cycle (RUC) Surface Assimilation System (RSAS; Miller et al. 2002) yields hourly, 15-km resolution analyses of surface variables. Hourly RSAS analyses of altimeter setting were interpolated to observation locations and the differences between the observation value and the analyses were computed during May–July 2011. This time period in the Pacific Northwest featured relatively tranquil weather conditions with few strong pressure gradients, allowing the RSAS analysis to be used to accurately estimate bias.

Differences between observations and RSAS analyses are evaluated for statistical significance. Where the mean difference is statistically nonzero with 95% confidence, it is treated as a bias and subtracted from future observation values. In total, 28% of the observations are found to have statistically significant differences from the RSAS analyzed values and are subject to bias correction. Furthermore, observations with a mean difference greater than 5 hPa or a standard deviation in the difference field of greater than 2 hPa are discarded as unrepresentative outliers. Out of all observation locations available for this study, 127 locations (approximately 7%) are rejected as outliers. The effect of this bias correction is examined later in this study.

b. Pressure tendency

The bias correction method described above relies on the RSAS analysis as “truth” for estimating observation bias and as such it is vulnerable to any deficiencies in that analysis. This potential source of error would be present regardless of whatever model-based analysis is chosen to represent truth. Furthermore, a long period of observations and analyses for a variety of weather scenarios is required to adequately sample the bias. Therefore, we consider using altimeter tendency as an alternative method of bias remediation. In the presence of a time-invarying or slowly changing bias, altimeter tendency, calculated as the difference between two subsequent altimeter setting observations, should not have a bias.

Little consideration of the potential of pressure or altimeter tendency for data assimilation has been given in the literature. Wheatley and Stensrud (2010) assimilate observed, 1-h pressure tendency for mesoscale convective systems and find that these observations offer some improvement to forecast convective cold pool strength and location, but not as much improvement as assimilating raw altimeter setting; the potential impact of altimeter tendencies of different lengths or in scenarios other than convective cold pools has yet to be evaluated. In preliminary tests using the UW-rtEnKF system (Torn and Hakim 2008), we find relatively high correlations between 3-h altimeter tendency observations and temperature, wind, and geopotential height below 500 hPa (not shown), suggesting that altimeter tendency observations could offer significant contributions to lower-tropospheric structure. To examine their role in mesoscale forecast scenarios, 3-h altimeter tendency is computed for all available altimeter observations and used in specific experiments described below.

c. Ensemble system

The ensemble design is based on the pseudo-operational UW-rtEnKF system (Torn and Hakim 2008; Ancell et al. 2011; Ancell 2012). This system features 64 ensemble members with a 4-km inner nested domain centered over the Pacific Northwest (Fig. 2) and assimilates observations every 3 h. In the UW-rtEnKF system, the 4-km nest is within a 36-km horizontal resolution domain covering much of the northeastern Pacific and western North America. However, for ensemble cycling experiments in this research, the outer domain resolution is increased to 12-km for smoother boundary condition transitions and only covers a limited area of western North America and the eastern Pacific Ocean (Fig. 2).

Fig. 2.

The 12-km domain used for the cycling experiments and the inner 4-km nest used for cycling and no-cycling experiments.

Fig. 2.

The 12-km domain used for the cycling experiments and the inner 4-km nest used for cycling and no-cycling experiments.

For experiments where the ensemble is cycled, the model state is integrated using version 3.4.1 of the Advanced Research core of the Weather Research and Forecasting Model (WRF-ARW; Skamarock et al. 2008). One-way nesting is used to transfer boundary conditions from the outer, 12-km resolution domain to the inner, 4-km nested domain. Lateral boundary conditions for the outer domain are provided by the National Oceanic and Atmospheric Administration (NOAA) Global Forecast System (GFS) forecasts from the most recently available GFS model run at the start of each cycle, simulating an operational environment. To maintain ensemble variance, the lateral boundary conditions are perturbed for each ensemble member using the fixed-covariance perturbation method described in Torn et al. (2006) with perturbations scaled using climatologically derived 6-h GFS forecast errors. Model parameterizations include the Kain–Fritsch convective scheme, WRF single-moment 3-class microphysics scheme (WSM3), Rapid Radiative Transfer Model (RRTM) longwave and shortwave radiation schemes, and Yonsei University (YSU) PBL physics. All ensemble members have identical parameterizations and model formulations, with a Rayleigh damping layer in the upper 7 km of the domain and a damping coefficient of 0.2. The 4-km domain time step is 24 s.

d. Data assimilation system

We use the Data Assimilation Research Testbed (DART) implementation of the ensemble square root adjustment filter, along with attendant utility programs from DART to analyze and process observations and ensemble output (Anderson et al. 2009). Observation error variances for all observation types are derived from a 2005 version of the GFS observation error tables included in the DART system. This includes a 1-hPa2 error variance assigned to all altimeter observations, regardless of their source. Altimeter tendency observations are assigned error variances of 1 hPa2 (3 h)−2.

Horizontal and vertical localization filters are applied to avoid long-range, spurious covariances that are not physically realistic (Anderson 2007a). The localization function used here follows Gaspari and Cohn (1999) with a horizontal localization radius of 800 km and a vertical localization radius of 14 km. To reduce the potential for filter divergence given the limited ensemble size (64 members) and the density and frequency of observations assimilated (3-h cycling), spatially varying and time-evolving inflation is applied during ensemble cycling (following Anderson 2007b).

