Influence of Surface Observations in Mesoscale Data Assimilation Using

APRIL 2014 H A A N D S N Y D E R 1489

Inﬂuence of Surface Observations in Mesoscale Data Assimilation Using an Ensemble Kalman Filter

SO-YOUNG HA AND CHRIS SNYDER National Center for Atmospheric Research, Boulder, Colorado

(Manuscript received 6 April 2013, in ﬁnal form 22 November 2013)

ABSTRACT

The assimilation of surface observations using an ensemble Kalman filter (EnKF) approach was success- fully performed in the Advanced Research version of the Weather Research and Forecasting Model (WRF) coupled with the Data Assimilation Research Testbed (DART) system. The mesoscale cycling experiment for the continuous ensemble data assimilation was verified against independent surface mesonet observations and demonstrated the positive impact on short-range forecasts over the contiguous U.S. (CONUS) domain throughout the month-long period of June 2008. The EnKF assimilation of surface observations was found useful for systematically improving the simulation of the depth and the structure of the planetary boundary layer (PBL) and the reduction of surface bias errors. These benefits were extended above PBL and resulted in a better precipitation forecast for up to 12 h. With the careful specification of observation errors, not only the reliability of the ensemble system but also the quality of the following forecast was improved, especially in moisture. In this retrospective case study of a squall line, assimilation of surface observations produced analysis increments consistent with the structure and dynamics of the boundary layer. As a result, it enhanced the horizontal gradient of temperature and moisture across the frontal system to provide a favorable condition for the convective initiation and the following heavy rainfall prediction in the Oklahoma Panhandle. Even with the assimilation of upper-level observations, the analysis without the assimilation of surface observations simulated a surface cold front that was much weaker and slower than observed.

1. Introduction boundary layer (PBL); the background error covariances are important because, in statistical data assim- Compared to other routine in situ observations, sur- ilation schemes, they specify how surface observations will face observations (2-m temperature and moisture, 10-m influence the analysis. Second, forecast models typically winds) have relatively good spatial and temporal reso- have limited skill in the PBL owing to significant errors lution, but their use in atmospheric data assimilation has in the parameterizations of vertical mixing, surface proven substantially more difficult. Indeed, many state- fluxes, and land surface processes. This paper concen- of-the-art systems do not assimilate surface observations trates on the first issue and presents results showing the directly; at the European Centre for Medium-Range effectiveness of the ensemble Kalman filter (EnKF) for Weather Forecasts (ECMWF), for example, surface the assimilation of surface observations, which we at- observations are used in a two-dimensional surface tribute to the EnKF’s use of background error co- analysis that forces the land surface model but are not variances computed from an ensemble of short-range used in the variational assimilation scheme for the at- forecasts and that are therefore consistent with the local mosphere (de Rosnay et al. 2013). The difficulty of as- structure and dynamics of the PBL. similating surface observations arises from at least two Specifying the background error covariances in the issues. First, the assimilation scheme must specify the PBL is challenging because of complexity and variability background error covariances, especially in the vertical in the structure of the PBL and in the relation of surface and in relation to the structure of the planetary variables to the rest of the atmosphere. Physical intuition suggests that observed surface variables may Corresponding author address: Dr. So-Young Ha, NCAR/ often have strong correlations with variables within the MMM, 3450 Mitchelle Ln., P. O. Box 3000, Boulder, CO 80301. PBL and weak correlations with flow outside the PBL, E-mail: [email protected] depending on the time of day and the boundary layer

DOI: 10.1175/MWR-D-13-00108.1

Ó 2014 American Meteorological Society 1490 MONTHLY WEATHER REVIEW VOLUME 142 regime. Nudging schemes for assimilation have incor- special treatment of surface observations, either through porated this intuition by requiring the nudging coefficients bias correction of the prior forecasts or by reducing their for surface observations to decrease to zero above the PBL observation-error variance, resulted in better fits of the top (Stauffer et al. 1991). Such modifications to the co- analysis and 6-h prior forecasts to unassimilated surface efficients are, however, heuristic, in that they are based on observations. These benefits did not, however, extend the developer’s assumptions for the relation of surface above the surface. variables to other variables and other levels. In the present study, we continue the exploration of Three-dimensional variational data assimilation assimilation of real surface observations in the EnKF (3DVAR) schemes typically use stationary covariances, approach using the Advanced Research version of the calculated from statistics of forecast differences, which Weather Research and Forecasting Model (WRF) can capture basic aspects of PBL dynamics such as the coupled with the Data Assimilation Research Testbed correlation between vertical vorticity and horizontal (DART) system (WRF–DART) and examine the in- divergence at synoptic scales (Derber and Bouttier fluence of surface observations on mesoscale analyses in 1999). Nevertheless, because they use stationary back- an optimized environment over the contiguous U.S. ground covariances, 3DVAR schemes generally do not (CONUS) domain. Our goal is to investigate the effec- account for any dependence of analysis increments on tiveness of the EnKF in the assimilation of surface ob- the instantaneous structure, the diurnal cycle, or the servations and demonstrate improvements of the regime of the PBL (e.g., Benjamin et al. 2004). following forecasts in the boundary layer and farther More sophisticated statistical assimilation approaches, aloft, as well as at the surface. For this purpose, we such as four-dimensional variational data assimilation perform experiments with and without assimilation of (4DVAR) and the EnKF schemes, extract dynamical surface observations, and make forecasts from the mean information directly from the forecast model and pro- EnKF analysis for 72 h. Through a mesoscale convective duce analysis increments that depend on the local state case study, we examine how the assimilation of surface of the PBL and its recent history. In the EnKF, the data can provide a favorable condition for the initiation relative weighting in the analysis between the back- and the development of severe convection associated ground and observations, and the influence of observa- with a squall line and to what extent the analysis in- tions away from the station locations, are determined by crement affects the structure and evolution of the anisotropic and inhomogeneous background covariances. boundary layer. These covariances, in turn, are estimated from an en- Surface pressure observations reflect the distribution semble of short-range forecasts using the full, nonlinear of mass vertically integrated through the depth of the model, and the covariances will, therefore, change as the entire atmosphere, and their use in the EnKF has been PBL evolves. explored by Dirren et al. (2007) and Compo et al. (2006), Several previous studies have specifically examined for example. It will not be a main focus here, since they the assimilation of surface observations with an EnKF are not particularly sensitive to or informative of con- method. Hacker and Snyder (2005) considered simu- ditions in the PBL and their assimilation is subject to lated surface observations in a single-column model. neither of the two main issues noted in the first part of Their results confirmed the physical intuition that sur- this session. face observations are strongly correlated with, and The data assimilation system implemented here therefore when assimilated informative of, conditions largely ignores forecast model errors, despite their po- throughout the well-mixed PBL or even into the pre- tential importance in the PBL and therefore in the as- vious day’s residual layer. Fujita et al. (2007) assimilated similation of surface observations. The system includes actual surface observations in the EnKF. In addition to only a spatially and temporally varying, adaptive in- examining the role of forecast model error, they dem- flation (Anderson 2009), which compensates for un- onstrated improvements to the analysis of important derestimation of the analysis variance by the EnKF as mesoscale features for two severe weather cases as well well as underdispersion in the prior ensemble forecasts as improvements lasting 6–12 h in the subsequent fore- owing to unrepresented model error. Other possible casts. Meng and Zhang (2008) also briefly examined approaches aimed specifically at representing model assimilation of surface observations in a case study of uncertainties within the EnKF are employing multiple a mesoscale convective vortex but found little influence suites of physical parameterizations (Fujita et al. 2007; on the short-range forecast. Results from a longer, Meng and Zhang 2008; Hacker et al. 2011) and including 5-week period of assimilation appeared in Ancell et al. an additive-noise or ‘‘backscatter’’ scheme within the (2011). They did not examine the influence of surface model (Hacker et al. 2011; Berner et al. 2011). We will observations on the analysis directly, but did show that examine these approaches in a subsequent study. APRIL 2014 H A A N D S N Y D E R 1491