In addition to altimeter and altimeter tendency observations, a set of standard observations are assimilated in some experiments to examine the additional contribution of pressure observations. These standard observations include METAR, ship, and buoy observations of 2-m temperature and 10-m winds, Aircraft Communications Addressing and Reporting System (ACARS) wind and temperature observations aloft, twice-daily radiosonde observations, and satellite-derived cloud-top wind observations. This observation set reflects the observations used by the UW-rtEnKF system.

e. No-cycling and cycling methods

There is a significant computational expense to cycling this data assimilation system through sufficient forecast cycles to examine the system’s effectiveness with statistical rigor. For some experiments we employ a cost-effective alternative “no cycling” method that uses the archived, 4-km resolution, 3-h forecasts from the UW-rtEnKF system valid at the desired analysis time to populate an ensemble. Data assimilation is performed on this ensemble forecast to produce a new analysis, and the process is repeated at each forecast time. The effect of assimilating particular observation sets can be rapidly evaluated and, since the prior ensemble states for all experiments are identical, the differences in the resulting analyses are exclusively the effect of any differences in observations assimilated or parameters changed. Because the forecast model is not run between different analysis times, there is no cumulative effect of repeated assimilations, and inflation cannot evolve between assimilation times. Both the analyses and forecasts for this method are sensitive to the source of these forecasts (here, the UW-rtEnKF system), and using different ensemble priors may yield different results.

f. Experiments

The time period for the study runs from 0000 UTC 10 November 2012 to 2100 UTC 10 December 2012. This period was marked by a particularly active weather pattern over the Pacific Northwest, including multiple strong frontal passages, shallow convective events over the Columbia Plateau, localized convective banding, and mountain wave activity (not shown).

To compare the impacts of different sets of observations, the available observations are broken into a series of subsets, summarized in Table 2. For verification purposes, 100 surface altimeter, 2-m temperature, and 10-m u- and υ-wind observations are withheld from assimilation at all times. These observations are randomly selected from all available observation sources (including both METAR and additional dense observations), but a spatial filter that limits verification observations to be no closer than 30 km apart is applied to ensure that these observations are well distributed throughout the domain. The altimeter observations used for verification are subjected to the bias correction method described above.

Table 2.

Observation types assimilated in each assimilation experiment. The collection of nonpressure-based observations used in the UW-rtEnKF system is abbreviated as “RT obs” in the table.

Observation types assimilated in each assimilation experiment. The collection of nonpressure-based observations used in the UW-rtEnKF system is abbreviated as “RT obs” in the table.
Observation types assimilated in each assimilation experiment. The collection of nonpressure-based observations used in the UW-rtEnKF system is abbreviated as “RT obs” in the table.

The no-cycling procedure is used to directly compare the differences in analyses resulting from the assimilation of various densities of altimeter observations and combinations of altimeter observations with other observations. Each 3-h ensemble forecast from the UW-rtEnKF during the month-long period of study serves as the prior state for no-cycle data assimilation with different observation sets. Tests are also performed to examine the performance of the altimeter bias correction procedure to improve the subsequent analyses. Evaluation of analysis errors is computed by taking domain-averaged root-mean-squared errors with respect to the withheld verification observations.

Experiments where the full ensemble was cycled every 3 hours using different observation sets are also conducted for the entire month-long period. This allows for rigorous evaluation of 3-h forecast errors, using not only the withheld verification observations but also any observations about to be assimilated at each assimilation step. Finally, an additional cycling study of a convergence zone event on 24 October 2011 is presented to highlight the effect of pressure assimilation on a mesoscale precipitation event.

3. No-cycling experiment

a. No-cycling analyses

Figure 3 shows the month-long 4-km domain-averaged root-mean-square error (RMSE) in the analysis altimeter field after assimilating different densities of altimeter observations. These altimeter observation sets were randomly selected from all available pressure observations in the same manner as the verification observations, with the spatial filter removed beyond 250 observations to allow greater station density. Results show, on average, an 18% reduction in the altimeter error from assimilating one altimeter observation and a 40% reduction for assimilating 100 observations. Error reduction decreases with increasing number of observations beyond 100–150 observations. We note that because these no-cycling experiments use prior ensemble states derived from the UW-rtEnKF cycling assimilation system, the error statistics would be different if another set of prior states were used.

Fig. 3.

Domain-averaged analysis RMSE at 100 unassimilated, bias-corrected, altimeter observation locations with varying numbers of altimeter observations assimilated throughout the domain in a no-cycling experiment. The standard deviations of the errors throughout the month-long period are shown by error bars.

Fig. 3.

Domain-averaged analysis RMSE at 100 unassimilated, bias-corrected, altimeter observation locations with varying numbers of altimeter observations assimilated throughout the domain in a no-cycling experiment. The standard deviations of the errors throughout the month-long period are shown by error bars.

Previous studies (Anderson et al. 2005; Dirren et al. 2007) have shown that given the large correlation length scale inherent in the pressure field, observation densities similar to the METAR network (in this study, 110 well-distributed observations) are sufficient to constrain the synoptic-scale pattern. As such, we hypothesize that the domain-averaged RMSE is more sensitive to synoptic-scale adjustments and less sensitive to the mesoscale adjustments from the additional pressure observations. Furthermore, because adjustments associated with mesoscale phenomena are highly localized, they may not be sampled by the verification network and their signal is easily lost by domain averaging. However, this month-long study provides a large sample to evaluate the domain-averaged RMSE for various fields and to compare between experiments with statistical rigor.

Table 3 summarizes the domain-averaged analysis RMSE over the entire month-long period for the experiments summarized in Table 2. The METAR_only and all_alts_only experiments are performed both with and without (“no_bc”) the altimeter bias correction procedure. There are statistically significant (at the 99% level) reductions in the analysis error when the bias correction is applied. Again, note that the verification pressure observations have been subject to the same bias correction procedure. Later results will show that this improvement persists through the 3-h forecasts, affirming the effectiveness of the bias correction method.

Table 3.

No-cycling domain-averaged analysis RMSE at verification observation sites.

No-cycling domain-averaged analysis RMSE at verification observation sites.
No-cycling domain-averaged analysis RMSE at verification observation sites.

Compared to errors in the control experiment with no observations assimilated, assimilating just the bias-corrected METAR altimeter observations reduces errors in the altimeter field by 44%, while adding the dense altimeter observations further reduces the error by another 5%, consistent with Fig. 3; the additional reduction is statistically significant at the 99% confidence level. In contrast, pressure assimilation alone at either density (the METAR_only and all_alts_only experiments) offers no significant reduction to the domain-averaged analysis errors for the surface temperature and wind fields. Experiments that assimilate surface temperature and wind observations (the rtenkf and all_obs experiments) show a reduction in analysis error over the control experiment, but the different densities of pressure observations do not have any statistically significant impact on the errors. This result is consistent with the findings of Wheatley and Stensrud (2010), who also note that assimilating METAR pressure observations seems to have little to no effect on the domain-averaged analysis error of surface temperature and wind.

Although domain-averaged surface temperature and wind analysis errors are unchanged after assimilating dense pressure observations, this does not mean that no adjustments were made to those fields. To compare the effect of assimilating different densities of altimeter observations on the temperature and wind analyses, at each assimilation step the analysis ensemble mean from the METAR_only experiment is subtracted from the analysis ensemble mean of the all_alts_only experiment. Since the prior ensemble states are identical at each assimilation step, any difference in the analysis is solely the result of assimilating different observation densities. Counts of the number of grid points at each assimilation time where the difference in the two temperature analyses exceed 1 K or the difference in the 10-m wind speed is greater than 2 m s−1 show significant differences between the two analyses, but also that these differences occur over, at most, 3%–5% of the domain. This result is consistent with mesoscale adjustments scattered about the domain.

Spatial maps of these differences reveal several interesting patterns surrounding mesoscale features. Figure 4 shows maps of differences in the analysis altimeter, 2-m temperature, and 10-m wind speed fields between the all_alts_only and METAR_only experiments for three times when a cold front is moving through the domain. In all cases, couplets are present in the difference plots in the vicinity of the front, indicating that the additional pressure observations make coherent adjustments to the position of the front. Furthermore, away from the front the differences in the temperature and wind speed fields are virtually zero, which reinforces the interpretation that these adjustments are highly localized. Difference plots at other times show similar localization of the adjustments surrounding convective bands and terrain-related features (not shown).

Fig. 4.

Analysis differences between the all_alts_only and METAR_only experiments for three frontal events. Differences in (top) pressure, (middle) 2-m temperature, and (bottom) 10-m wind speed.

Fig. 4.

Analysis differences between the all_alts_only and METAR_only experiments for three frontal events. Differences in (top) pressure, (middle) 2-m temperature, and (bottom) 10-m wind speed.

b. 12-h deterministic forecasts

While additional altimeter observations appear to make localized adjustments surrounding mesoscale phenomena, it is unclear whether these adjustments persist for a significant period beyond the initialization time or whether these adjustments lead to improved forecasts. We investigate this issue for the METAR_only and all_alts_only experiments. At each no-cycling update time in these experiments, the ensemble member closest to the ensemble mean is defined by the following metric. A root-mean-square difference between each ensemble member’s state and the ensemble mean is computed, normalized by the variance of each state variable. The ensemble member with the lowest root-mean-square difference is chosen as the member closest to the ensemble mean. This ensemble member is then used to initialize a 12-h forecast using WRF with boundary conditions provided by the most recently available GFS forecast. The boundary conditions at the beginning of the model integration are adjusted to match each analysis, and the boundary time tendencies for the first 3 hours are recomputed to approach the GFS boundary conditions three hours into the forecast. After 3 hours, the boundary conditions in both experiments are identical.

The difference in RMSE between the deterministic forecasts from the all_alts_only and METAR_only experiments is computed from the verification observations. The domain-averaged RMSE for all surface fields in the all_alts_only experiment at 3 hours into the forecast is lower than in the METAR_only experiment for all surface fields, but this reduction is not statistically significant (not shown). Differences in RMSE at longer forecast lead times are smaller and likewise statistically insignificant. A more rigorous examination of the RMSE differences for experiments assimilating both altimeter observation densities is conducted in the fully cycling experiments below.

With such small differences in the forecast RMSE between these two deterministic forecast experiments, the spatial extent and evolution of the differences in these forecasts is examined to determine if the adjustments made during the assimilation step lead to diverging forecasts. Figure 5 illustrates the domain-averaged RMS difference in the surface pressure, 2-m temperature, and 10-m u-wind speed as a function of forecast hour between the all_alts_only and METAR_only forecasts. In the surface pressure field, initially large differences are halved 1 hour into the forecast, most likely as assimilation increments inconsistent with the WRF attractor are dispersed. After the first hour, forecast differences grow through 3 hours, indicating that some of the adjustments made by assimilating dense altimeter observations project onto modes of growth in the model and the forecasts diverge. This forecast divergence is more notable in the 2-m temperature and 10-m u-wind plots where differences grow directly from the initialization time and increase by 130% for temperature and 60% for wind speed through the first 3 hours of the forecast. Thus, even though the additional pressure observations make small, localized adjustments to the wind and temperature fields, they still lead to significant differences in subsequent forecasts. After 3 hours, the differences in all three fields decrease, most likely due to the identical boundary conditions spreading into the small domain; we expect that the influence of the boundary conditions on the domain is strongly dependent on the size of the domain used.