FIG. 1. Model domains with aviation routine weather report (METAR) stations reported at 0000 UTC 9 Jun 2008. Stations that are far away from the model terrain more than 100 m but less than 300 m marked as an open circle (o), and the ones with more than 300-m height difference marked as a crisscross (3). Stations shown as ﬁlled dots have the height discrepancy less than 100 m.

The outline of the paper is as follows. The experi- conditions for the next 3-h forecasts. Both the analysis mental design is outlined in section 2, including our and forecast steps used a 50-member ensemble in all configuration of the Advanced Research version of experiments. WRF (Skamarock et al. 2005) and details of the EnKF as a. The model configuration implemented in the DART (Anderson et al. 2009). Section 3 covers the observations used in the assimila- The forecast step employed WRF version 3.2 on the tion and the observation processing. In particular, we domain depicted in Fig. 1, with a horizontal grid of 123 3 describe in detail how the error variances for surface 99 points at 45-km resolution for the outer domain (D1) observations were assigned because of their importance and of 163 3 106 points at 15-km for the inner domain to our results. Results from the month-long assimilation (D2). The 3-h ensemble forecasts in the two domains ran experiments with and without surface observations ap- with two-way nesting. We used 41 vertical levels to pear in section 4, followed by a case study of how as- discretize the atmosphere between the surface and similation of surface observations affects the forecast of 50 hPa, with about 10 layers in the lowest 2 km. On both severe convection that occurred during the assimilation domains, the model physics configuration included the cycling period. A summary and conclusions are pro- Yonsei (YSU) PBL scheme (Hong et al. 2006), the vided in section 6. Kain–Fritsch (Kain 2004) cumulus parameterization, the WRF single-moment 5-class microphysics scheme (WSM5; Hong et al. 2004), the Noah land surface model 2. Experiments (LSM; Chen and Dudhia 2001), Dudhia shortwave Our data assimilation experiments were performed in (Dudhia 1989), and Rapid Radiative Transfer Model a full cycling mode for a 1-month period of June 2008, (RRTM) longwave (Mlawer et al. 1997) radiation starting from 0000 UTC 1 June, at a 3-h interval. That is, schemes. In the current WRF version, each PBL scheme is observations were assimilated every 3 h using the 3-h tied to a particular surface-layer scheme, and the YSU forecast ensemble from the previous cycle to compute PBL uses the fifth-generation Pennsylvania State Univer- the prior (or background) ensemble covariance, and sity–National Center for Atmospheric Research (PSU– producing the analysis ensemble as an ensemble of initial NCAR) Mesoscale Model (MM5) surface-layer similarity 1492 MONTHLY WEATHER REVIEW VOLUME 142

(Zhang and Anthes 1982). Here, all the ensemble each member, which were diagnosed in the PBL and members use the same model configuration. surface-layer schemes as part of the normal WRF in- The ensemble-mean lateral boundary conditions tegration. Within the EnKF, these quantities were first (LBCs) on D1 were given by the 18318 National used, if necessary, to calculate observed variables (in the Centers for Environmental Prediction (NCEP) Global case of Td2 and altimeter setting), and then interpolated Forecast System (GFS) 6-h forecast valid at the appro- horizontally to the observation location to provide priate times, consistent with what would be available a prior (or background) estimate of the observations were the system running in real time. We accounted for from each member. These forward operators are thus the uncertainty in the LBCs, following Torn et al. (2006), consistent with the PBL and surface-layer schemes in by specifying each member’s LBCs as the sum of the the forecast model. ensemble-mean LBCs and random, spatially correlated Fujita et al. (2007) followed a similar approach, but perturbations. The perturbations were generated as diagnosed T2, Td2, U10, and V10 within the EnKF using draws from a Gaussian random vector, using the back- surface-layer similarity relations. Note also that these ground covariance with random_cv 5 3 in the WRF data forward operators ignored differences between model assimilation system (Huang et al. 2009), and variances, terrain height and station elevations, unlike other stud- and horizontal and vertical length scales tuned as in ies where T2 was corrected based on the local lapse rate Torn and Hakim (2008). (Benjamin et al. 2004). At the beginning of the analysis cycle, initial condi- Besides the check on the station elevation to ensure tions are required for each member. On D1, these were that it is close enough to the model’s orography (which generated in the same way as the ensemble of LBCs, but will be described in section 3), the ensemble data as- centered on the GFS analysis from 0000 UTC 1 June similation system also discards observations based on 2008. On the inner domain, there were no initial per- the innovation magnitude (a so-called ‘‘outlier check’’). turbations; these spin up over 1 or 2 days through the An observation is rejected if the absolute value of its interaction with the outer domain. The initial soil states difference from the mean prior observation is greater for each member are simply the GFS analysis, without than 3 times the total spread, which is given by the additional perturbations. The details of the ensemble square root of the sum of the prior ensemble variance initial conditions at 0000 UTC 1 June have little effect and the observation-error variance. on our results, as the assimilation of observations and To counteract the detrimental effects of sampling er- random perturbations in the LBCs cause the ensemble ror, we utilized a distance-dependent covariance local- to become largely independent of initial conditions ization (Houtekamer and Mitchell 2001; Hamill et al. within a few days. 2001). For the algorithm used here, in which observations are processed serially within the EnKF, covariance b. The EnKF configuration localization amounts to multiplying the observation- For assimilation, we chose the EnKF, implemented in state covariance estimated from the ensemble by DART as an ensemble adjustment Kalman filter a function of the distance between the observation and (Anderson 2001), one variant in the class of deterministic state variables, which decreases smoothly to zero at ‘‘square root’’ EnKFs (Tippett et al. 2003). The analysis a finite distance according to the compactly supported variables were the three components of velocity, po- correlation function of Gaspari and Cohn (1999). Based tential temperature, column mass of dry air, geo- on limited sensitivity tests for a two-week period (not potential and mixing ratios for water vapor and shown), we took that distance to be 600 km horizontally microphysical species, as well as various surface di- and 8 km vertically. agnostic variables available in WRF (as specified in the Even using the covariance localization, however, following paragraph). Soil states were not updated by other factors contribute to the underestimation of the assimilation system and therefore evolve over the analysis uncertainties in our system, including de- month of cycling according to the atmospheric forcing ficiencies in the forecast model (which we ignored in from each member. running ensemble forecasts), effects of nonlinearity and Surface observations assimilated were 2-m tempera- non-Gaussianity, and remaining sampling errors. The ture and dewpoint (T2 and Td2, respectively), the 10-m influence of these will vary spatially and temporally horizontal wind components (U10 and V10), and the depending on the details of flow and of the observation surface altimeter. The forward operators for these ob- network. To account for the combined effects of these servations were computed as follows. At the end of the factors, we employed spatially and temporally varying, forecast step, we output gridded fields of 2-m water va- adaptive inflation in the model state space (Anderson por mixing ratio, T2, U10, V10, and surface pressure for 2009). APRIL 2014 H A A N D S N Y D E R 1493