Fig. 5.

Time series of the time mean root-mean-square difference between the all_alts_only experiment and the METAR_only experiments for surface pressure, 2-m temperature, 10-m u-wind component, and 10-m υ-wind component as a function of forecast hour. The forecasts used for these averages were limited to those for which the initial difference in 10-m wind speed exceeded 2 m s−1 at more than 500 grid points.

Fig. 5.

Time series of the time mean root-mean-square difference between the all_alts_only experiment and the METAR_only experiments for surface pressure, 2-m temperature, 10-m u-wind component, and 10-m υ-wind component as a function of forecast hour. The forecasts used for these averages were limited to those for which the initial difference in 10-m wind speed exceeded 2 m s−1 at more than 500 grid points.

The differences between the forecast geopotential height fields of these two experiments are also evaluated. Forecast difference evolution patterns similar to the surface pressure difference evolution are evident in the geopotential height field below 700 hPa (not shown). Above 700 hPa, at no forecast time did the differences in geopotential height between the all_alts_only and METAR_only experiments increase, suggesting that the additional dense altimeter observations have a smaller effect on the domain-averaged mid- and upper-tropospheric structure.

Model time series are compared to 1-min observations to evaluate the performance of the different forecasts. Figure 6 shows a comparison of the 10-m u-wind component at three observation sites along the Columbia River for the frontal passage shown in the left column of Fig. 4. As seen from the difference patterns in Fig. 4, the additional altimeter observations shift the frontal boundary approximately 50 km east. This shift is apparent in bands of higher pressure, lower temperature, and increased wind speed in the difference plots. At all three sites in Fig. 6, the forecast from the all_alts_only experiment improves the timing of the frontal passage by as much as 20–45 min over the experiment where only METAR altimeters were assimilated. This improvement in timing is apparent through 9 hours into the forecast. Similar improvements in the timing of the frontal passages are noted for all six frontal passage cases examined (not shown). These findings indicate that additional dense pressure observations improve the position of frontal boundaries beyond METAR observations alone, and that these changes result in forecast improvements.

Fig. 6.

(top three panels) Time series of u-wind components (m s−1) from 1-min observations (black) and WRF time step traces from an all_alts_only assimilation (red) and a METAR_only assimilation (green) for forecasts beginning at 1800 UTC 17 Nov 2012. The time series have been smoothed by 5-min running means. (bottom) The locations of the three observations are identified. Sites BGDH and BHOR are wind observations from the Bonneville Power Administration network and are not used for assimilation in any experiments.

Fig. 6.

(top three panels) Time series of u-wind components (m s−1) from 1-min observations (black) and WRF time step traces from an all_alts_only assimilation (red) and a METAR_only assimilation (green) for forecasts beginning at 1800 UTC 17 Nov 2012. The time series have been smoothed by 5-min running means. (bottom) The locations of the three observations are identified. Sites BGDH and BHOR are wind observations from the Bonneville Power Administration network and are not used for assimilation in any experiments.

4. Cycling experiments

The 12-h forecasts of the ensemble member closest to the mean from the no-cycling update experiments allow for rapid evaluation of the forecast improvement from assimilating additional dense pressure observations. However, since each no-cycling analysis is nearly independent of all other analyses, the cumulative effect of cycling an ensemble while assimilating additional pressure observations cannot be evaluated. Furthermore, single deterministic forecasts are typically inferior to ensemble mean forecasts. To evaluate the cumulative effect of assimilating dense pressure observations with full ensemble forecasts, 3-h forecast errors are evaluated for the entire month of study using a 3-h cycling ensemble.

A summary of the domain-averaged ensemble-mean 3-h forecast RMSE for all cycling experiments and all surface fields is given in Table 4 with the mean reductions in error between select pairs of cases shown in Fig. 7. Two experiments assimilating different densities of altimeter observations are performed with (METAR_only and all_alts_only) and without (METAR_only_nobc and all_alts_only_nobc) bias correction for the altimeter observations. At both observation densities, the domain-averaged ensemble-mean 3-h altimeter forecast errors improve with 99% confidence when the observations are bias corrected: a 21% reduction in errors for the all_alts_only experiment and a 6% reduction in errors for the METAR_only experiment. These results indicate that the bias correction method is effective, even for METAR observations. Both the all_alts_only and all_obs experiments (which include the additional altimeter observations) show improvement in the 3-h forecast RMSE of the surface altimeter field over their counterpart experiments that just assimilate the METAR altimeter observations (METAR_only and rtenkf experiments, respectively). Comparing these experiments reveals that for surface temperature and wind forecasts, pressure observations alone are no substitute for assimilating surface temperature and wind observations. The 3-h forecast errors for surface temperature and wind components are much improved when observations of those fields are assimilated. When temperature and wind observations are assimilated, assimilating additional dense altimeter observations does improve the forecast: statistically significant (at the 95% level) reductions in error are noted for the altimeter, 2-m temperature, and 10-m υ-wind component (but not the 10-m u-wind component).

Table 4.

Cycled domain-averaged 3-h forecast RMSE at verification observation sites.

Cycled domain-averaged 3-h forecast RMSE at verification observation sites.
Cycled domain-averaged 3-h forecast RMSE at verification observation sites.
Fig. 7.

The month-long mean difference in domain-averaged 3-h forecast RMSE between selected pairs of cases, with the 95% confidence interval shown. The experiments compared are [all_alts_only − METAR_only] (blue), [all_obs − rtenkf], and [rtenkf+tend − rtenkf] (green).

Fig. 7.

The month-long mean difference in domain-averaged 3-h forecast RMSE between selected pairs of cases, with the 95% confidence interval shown. The experiments compared are [all_alts_only − METAR_only] (blue), [all_obs − rtenkf], and [rtenkf+tend − rtenkf] (green).

The 3-h forecasts of surface pressure produced by the ensemble are used to compute 3-h altimeter tendency forecasts and additional experiments are performed where observed 3-h altimeter tendencies are assimilated. Forecast errors for experiments assimilating only 3-h altimeter tendency at all available observation locations (alt_tend_only) and assimilating a standard set of observations with the dense 3-h altimeter tendencies (rtenkf+tend) are also shown in Table 4. Interestingly, the rtenkf+tend experiment shows the lowest RMSE for 3-h forecasts of 2-m temperature and an identical RMSE for 10-m u-wind component to the all_obs experiment. The improvements in the 2-m temperature forecasts are statistically significant (at the 95% level) over the rtenkf case while the improvements in the 10-m u wind are not (Fig. 7). The ability of altimeter tendency to constrain surface temperature errors was noted in Wheatley and Stensrud (2010), where they hypothesize that pressure tendency observations are particularly sensitive to mesoscale temperature patterns. Given that bias correction was needed to better exploit the raw altimeter observations but the altimeter tendency observations required no bias correction, these similar error reductions suggest altimeter tendency observations could be considered as an alternative to raw altimeter observations without the need for bias correction.

Forecast errors for upper-level variables are also evaluated from these experiments. Figure 8 shows vertical profiles of the domain-averaged RMSE verified against ACARS observations. Note that all experiments shown, except the control, assimilate upper-air observations (ACARS, radiosondes, and satellite cloud-track winds); the only differences are in the number or type of pressure observations assimilated. While all experiments show improvement over the control with no data assimilation, reduced errors at low- to midtropospheric levels are noted for wind and temperature in the all_obs experiment over the rtenkf experiment. The rtenkf+tend experiment has an error profile nearly identical to the rtenkf experiment, suggesting that the adjustments made by assimilating 3-h altimeter tendencies do not project strongly on structures aloft. At and above 300 hPa, all three experiments have nearly identical errors, which agrees with the lack of growing differences in the upper-troposphere geopotential height fields noted in the deterministic 12-h forecast comparisons. Similar error profiles are seen when evaluated against radiosonde observations (not shown).

Fig. 8.

Domain-averaged 3-h forecast root-mean-square errors for (left) temperature and (middle) u- and (right) υ-wind components at various levels as computed against ACARS observations. The experiments are the control (red), rtenkf (green), rtenkf+tend (purple), and all_obs (orange).

Fig. 8.

Domain-averaged 3-h forecast root-mean-square errors for (left) temperature and (middle) u- and (right) υ-wind components at various levels as computed against ACARS observations. The experiments are the control (red), rtenkf (green), rtenkf+tend (purple), and all_obs (orange).

Convergence zone case study

To further evaluate the ability of dense pressure observations to improve the analysis and short-term forecast of small-scale, high-impact weather phenomena, a case study was chosen where there was a failure in operational deterministic and ensemble forecasts. A Puget Sound convergence zone (Mass 1981) formed over northern King County, Washington, during the morning of 24 October 2011 and drifted south from 1200 to 1800 UTC. This isolated convective band brought heavy showers and small hail to the Seattle, Washington, area, disrupting the morning commute and contributing to several automobile accidents. Figure 9 shows the composite reflectivity image at 1502 UTC during the peak of the event. This event was poorly forecast by the University of Washington WRF deterministic forecasts, even at 4-km horizontal resolution. The combination of a forecast failure, a very localized, high-impact weather event, and the location of this convergence zone over a region with a high density of available pressure observations (Fig. 1) makes this an attractive case to test the contributions of dense pressure observations to the forecast.

Fig. 9.

Composite reflectivity image from the KATX Weather Surveillance Radar-1988 Doppler (WSR-88D) at 1502 UTC 24 Oct 2011. The convergence zone convective showers are circled by a red line. Many of the remaining radar returns are ground clutter.

Fig. 9.

Composite reflectivity image from the KATX Weather Surveillance Radar-1988 Doppler (WSR-88D) at 1502 UTC 24 Oct 2011. The convergence zone convective showers are circled by a red line. Many of the remaining radar returns are ground clutter.

Given the limited time duration of this event, the number of ensemble members is increased to 80 to better sample local covariances. The inner, 4-km nested domain remains the same as for previous experiments, but the outer domain duplicates the 36-km outer domain used in the UW-rtEnKF system and in Ancell et al. (2011). All other parameters remain the same as described before, except for the microphysics scheme, which is changed to the WSM 5-class scheme for better representation of convective microphysical processes.