The WRF-DART system updates multiple, nested domains in a straightforward fashion. First, the forward operators are computed using the fields from the finest grid that contains the observation. Then, the analysis proceeds on all domains simultaneously, with each observation updating all grid points within the localization radius, regardless of the domain to which they belong. Thus, observations within the inner 15-km domain can also affect the analysis on the outer domain, and vice versa. Since surface observations are typically reported at least every hour, the frequency of assimilation is a po- tentially important question. Our choice of a 3-hourly FIG. 2. A time series of rms innovations (thick solid lines) and assimilation cycle was based on initial tests over the first total spread (thin dashed lines) in u wind at 10 m is shown for the two cases where the NCEP observation-error variance (gray) and 10-day period of cycles that compared the performance the newly adjusted errors (black) are used for METAR observa- of 1-, 3-, and 6-hourly cycling frequencies and revealed tions in domain 1. Both posterior (lower points) and prior obser- that innovations for surface observations were not very vations (higher points) were put together at each cycle making sensitive to the frequency of assimilation (not shown). a sawtooth pattern. Choosing 3-hourly cycles was a compromise between increasing the number of surface observations and the Although more sophisticated approaches are avail- indications from some previous studies that more fre- able (Desroziers and Ivanov 2001; Desroziers et al. quent assimilation cycles lead to spurious noise in the 2005), we tuned the observation-error variances using background forecast (Benjamin et al. 2004). the fact that, when observation and background errors are unbiased and uncorrelated, 3. Observations tr(hddTi) 5 tr(P) 1 tr(R), (1) In this study, we collected observations from the Meteorological Assimilation Data Ingest System where d 5 y 2 hH(x)i is an innovation, y is the obser- (MADIS) of the National Oceanic and Atmospheric vation vector, x is the state vector, H is the forward Administration (NOAA) for different observation operator, P 5 cov[H(x)] is the background covariance types. Radiosonde soundings, Aircraft Communications matrix for the observed variables, R is the observation- and Reporting System (ACARS), marine, and METAR error covariance matrix, and angle brackets denote an data were used for assimilation. We also collected sur- expectation over both the observation-error and the face observations from various mesonet archived in background distribution. This equation is routinely used MADIS. Because they have slightly lower quality than in both ensemble forecasting and ensemble data assim- the METAR observations but show better coverage ilation, with the expectation on the rhs approximated by over the CONUS domain, we withheld the mesonet a time average and that on the lhs by averaging over the observations from the assimilation and used them as ensemble and over time (e.g., Mitchell and Houtekamer independent observations to evaluate the quality of the 2000; Wang and Bishop 2003; Aksoy et al. 2009). Dee surface analyses and forecasts. The mesonet observa- (1995) discusses in detail the use of Eq. (1) for co- tions were otherwise treated identically to the METAR variance estimation. observations: the same forward operators, the same Figure 2 depicts, for the first 10 days of June 2008, the observation-error variances, and the same quality con- square root of the rhs of Eq. (1) with P estimated from trol were applied. The ACARS observations were as- the prior ensemble, which we term the total spread, and similated as a fewer set of superobservations, given by the rms innovations [i.e., the square root of the lhs of Eq. averages over boxes 45 km on a side and 50 hPa deep. (1)]. With the NCEP observation errors, the rms in- Specification of observation-error variances is an novations between the 3-h ensemble mean forecast and important element of the successful data assimilation METAR observations of U10 were much smaller than 2 of real observations. For all observations except sur- the total spread (gray dashed line around 3.6 m s 1). face data, we employed observation-error variances When we reduced the observation-error variance from 2 2 taken from the gridpoint statistical interpolation (GSI) (3.5 m s 1)2 to (1.5 m s 1)2, the forecast fit to the obser- 2 system (Kleist et al. 2009). For surface observations, we vations improved by ;0.6 m s 1 on average, making the have conducted additional tuning based on observation- innovation comparable to total spread. The y wind at space diagnostics. 10 m yielded very similar results when the same observation 1494 MONTHLY WEATHER REVIEW VOLUME 142

TABLE 1. The rms errors and total spread of the 3-h forecast in terms of 2-m dewpoint temperature before (‘‘Old’’) and after (‘‘New’’) adjusting dewpoint observation-error variance. The values are computed by averaging over the METAR stations in domain 1 for 1–10 Jun 2008.