The ensemble is initialized with member states from the UW-rtEnKF system archive with linear combinations of existing UW-rtEnKF ensemble member states used to increase the ensemble population to 80 members. The ensemble forecasts start at 0000 UTC 24 October 2011 and run through 0600 UTC without data assimilation to allow for model spinup. Data assimilation for each experiment commences at 0600 UTC and continues with 3-hourly cycling through 1500 UTC. In addition to the 3-h cycling experiments, the impact of increasing the frequency of assimilated altimeter observations is investigated with two hourly assimilation experiments: one using only the dense altimeter observations every hour and another using only the computed 1-h altimeter tendencies every hour. The 3-h ensemble forecasts starting at each assimilation time are also made in these hourly cycling experiments so that the forecast results are directly comparable to other experiments.

Figure 10 shows a comparison of the 3-h ensemble-mean forecasts of simulated composite reflectivity from six experiments (control, rtenkf, alt_tend_only with 3-h cycling, alt_tend_only with 1-h cycling, all_alts_only with 3-h cycling, and all_alts_only with 1-h cycling) valid at 1500 UTC. These simulated reflectivity images may be compared with the observed radar composite reflectivity at 1502 UTC (Fig. 9). Figure 11 shows the number of ensemble members with a local maxima in simulated composite reflectivity of at least 30 dBZ at each grid point to illustrate the ensemble spread in the placement of the precipitation maxima at 1500 UTC.

Fig. 10.

The 3-h forecasts of ensemble-mean simulated composite reflectivity (dBZ) valid at 1500 UTC 24 Oct 2011 for the (top left) control, (top middle) rtenkf, (top right) alt_tend_only with 3-h cycling, (bottom left) alt_tend_only with 1-h cycling, (bottom middle) all_alts_only with 3-h cycling, and (bottom right) all_alts_only with 1-h cycling.

Fig. 10.

The 3-h forecasts of ensemble-mean simulated composite reflectivity (dBZ) valid at 1500 UTC 24 Oct 2011 for the (top left) control, (top middle) rtenkf, (top right) alt_tend_only with 3-h cycling, (bottom left) alt_tend_only with 1-h cycling, (bottom middle) all_alts_only with 3-h cycling, and (bottom right) all_alts_only with 1-h cycling.

Fig. 11.

Number of ensemble members with a local maxima in composite reflectivity of at least 30 dBZ at each grid point at 1500 UTC 24 Oct 2011 for all cases. Local maxima are determined with a 40-km exclusion radius to identify independent maxima. The thick black line follows the 10-dBZ contour in the ensemble-mean composite reflectivity.

Fig. 11.

Number of ensemble members with a local maxima in composite reflectivity of at least 30 dBZ at each grid point at 1500 UTC 24 Oct 2011 for all cases. Local maxima are determined with a 40-km exclusion radius to identify independent maxima. The thick black line follows the 10-dBZ contour in the ensemble-mean composite reflectivity.

The control experiment without data assimilation shows no precipitation band in the vicinity of the convergence zone, consistent with the forecast failure of the UW WRF real-time deterministic model. The rtenkf experiment, which closely mirrors the performance of the UW-rtEnKF system, introduces a band of precipitation over northern Puget Sound, but the band is north of the correct position. The relatively diffuse pattern of the ensemble mean in the rtenkf experiment is indicative of considerable spread in the ensemble members as to the location and intensity of the precipitation. This is further highlighted by the dispersion in the locations and reduced number of reflectivity maxima greater than 30 dBZ in Fig. 11. The alt_tend_only experiment with 3-h cycling appears to produce little improvement over the control. There is a greater southwestern extent of the precipitation and some hint of a more banded structure, but overall it remains similar to the control experiment. Increasing the cycling frequency to hourly and using 1-h altimeter tendencies also does not provide much improvement over the control case, though a small secondary maxima in the mean composite reflectivity occurs over the correct location of the convergence zone precipitation. However, only a small number of ensemble members appear to have a local maxima in reflectivity at that location.

In contrast to these experiments, the all_alts_only experiment with 3-h cycling shows that assimilating dense pressure observations alone yields a more concentrated band of precipitation in the 3-h forecast, and that this band is slightly closer to the observed position. Finally, the 3-h forecast from the all_alts_only experiment with 1-h cycling matches the observed precipitation band well. Not only is the location of the precipitation nearly correct, but the high intensity of the ensemble-mean composite reflectivity reflects relatively small spread in the ensemble members as to the location and intensity of the precipitation band (Fig. 11). It is hypothesized that the more frequent assimilation of the observations predisposed the ensemble to smaller, more transient mesoscale features relevant to the convergence zone and to low ensemble spread, such that many members exhibit a precipitation maximum in nearly the correct location. The success of this 1-h cycling experiment suggests further study on the potential utility of frequent assimilation intervals to improve short-term mesoscale forecasts. The improved depiction of the precipitation band in the all_alts_only cases at either cycling frequency over the alt_tend_only cases also suggests that the raw altimeter observations may be more sensitive to the mesoscale structures surrounding this feature than the altimeter tendency observations.

A more direct comparison of the mesoscale impacts of these two observation types can be seen in the assimilation increments. The mean assimilation increments in the 850-hPa wind fields at the time of the maximum intensity of the convergence zone (1500 UTC) are shown for three cases in Fig. 12. In all of these cases, the location of the reflectivity maxima at 1500 UTC was misplaced to varying degrees as noted previously. The increment, where the rtenkf observations were assimilated (which include upper-air observations of wind and temperature and the METAR altimeter observations), makes larger-scale adjustments to the flow. However, a computation of the change in convergence (not shown) in the increment of the rtenkf experiment reveals no increase in convergence in the vicinity of the convergence zone. In the all_alts_only experiment (with 3-h cycling), the dense altimeter observations directly increase the low-level convergence in the vicinity of the convergence zone by increasing southerly winds to the south of the convergence zone. The 3-h altimeter tendency observations in the alt_tend_only case also increase southerly flow, but not nearly to the same extent as the raw altimeter observations. These assimilation increments provide another example of how dense pressure observations can be sensitive to mesoscale features that coarser surface observation densities are unable to capture.

Fig. 12.

850-hPa wind increments (m s−1) at 1500 UTC 24 Oct 2011 for the rtenkf, alt_tend_only, and all_alts_only cases, all with 3-h cycling intervals. The vector wind increment is shown by arrows. The magnitude of the wind speed increment is illustrated by the size of the arrows and the color shading. Gray areas are locations where the 850-hPa surface is below the terrain.

Fig. 12.

850-hPa wind increments (m s−1) at 1500 UTC 24 Oct 2011 for the rtenkf, alt_tend_only, and all_alts_only cases, all with 3-h cycling intervals. The vector wind increment is shown by arrows. The magnitude of the wind speed increment is illustrated by the size of the arrows and the color shading. Gray areas are locations where the 850-hPa surface is below the terrain.

5. Conclusions

This research evaluates the potential of dense surface pressure observations for describing the atmospheric details necessary for better analyses and forecasts of mesoscale, short-term, high-impact weather events. A large number of pressure observations are available on a regular basis from extant networks in the Pacific Northwest, increasing the number by an order of magnitude over METAR observations. Quality control and bias correction methods were developed to improve the utility of the pressure observations, and these methods were shown to be effective at reducing errors in the analysis altimeter fields and 3-h forecasts when verified against independent observations.

Assimilating all available pressure observations reduced the domainwide error in the resulting surface pressure analysis by an additional 5% as compared to just assimilating METAR observations, but these additional pressure observations did little to change the domainwide analysis error in the surface temperature or wind fields. Closer scrutiny of the differences in analysis increments between the low-density (METAR) and dense pressure network experiments revealed that substantial, small-spatial-scale modifications were made to surface temperature, wind, and pressure fields in the vicinity of mesoscale features with the addition of more pressure observations. For example, assimilating additional pressure observations made coherent adjustments to frontal features, convective bands, and wind patterns around a Puget Sound convergence zone. Forecasts based on these analyses showed that the adjustments introduced by additional pressure observations persisted and grew through the first 3 hours of the forecast. Furthermore, forecasts following the assimilation of additional pressure observations had significantly improved timing (20–45 min) of frontal passages.

Experiments with fully cycling ensembles revealed statistically significant improvements in domain-averaged errors for surface pressure, temperature, and meridional wind forecasts with the addition of dense pressure observations. Improvements were also noted in the temperature and wind errors throughout the low- to midtroposphere. The 3-h altimeter tendency observations successfully contributed to the reduction of forecast errors, most notably in the surface temperature field. However, these improvements did not appear to extend above the surface. Further research may investigate how to best exploit pressure tendency observations, for instance by determining the optimal tendency time frame that provides the most useful information for the scales of interest.

A case study of a specific occurrence of a mesoscale precipitation band showed the potential for dramatic improvement of the short-term forecast by assimilating dense surface pressure observations. In particular, hourly assimilation of dense pressure observations led to a successful forecast of this event, suggesting that frequent assimilation of dense pressure observations may improve short-term forecasts of high-impact mesoscale events.

While bias correction methods were developed to improve observations, no attempt was made here to correct model bias. Several studies (Ancell 2012; Mass et al. 2008) have shown that systematic estimation and removal of model bias in surface wind and temperature fields can lead to more effective data assimilation and improved forecasts. It is unclear at this time whether significant biases are present in the surface pressure field; future studies could determine if a model bias removal technique would further improve the impact of dense pressure observations.

Acknowledgments

We express our gratitude for funding from Microsoft, NOAA CSTAR Award NA10OAR4320148 AM63, and NSF Grant AGS-1041879, which supported this research. Additional appreciation is expressed to Jeff Anderson and the NCAR DART team for conversations and suggestions regarding this research. We thank Mark Albright, Neal Johnson and David Ovens for technical support in obtaining additional pressure observations. We also would like to acknowledge the suggestions of three anonymous reviewers who helped clarify the presentation of this research; their comments were much appreciated.