Old New rmse 2.2 2.0 Spread 3.6 2.0

humidity are 610% and 10%, respectively, or 65% and 20%, while the assumed temperature uncertainty re- mains as 61 K. In the former case, most dewpoint observations are assigned error standard deviations between 2 and 3 K but the maximum error can be as large as 14 K when relative humidity becomes small (in the transparent bar graph). In the latter case, the observation-error standard deviations generally lie between 1.4 and 2.2 K and are no larger than 4 K. We FIG. 3. Histogram of the observation errors for METAR dew- summarize in Table 1 how these choices for the error point temperature sampled from domain 1 at 0000 UTC 1 Jun 2008. variances affect the performance of the forecast over The original error distribution estimated from Lin and Hubbard the 10-day test period. Averaged over roughly 1000 (2004) is shown as transparent bar graph while the newly adjusted METAR stations in the CONUS domain, the larger error distribution in bar graph filled in black. error variances result in total spread that is substantially larger than the rms innovations for 2-m dewpoint. errors were applied. For 2-m temperature, rms innova- The smaller error variances reduce the rms innovations tions improved more than 0.5 K with the observation- by 10% and give a total spread that matches the rms error variance reduced from (2.5 K)2, as specified in innovations. the NCEP error table, to (1.5 K)2 (not shown). Based on Based on this result, we applied the new error speci- these experiments, those reduced observation-error var- fication regardless of vertical levels or instrument types iances were applied for all the subsequent experiments. in all experiments. This was done partly because of Moisture observations were assimilated as dewpoint simplicity and partly because our main focus was on temperature regardless of the observation type, follow- surface data assimilation, but Fig. 4 indicates that the ing Fujita et al. (2007). As the MADIS provides mois- mean bias and the rms errors of the 3-h forecast (i.e., ture observations as dewpoint, we could avoid an prior) are greatly reduced in the entire atmosphere by additional conversion error in the observation operator reducing dewpoint observation errors at all levels. The and take advantage of the Gaussian characteristics of improved moisture assimilation, however, does not no- the observation and background errors for dewpoint. ticeably affect the horizontal wind or temperature in the The dewpoint observation-error variance was speci- background forecast (not shown). fied following Lin and Hubbard (2004). They estimated Surface observations are subject to considerable errors dewpoint errors based on the sensitivity of dewpoint to of representativeness, especially in regions of complex small perturbations in temperature and relative hu- terrain. This means that the variance of observation midity, with the result that the error variance depends error, which includes both the instrumental error and on the relative humidity and temperature and increases errors of representativeness, is uncertain for some sur- as the relative humidity decreases. There are three pa- face observations. Moreover, innovation statistics for rameters to be specified: assumed uncertainties for surface observations often exhibit significant bias, which temperature and relative humidity, and a minimum may arise either from deficiencies in the forecast model value allowed for the relative humidity, since the sen- or peculiarities of where the observations are located sitivity of dewpoint to relative humidity becomes ex- and that bias is not accounted for in the formulation of ponentially large as relative humidity approaches the assimilation system. We addressed these issues in zero. Figure 3 shows histograms of the estimated dew- only the simplest way, by assimilating just those obser- point observation-error variance when the assumed vations for which the station elevation was within a cer- relative humidity uncertainty and the minimum relative tain distance from the model topography. Given that our APRIL 2014 H A A N D S N Y D E R 1495

As our observation-space diagnostics showed no consistent advantage for one threshold over the other, we chose the 300-m threshold in our assimilation to beneﬁt from the dense observing network at the surface.

4. Results

Following the careful specification of observation errors and the filter design for mesoscale cycling, we ran experiments with and without surface observations for the month of June 2008. The experiments are named SFCDA and NO_SFCDA, respectively, and both use all other conventional observations discussed in section 3. Figure 5 shows the diagnostics for the two runs in terms of rms innovations, ensemble spread, and model- minus-observation bias errors verified against independent mesonet observations in domain 1. With the assimilation of surface observations (SFCDA), both the prior and posterior ensemble mean fit the observations better throughout the period. A similar advantage of

FIG. 4. Vertical profile of the 3-h forecast rms innovations (solid SFCDA is also clear in the METAR verification where line), total spread (dotted line), and mean forecast-minus-observation both experiments produced better fits than in the mes- bias errors (dashed line) verified against sounding dewpoint ob- onet verification by ;10% (not shown). servations in domain 1. The experiments before and after ad- Among all surface fields, the largest improvement justing dewpoint observation errors are shown as gray and black lines, respectively. from the assimilation of surface observations is made in dewpoint temperature; the rms difference of the prior forecast from the observations decreases by 0.4 K when horizontal resolution (for either 45- or 15-km grid) does averaged over all the common stations between the two not capture the scales associated with narrow valleys, we experiments for the entire cycling period (Fig. 5c). The tested two different threshold values of 100 and 300 m in rms innovations and the bias of the 3-h ensemble mean rejecting surface observations. Compared to a more forecast, and ensemble spread, averaged over the cycles, generous threshold of 300 m, the use of 100-m threshold are summarized in Table 3. discarded more surface observations (up to 14%), pre- A large portion of rms innovations in the surface dominantly over the western United States; over the thermodynamic fields comes from bias. In comparison to Great Plains, however, most observations were retained mesonet observations, the 3-h forecast from NO_SFCDA (Fig. 1). The verification of 3-h forecasts against surface presents cold and wet bias during the daytime while it is METAR observations for the two different thresholds is slightly warmer and much drier in stable boundary summarized in Table 2. For the 10-day test period, rms conditions. The assimilation of surface observations ef- innovations were not significantly reduced by rejecting fectively decreases the magnitude and diurnal variability more stations with the 100-m threshold, while the cold of the mean bias near the surface. The 3-h forecast from bias was decreased in 2-m temperature by about 20%. the SFCDA reduces the cold bias by up to 0.4 K in the

TABLE 2. The rms innovations and bias errors of the prior (i.e., 3-h forecast) verified against METAR observations computed over all the stations in domain 1 for the first 10 days of June 2008. The errors are compared between two experiments with different threshold values—dh100 for 100 m and dh300 for 300 m—as the maximum height difference between the stations and the model height to be assimilated. In the first column, ‘‘Nobs’’ stands for the total number of stations assimilated.

2 2 U10 (m s 1) V10 (m s 1) T2 (K) Td2 (K) dh100 dh300 dh100 dh300 dh100 dh300 dh100 dh300 rmse 1.93 1.98 1.95 2.00 2.00 2.01 1.94 1.98 Bias 0.48 0.54 0.38 0.30 20.26 20.32 0.15 0.14 Nobs 936 1082 932 1079 961 1118 973 1129 1496 MONTHLY WEATHER REVIEW VOLUME 142

evident that the assimilation of surface observations effectively alleviates the surface thermodynamic bias errors (with the maximum time-mean decrease of ;0.5 K in 2-m dewpoint). The spread of the 3-h ensemble forecasts is larger in NO_SFCDA than in SFCDA, but compared to the specified observation-error standard deviation (as described in the previous section), the spread difference between the two experiments is not considerable. That is, the observation uncertainty is dominant over the difference in the forecast uncertainties between the two runs. Diagnostics in the nested domain illustrate similar advantages in SFCDA to the ones in domain 1 in all surface fields. Moisture assimilation makes the biggest improvement again—about 0.8 K in rms innovation and 0.7 K in bias errors on average—as shown in Fig. 6. Improvements from the assimilation of surface observations are generally larger in domain 2 partly because it mostly covers smooth terrain and partly because innovations are computed using the atmospheric states from the finest grid as discussed in section 2. Another interesting aspect in the ensemble data assimilation system is to compare the number of observations actually used in each experiment. In general, our experiments use 80%–90% of surface observations depending on the observation type, both in the assimilation (e.g., for METAR) and in the verification (e.g., for mesonet observations). Without assimilation of METAR observations, the system rejects more mesonet observations (up to 10%) especially when the model has a strong bias with respect to the surface observations due to the failure of outlier check described in section 2b. The fact that SFCDA retains more observations is an another sign that the SFCDA analyses and forecasts are FIG. 5. A time series of mesonet verification for SFCDA (red) improved relative to those from NO_SFCDA. and NO_SFCDA (blue) in terms of (a) u wind at 10 m, (b) tem- To examine the vertical impact of the assimilation of perature, and (c) dewpoint at 2 m, computed over the common surface data, we plot vertical profiles of the rms forecast stations between the two experiments in domain 1. Root-mean- square innovations (rmsi) are shown in thick solid lines and en- differences from radiosonde observations for both ex- semble spread in thin solid lines. Bias errors (dashed lines) are periments in the nested domain in Fig. 7. With the as- shown in black and gray for SFCDA and NO_SFCDA, re- similation of surface data, the 3-h ensemble mean spectively. forecast shows better fits to sounding observations as well as improved spread-error relationships for all four afternoon and dry bias by up to 1.6 K during nighttime fields in the entire atmosphere. The superiority of (Figs. 5b and 5c). These bias errors in the surface fields SFCDA is also evident in domain 1, though it is slightly may have various causes including the representative- smaller (not shown). The rms innovations from radio- 2 ness error due to the smoother model terrain height than sonde observations decrease by up to 0.4 m s 1 for hor- the actual orography, errors in the soil initialization, izontal winds at 850 mb and up to 0.4 K in temperature at systematic deficiencies of the surface-layer scheme in 925 mb. Bootstrapping over 3000 resamples, these dif- representing the surface exchange coefficients (which ferences between the two experiments are statistically are then used to compute surface fluxes), the LSM in significant at 95% confidence intervals. The positive representing the evolution of soil states, and the PBL impact of the surface data assimilation is not limited to scheme in parameterizing the subgrid-scale turbulent the surface or the boundary layer but improves the mixing. Despite all these different error sources, it is forecasts of the entire troposphere in all variables. APRIL 2014 H A A N D S N Y D E R 1497

TABLE 3. Mesonet veriﬁcation of 3-h forecast, averaged over all common stations in domain 1 for the month of June 2008. SFC stands for SFCDA and NOSFC represents NO_SFCDA.

2 2 U10 (m s 1) V10 (m s 1) T2 (K) Td2 (K) Altimeter (hPa) SFC NOSFC SFC NOSFC SFC NOSFC SFC NOSFC SFC NOSFC rmse 2.09 2.19 2.06 2.14 2.05 2.21 2.26 2.66 1.47 1.53 Bias 0.78 0.78 0.17 0.10 20.06 20.30 0.02 20.45 0.07 0.15 Spread 0.56 0.75 0.57 0.78 0.48 0.66 0.70 1.14 0.61 0.77

Now we return to the surface bias of the analyses and surface (S.-Y. Hong and J. Dudhia 2013, personal short-range forecasts. As shown in Fig. 5, the assimila- communication). As described in section 2b, our soil tion of surface observations reduces the forecast bias states freely evolve throughout the month-long period. errors in surface thermodynamic fields more effectively Further investigation is required for this significant issue than in surface winds, we therefore focus on tempera- on how to properly treat the soil states in the assimila- ture and dewpoint at 2 m. tion of surface observations in the cycling context. Figure 8 illustrates the time-mean innovation at each In terms of surface winds, there was stronger southerly METAR station in the SFCDA experiment. Surface wind over Texas and stronger westerly wind over the temperature (Fig. 8a) and dewpoint (Fig. 8b) in the 3-h Rockies compared to METAR observations (not ensemble mean forecast are too warm and slightly dry in shown). But as illustrated in Fig. 5 and Table 3, SFCDA the central United States but are colder and wetter than had little effect on the domain-averaged surface wind observations elsewhere. To see how the bias varies bias in the month-long statistics. through the diurnal cycle, we plot a time series of the To confirm that the assimilation of surface observa- bias averaged over all stations at different UTC cycles in tions updated the thermodynamic fields in the right di- Fig. 8c. It shows prominent warm and dry bias near the rection, Fig. 9 shows time-mean analysis increments (i.e., surface during nighttime (from 0300 to 1200 UTC) while analysis minus background) in the SFCDA experiment. the cold and wet bias is dominant during daytime (from The horizontal distribution of the time-mean analysis 1500 to 0000 UTC). This implies that the amplitude of increments (Fig. 9a) matches well with that of the mean the diurnal cycle for surface temperature forecast is bias errors (Fig. 8a) but with opposite sign, which generally underestimated and the moisture diurnal cycle demonstrates that SFCDA tends to reduce the bias er- is out of phase between the forecast and the observa- rors in the background forecast. Separating the cycles tions for this summer month. according to the time of day shows that the analysis These bias errors are consistent with those found in decreases the warm and dry bias during the nighttime the previous studies. Shin and Hong (2011) compared (Fig. 9b) and reduces the cold and wet bias before the five different PBL schemes available in WRF and found PBL collapses, especially in the western United States the warm bias in the stable boundary condition common (Fig. 9c). in all the PBL parameterizations owing to the un- We also investigated the systematic impact of the derestimation of the surface cooling rate. The cold and surface changes on the boundary layer structure by wet bias in the convective boundary layer was also re- computing the ensemble- and time-mean vertical pro- ported in Hu et al. (2010) for three different PBL files of potential temperature across the domain. At the schemes in WRF including the YSU PBL during the summertime. In a similar cycling study using WRF– DART for mesoscale assimilation over the CONUS domain, Romine et al. (2013) also found colder and wetter surface bias with the Mellor–Yamada–Janjic (MYJ) scheme (Janjic 1994) than with YSU in their short-range forecasts. The reason for the under- prediction of surface temperature in the convective regime is, however, still not clear. As well as the model errors in the surface layer and the PBL schemes and the imperfection of the surface diagnostic algorithm, the initialization of ensemble soil states during the cycling FIG. 6. As in Fig. 5c, but for all the mesonet stations used in each can also be an important factor in characterizing the experiment in domain 2 at 15-km grid resolution. Ensemble spread diurnal variation of thermodynamic bias errors near the is not depicted. 1498 MONTHLY WEATHER REVIEW VOLUME 142