REFERENCES

REFERENCES
Ancell
,
B.
,
2012
:
Examination of analysis and forecast errors of high-resolution assimilation, bias removal, and digital filter initialization with an ensemble Kalman filter
.
Mon. Wea. Rev.
,
140
,
3992
4004
, doi:.
Ancell
,
B.
,
C.
Mass
, and
G.
Hakim
,
2011
:
Evaluation of surface analyses and forecasts with a multiscale ensemble Kalman filter in regions of complex terrain
.
Mon. Wea. Rev.
,
139
,
2008
2024
, doi:.
Anderson
,
J. L.
,
2007a
:
An adaptive covariance inflation error correction algorithm for ensemble filters
.
Tellus
,
59A
,
210
224
, doi:.
Anderson
,
J. L.
,
2007b
:
Exploring the need for localization in ensemble data assimilation using a hierarchical ensemble filter
.
Physica D
,
230
,
99
111
, doi:.
Anderson
,
J. L.
,
B.
Wyman
,
S.
Zhang
, and
T.
Hoar
,
2005
:
Assimilation of surface pressure observations using an ensemble filter in an idealized global atmospheric prediction system
.
J. Atmos. Sci.
,
62
,
2925
2938
, doi:.
Anderson
,
J. L.
,
T.
Hoar
,
K.
Raeder
,
H.
Liu
,
N.
Collins
,
R.
Torn
, and
A.
Avellano
,
2009
:
The Data Assimilation Research Testbed: A community facility
.
Bull. Amer. Meteor. Soc.
,
90
,
1283
1296
, doi:.
Benjamin
,
S.
,
B.
Schwartz
, and
R.
Cole
,
1999
:
Accuracy of ACARS wind and temperature observations determined by collocation
.
Wea. Forecasting
,
14
,
1032
1038
, doi:.
COESA
,
1976
: U.S. Standard Atmosphere, 1976. NOAA, 227 pp.
Colle
,
B.
,
C.
Mass
, and
D.
Ovens
,
2001
:
Evolution of the timing and strength of MM5 and Eta surface trough passages over the eastern Pacific
.
Wea. Forecasting
,
16
,
553
572
, doi:.
Compo
,
G. P.
, and Coauthors
,
2011
:
The twentieth century reanalysis project
.
Quart. J. Roy. Meteor. Soc.
,
137
,
1
28
, doi:.
Dirren
,
S.
,
R.
Torn
, and
G.
Hakim
,
2007
:
A data assimilation case study using a limited-area ensemble filter
.
Mon. Wea. Rev.
,
135
,
1455
1473
, doi:.
Eckel
,
F.
, and
C.
Mass
,
2005
:
Aspects of effective mesoscale, short-range ensemble forecasting
.
Wea. Forecasting
,
20
,
328
350
, doi:.
Engerer
,
N.
,
D.
Stensrud
, and
M.
Coniglio
,
2008
:
Surface characteristics of observed cold pools
.
Mon. Wea. Rev.
,
136
,
4839
4849
, doi:.
Gaspari
,
G.
, and
S. E.
Cohn
,
1999
:
Construction of correlation functions in two and three dimensions
.
Quart. J. Roy. Meteor. Soc.
,
125
,
723
757
, doi:.
Goff
,
R. C.
,
1976
:
Vertical structure of thunderstorm outflows
.
Mon. Wea. Rev.
,
104
,
1429
1440
, doi:.
Hanley
,
K.
,
D.
Kirshbaum
,
N.
Roberts
, and
G.
Leoncini
,
2013
:
Sensitivities of a squall line over central Europe in a convective-scale ensemble
.
Mon. Wea. Rev.
,
141
,
112
133
, doi:.
Janjic
,
T.
, and
S.
Cohn
,
2006
:
Treatment of observation error due to unresolved scales in atmospheric data assimilation
.
Mon. Wea. Rev.
,
134
,
2900
2915
, doi:.
Mass
,
C. F.
,
1981
:
Topographically forced convergence in western Washington state
.
Mon. Wea. Rev.
,
109
,
1335
1347
, doi:.
Mass
,
C. F.
, and
G.
Ferber
,
1990
:
Surface pressure perturbations produced by an isolated mesoscale topographic barrier. Part I: General characteristics and dynamics
.
Mon. Wea. Rev.
,
118
,
2579
2596
, doi:.
Mass
,
C. F.
,
J.
Baars
,
G.
Wedam
,
E.
Grimit
, and
R.
Steed
,
2008
:
Removal of systematic model bias on a model grid
.
Wea. Forecasting
,
23
,
438
459
, doi:.
Melhauser
,
C.
, and
F.
Zhang
,
2012
:
Practical and intrinsic predictability of severe and convective weather at the mesoscales
.
J. Atmos. Sci.
,
69
,
3350
3371
, doi:.
Miller
,
P.
,
M.
Barth
, and
L.
Morone
, cited
2002
: RSAS technical procedures bulletin. [Available online at http://msas.noaa.gov/rsas_tpb.html.]
Roebber
,
P.
,
D.
Schultz
,
B.
Colle
, and
D.
Stensrud
,
2004
:
Toward improved prediction: High-resolution and ensemble modeling systems in operations
.
Wea. Forecasting
,
19
,
936
949
, doi:.
Skamarock
,
W. C.
, and Coauthors
,
2008
: A description of the Advanced Research WRF version 3. Tech. Rep. NCAR/TN-475+STR, National Center for Atmospheric Research, Boulder, CO, 125 pp. [Available online at http://www.mmm.ucar.edu/wrf/users/docs/arw_v3.pdf.]
Torn
,
R.
, and
G.
Hakim
,
2008
:
Performance characteristics of a pseudo-operational ensemble Kalman filter
.
Mon. Wea. Rev.
,
136
,
3947
3963
, doi:.
Torn
,
R.
,
G.
Hakim
, and
C.
Snyder
,
2006
:
Boundary conditions for limited-area ensemble Kalman filters
.
Mon. Wea. Rev.
,
134
,
2490
2502
, doi:.
Wheatley
,
D.
, and
D.
Stensrud
,
2010
:
The impact of assimilating surface pressure observations on severe weather events in a WRF mesoscale system
.
Mon. Wea. Rev.
,
138
,
1673
1694
, doi:.
Whitaker
,
J. S.
,
G. P.
Compo
,
X.
Wei
, and
T. M.
Hamill
,
2004
:
Reanalysis without radiosondes using ensemble data assimilation
.
Mon. Wea. Rev.
,
132
,
1190
1200
, doi:.