FIG. 7. Sounding verification of SFCDA (black) and NO_SFCDA (gray) in terms of (a) u wind, (b) y wind, (c) temperature, and (d) dewpoint in domain 2 for the whole month of June 2008. Total number of observations used in the verification for both experiments are marked as an open circle (o) with the x axis on top. location of the large positive analysis increment at analysis increment at the surface in SFCDA (Fig. 9c) 2100 UTC [marked as a star (+) in northwestern New resulted in raising (lowering) the PBL height in the Mexico in Fig. 9c], the reduction of the cold surface bias following forecast compared to NO_SFCDA, as shown in the analysis enhances the development of PBL and in Fig. 10b. makes the whole PBL structure warmer in the 3-h fore- Another important question is how long the influence cast during the daytime (Fig. 10a). The corresponding of surface data assimilation lasts in forecasts. Therefore, wind profile indicates that turbulent mixing was much we extended forecasts from the ensemble mean analyses stronger throughout the convective boundary layer in valid at 0000 and 1200 UTC every other day for 72 h SFCDA (not shown). These changes result in raising the and verified them against METAR surface observations PBL height by about 300 m as shown by the thin hori- over ;1200 stations. In Fig. 11, averaged over the 30 zontal lines in Fig. 10a. The positive (negative) mean extended deterministic forecasts, the assimilation of APRIL 2014 H A A N D S N Y D E R 1499

FIG. 8. The horizontal distribution of the mean bias (f 2 o)in (a) temperature and (b) dewpoint at 2 m averaged over all cycles for June 2008 where o is the METAR observation and f is the 3-h ensemble forecast mean in SFCDA. (c) A time series of domain- averaged bias errors at each UTC time. surface observations improved the forecast over the first 6 h in 10-m wind and 2-m temperature and for up to 72 h FIG. 9. The horizontal distribution of ensemble mean analysis in 2-m dewpoint (although we did not show all 72-h increments in 2-m temperature (color) and dewpoint (contour) in SFCDA averaged over (a) all cycles, (b) 0900 UTC cycles, and forecast lead times to zoom in the differences for the first (c) 2100 UTC cycles. Negative moisture increments are shown as 6 h). As shown in Fig. 5 and Table 3, forecasts of surface a dashed line. dewpoint fit observations better than any other variables through the cycles thanks to the more sophisticated 1500 MONTHLY WEATHER REVIEW VOLUME 142

FIG. 10. (a) Vertical profile of the 3-h ensemble forecast mean in potential temperature at the location marked as a star (+) in northwestern New Mexico in Fig. 9c, averaged over all the cycles at 2100 UTC. Thin black and gray horizontal lines around 3 km indicate the PBL height in SFCDA and NO_SFCDA, respectively. (b) The horizontal distribution of the PBL height difference (m) between SFCDA and NO_SFCDA in the 3-h ensemble forecast mean, averaged over all 2100 UTC cycles. observation-error specification at each station, which Figure 12 shows the FSS for the 3-h accumulated resulted in the biggest and the longest improvement in precipitation forecast verified against NCEP stage IV the extended forecast. These differences were statisti- precipitation data at each forecast lead time. Because cally significant up to at least 24-h forecasts at the 95% the rainfall verification was done in domain 2, stage IV confidence intervals in our bootstrapping. Separating the fields (with a native resolution of 4.7 km) are in- forecasts from 0000 and 1200 UTC initialization (not terpolated to the 15-km grid before computing the FSS. shown) revealed that bias in either subset was as large as The radius of influence was chosen as 105 km since it is ;61 K in both 2-m temperature and dewpoint, but ow- comparable to the actual resolvable scale (roughly six ing to their opposite signs, the bias averaged over both grid intervals) of the 15-km grid. Overall, the scores initialization times was small, as shown in Figs. 11c and were not very sensitive to the radius of influence or to 11d. Although their bias errors varied strongly over di- the initialization time of ensemble forecast. Here, we urnal cycle, the quality of forecasts from the 0000 and show the FSS for precipitation thresholds of 1.0 and 1200 UTC analyses were similar in terms of both rms 10.0 mm over all forecasts from the 0000 UTC initialization. and bias errors (not shown). Forecasts from SFCDA analyses produced higher FSS We have also examined the quality of precipitation than the forecasts from NO_SFCDA analyses for both forecasts in the experiments using the fractions skill light and heavy rain cases up to 12-h lead time. Although score (FSS; Roberts and Lean 2008; Schwartz et al. FSS from SFCDA for the 10-mm threshold is degraded 2009). Unlike traditional point-by-point verification after 12 h, the benefits at shorter lead times together approaches, the FSS relaxes the requirement that fore- with evidence from the case study in the next section cast and observed events match at the grid scale for demonstrate a positive impact of surface data assimila- a forecast to be considered statistically perfect. Com- tion on the forecast. puting the FSS requires specifying a radius of influence r and precipitation exceedance threshold q. The number 5. A case study—9 June 2008 of grid boxes where both the forecast and observed precipitation exceed q within the radius r is determined Next we examine how the assimilation of surface ob- and divided by the total number of grid points available servations influenced the boundary layer structure and within the radius. In other words, the points that are ob- the following forecast through a severe convective case served and forecasted at the same time are transformed into associated with a strong surface front situated across the fractional values for the particular combination of q and r. model domain for several days. For this case study, we The FSS varies from 0 to 1, with a perfect forecast being 1. present all the forecasts from the 15-km grid (e.g., in D2). APRIL 2014 H A A N D S N Y D E R 1501

FIG. 11. A time series of rms innovations (solid lines) and bias (f 2 o) errors (dashed lines) computed for extended forecasts from the mean EnKF analyses with (i.e., SFCDA) and without assimilation of surface data (i.e., NO_SFCDA), verified against ;1200 METAR stations over 30 cycles (twice daily, every other day) for the month of June 2008 in terms of (a) u wind and (b) y wind at 10 m and (c) temperature and (d) dewpoint at 2 m. a. Case description and experiment design The strong southerly LLJ transported moisture and contributed to the instability with the mixed-layer con- An upper-level trough approached over the Rockies vective available potential energy (CAPE) on the order 2 and the surface low was located in central Colorado at of 2000 J kg 1 ahead of the frontal boundary. Within 1200 UTC 8 June 2008 developing a strong surface cold a couple of hours, the squall line has elongated into front over the central plains (not shown). At that time, northern Oklahoma along the front. By 0000 UTC 9 there was a strong southwesterly low-level jet (LLJ) ahead June 2008, the squall line was further intensified in the of the front reporting the maximum speed of 63 kt southern plains, producing the maximum radar re- 2 (32.4 m s 1) at Topeka, Kansas. At 1800 UTC 8 June flectivity of 75 dBZ in western Texas (Fig. 13b). This also 2008, the cold front extended from western Iowa to the resulted in high winds, large hail, and heavy rainfall 2 Oklahoma–Texas Panhandle. A squall line was initiated reaching a maximum of 66 mm (6 h) 1 at Gage, Okla- along the surface front in northwestern Kansas (Fig. 13a). homa. As the boundary layer began to cool and 1502 MONTHLY WEATHER REVIEW VOLUME 142

21 FIG. 12. FSS as a function of time for precipitation thresholds of (a) 1.0 and (b) 10.0 mm (3 h) computed over the entire 15-km domain for a 105-km radius of inﬂuence, aggregated over all forecasts from 0000 UTC initialization time. decouple in the evening, the LLJ increased across the importance of surface observations at the analysis southern plains into central Oklahoma. With veering when the background forecast already contains all the wind shear and strong instability ahead of the slowly surface information from the previous cycles in the propagating cold front, the squall line merged into a se- second experiment. vere mesoscale convective system (MCS) that persisted b. Inﬂuence of surface observations for the next 24 h (not shown).

To investigate how the assimilation of surface obser- Figure 14 shows equivalent potential temperature (ue) vations affected the initiation of the MCS, we selected at 2 m and the CAPE from the ensemble mean analysis 2 the analysis cycle at 1800 UTC 8 June 2008, which is in the two experiments. In SFC , there was a large about 3 h before convective initiation over northern horizontal ue gradient associated with the strong surface Oklahoma. front ranging from central Kansas to the Oklahoma– Isolating the effects of surface observations is subtle, Texas Panhandle. The large CAPE ahead of the surface because they influence the analysis directly, at the time front was indicative of a favorable condition for con- 1 they are assimilated, and indirectly, by contributing to vective initiation at 1800 UTC (Fig. 14a). In NOSFC , improvements in the prior forecasts as the assimilation even with the good background with all the previous system cycles. Thus, we performed two additional ex- surface observation information and the assimilation 2 periments. The first experiment, named SFC , assimi- of all other upper-level observations, the surface tem- lated surface observations as well as all other observations perature and moisture gradients were weaker than in 2 but used the prior ensemble forecasts from NO_SFCDA SFC , as was the convective instability in the warm (which were unaffected by previous surface observa- sector (Fig. 14b). 1 tions). The second experiment, named NOSFC , ex- Figure 15 illustrates the ensemble mean analysis included surface observations from the assimilation but crement (i.e., posterior–prior) at the lowest model level 2 used the prior ensemble forecasts from SFCDA (which in SFC , where negative temperature increments were depended on surface observations before the 1800 UTC as large as 3 K and horizontal wind vector increments 2 cycle). That is, the experiment name itself indicates were bigger than 3 m s 1. The structure of these in- whether or not the surface data was assimilated in the crements enhanced the surface cold front and displaced 1 1800 UTC cycle, and the superscript implies that the it farther to the southeast. In NOSFC , there was no background was generated from the analysis with (1) noticeable increment in all the fields at the lowest level and without (2) the assimilation of surface observa- (not shown), indicating that the low-level analysis in- 2 tions in the previous cycles. Here the idea is to ex- crements in SFC were produced almost exclusively by amine how much information we can get from the surface observations. surface data assimilation when the prior is bad near Figure 16 shows the cross section of analysis in- the surface in the first experiment, then to check the crements in temperature and vertical velocity along the APRIL 2014 H A A N D S N Y D E R 1503

FIG. 13. Surface observations with radar reﬂectivity valid at (a) 1800 UTC 8 Jun and (b) 0000 UTC 9 Jun 2008.

2 line AB in Fig. 15. In SFC , the biggest increments ap- velocity, strong and narrow positive increments oc- peared behind the cold front and spanned ;8 model curred in the entire atmosphere above the front with the 2 levels up to 2-km height. Crucially, the horizontal and maximum up to ;7.5 cm s 1, while smaller negative in- 2 vertical extent of the analysis increments in SFC were crements were present behind the front at lower levels, not predeﬁned like in classical nudging techniques, but consistent with the displacement of the cold front were objectively determined by ensemble covariances (Fig. 16a). Without surface data assimilation, upper-level that reﬂected the mesoscale structure and dynamics of observations still produced positive w increments in the the PBL as simulated in the model. In terms of vertical upper troposphere but they were somewhat broader and 1504 MONTHLY WEATHER REVIEW VOLUME 142

21 FIG. 14. Equivalent potential temperature at 2 m (contoured every 5 K) and CAPE (J kg ) (shaded) in the analysis 2 1 valid at 1800 UTC 8 Jun 2008 from (a) SFC and (b) NOSFC experiments. shifted downstream (Fig. 16b). More importantly, 15-km grid simulated the squall line a bit broader than 1 NOSFC failed to provide the significant temperature NCEP stage IV data (which was processed at 4.7-km 2 and wind changes necessary for the development of the resolution), but SFC accurately predicted its location 1 surface cold front in the right place at this analysis time. near the Oklahoma–Kansas border. NOSFC predicted These results prove that surface observations play heavy rainfall in southern Iowa, but missed the squall a dominant role in making a considerable difference line in northwestern Oklahoma that developed into the near the surface and contribute to the correct timing and MCS for the next 24 h in the central plains. The en- location of the surface front. The horizontal and the semble mean forecast is superimposed by the probabil- 2 vertical range within which surface data affects in EnKF ity of the heavy rainfall [$50 mm (6 h) 1] in the vary depending on the model error structure in the ensemble forecast in both experiments contouring every simulated PBL and the localization radius we specified 10% in Figs. 17a,b. They indicate more than 5 members based on the ensemble size and the observation network among 50 predicted the accumulation larger than 50 mm in use. in the contoured area. As shown in Fig. 17b, most To understand the relative impact of the background, we also present the same increments from the original SFCDA experiment that had continuous cycling with all available observations including surface data (Fig. 16c). Note this figure does not show the actual fields but their increments. As the prior was already good (with all the surface information digested in the previous times), the increments were slightly smaller. One noticeable thing is that the SFCDA analysis showed cold air slightly shifted toward B in the downstream region and the vertical motion associated with the surface front was right ahead of the negative temperature increment. This indicates that the front was located about 30 km ahead in SFCDA 2 and moving faster than in SFC when the background was consistent with the surface information in the previous time. The impact of background was thus not as 2 big as SFC in this particular case. The better analysis leads to better precipitation forecasts at longer forecast lead times. Figure 17 depicts the ensemble-mean accumulated rainfall (mm) for 6 h in 2 1 FIG. 15. Analysis increments in temperature (K, shaded) in 2 SFC , NOSFC , and NCEP stage IV data at 0000 UTC SFC at the lowest model level at 1800 UTC 8 Jun 2008. Arrows 2 9 June 2008. The mean of the 6-h ensemble forecast at the indicate horizontal wind vector increments larger than 3 m s 1. APRIL 2014 H A A N D S N Y D E R 1505

FIG. 16. Cross section of analysis increments in temperature (K, 2 shaded) and vertical velocity w (contoured every 2.5 cm s 1) in (a) 2 1 SFC , (b) NOSFC , and (c) SFCDA experiments. Negative increments in temperature and w are represented as shading and contours with dashed lines, respectively.

FIG. 17. Accumulated rainfall from the mean of 6-h ensemble 2 forecasts valid at 0000 UTC 9 Jun 2008 in (a) SFC and (b) 1 members overpredicted heavy precipitation over Iowa NOSFC experiments, compared to (c) NCEP stage IV data. 2 and missed the rainband over Oklahoma and Kansas in Probability (%) of accumulated rainfall over 50 mm (6 h) 1 in the 1 NOSFC . The original SFCDA and NO_SFCDA ex- ensemble forecast is contoured from 10% every 10% interval in (a) periments showed a similar comparison in terms of the and (b). precipitation forecast, but were not presented here. 1506 MONTHLY WEATHER REVIEW VOLUME 142

6. Summary and conclusions surface observations also increased convective instability ahead of the front. These changes in the PBL Surface observations (e.g., 2-m thermodynamic fields and mesoscale flow near the surface led to improve- and 10-m winds) resolve mesoscale variations both spatially ments in the subsequent precipitation forecast. and temporally. This paper has explored assimilation To see benefits from surface observations, some care of these surface observations with an EnKF employing is required in assigning observation-error variances. WRF implemented via DART. The assimilation system Comparing rms innovations against the square root of cycles for one summer month of June 2008, producing the sum of the ensemble variance and observation-error analyses every 3 h over the contiguous United States and variance showed that the variances assigned to surface including an embedded nested domain at 15-km resolution. observations in the GSI were significantly too large in Relative to experiments using only other conventional our system. Based on the innovation statistics, we re- 2 observations, assimilation of surface observations im- duced observation-error standard deviations to 1.5 m s 1 proves both analyses and short-term forecasts of surface and 1.5 K for 10-m winds and 2-m temperature, re- variables. Improvement of analyses, evaluated by com- spectively, and modified the observation-error variance paring against independent (e.g., unassimilated) meso- for dewpoint as a function of relative humidity and net observations, is as much as 16% in the components temperature. The new observation-error statistics im- of 10-m winds, 15% in 2-m temperature, and 28% in 2-m proved not only the reliability of the ensemble system dewpoint. Assimilation of surface observations also but also the quality of the following forecast, with the improves the fit of short-range forecasts to surface winds greatest enhancements following from the formulation and temperature for 6 h, and for substantially longer of the error variance for dewpoint observations. hours for dewpoint temperature. These results are all There are three issues that offer clear potential for statistically significant at the 95% confidence intervals. further improvements in assimilating surface observa- In contrast to previous results with the EnKF, assim- tions, which we leave as topics for future work. First, ilation of surface observations improves forecasts aloft representativeness errors are particularly acute for sur- as well, as seen by reduced differences from radiosondes face observations. We have accounted for these only through at least 500 hPa, with the biggest improvement crudely and indirectly, by rejecting observations at in horizontal winds at 850 hPa. These improvements which the station elevation differs from the model translate into better precipitation forecasts, as measured orography by more than 300 m. by the FSS, through 12-h lead times. Second, coupling to the land surface strongly in- Bias accounts for a significant portion of the forecast fluences atmospheric variables within the PBL but we error for surface variables. In 2-m temperature, the ab- have not updated or adjusted the land surface state solute value of the time-averaged 3-h forecast difference during assimilation. This is likely far from optimal, as from observations is greater than 0.5 K at most stations. much previous work, beginning with Mahfouf (1991), The bias has a pronounced diurnal variation, whose has emphasized the role of surface observations in es- phase is consistent with an underestimation of the di- timation of the land surface state. An important re- urnal cycle on average, and also varies coherently across search problem is the development of techniques for the CONUS. Both these characteristics point to de- jointly updating the atmospheric and land surface states ficiencies in the forecast model as the source of the bias. when assimilating surface observations. Assimilating surface observations reduces biases in Finally, the magnitude of forecast bias relative to surface variables and also leads, through the analysis- surface observations indicates that, not surprisingly, the forecast cycle, to accompanying, systematic changes in forecast model has important deficiencies and errors. the structure and depth of the PBL. Forecast model errors affect our results by degrading The benefits of assimilating surface observations both the 3-h ensemble mean prediction of the observa- come in part from the ability of the EnKF to produce tions and the forecast covariances that are used in the analysis increments that are consistent with the structure EnKF, yet the assimilation system implemented here and dynamics of the boundary layer. In addition to the only implicitly accounts for those errors, through an systematic effects of surface observations on the depth adaptive inflation of the forecast covariances. Direct of the PBL, our case study of a squall line illustrated how improvements to the model are certainly helpful and the EnKF used information in the surface observations have been the focus of several recent studies with WRF- to modify the intensity of the cold boundary layer be- DART (Cavallo et al. 2011; Torn and Davis 2012). In hind the surface front and to displace the frontal addition, we are actively investigating various model boundary, together with the associated wind direction, error schemes for the EnKF and will report results in to be consistent with observations. Assimilation of a subsequent study. APRIL 2014 H A A N D S N Y D E R 1507

Acknowledgments. We thank David Dowell, Glen Dirren, S., R. D. Torn, and G. J. Hakim, 2007: A data assimilation Romine, Jimy Dudhia, and Jeff Anderson for helpful case study using a limited-area ensemble Kalman filter. Mon. discussions on this work. Observations were provided by Wea. Rev., 135, 1455–1473. Dudhia, J., 1989: Numerical study of convection observed dur- NOAA/MADIS. We appreciate the support from Tim ing the winter monsoon experiment using a mesoscale two- Hoar and Nancy Collins on the data processing in the dimensional model. J. Atmos. Sci., 46, 3077–3107. DART team at NCAR, Glen Romine on the image Fujita, T., D. J. Stensrud, and D. C. Dowell, 2007: Surface data processing, Dave Gill for the WRF preprocessing, and assimilation using an ensemble Kalman filter approach with Craig Schwartz on the precipitation verification. This initial condition and model physics uncertainties. Mon. Wea. Rev., 135, 1846–1868. study was partially funded by the U.S. Air Force Gaspari, G., and S. E. Cohn, 1999: Construction of correlation Weather Agency. functions in two and three dimensions. Quart. J. Roy. Meteor. Soc., 125, 723–757. Hacker, J. P., and C. Snyder, 2005: Ensemble Kalman filter as- REFERENCES similation of fixed screen-height observations in a parameter- Aksoy, A., D. C. Dowell, and C. Snyder, 2009: A multicase com- ized PBL. Mon. Wea. Rev., 133, 3260–3275. parative assessment of the ensemble Kalman filter for assim- ——, and Coauthors, 2011: The U.S. Air Force Weather Agency’s ilation of radar observations. Part I: Storm-scale analyses. mesoscale ensemble: Scientific description and performance Mon. Wea. Rev., 137, 1805–1824. results. Tellus, 63, 625–641. Ancell, B. C., C. F. Mass, and G. J. Hakim, 2011: Evaluation of Hamill, T. M., J. S. Whitaker, and C. Snyder, 2001: Distance- surface analyses and forecasts with a multiscale ensemble dependent filtering of background error covariance estimates Kalman filter in regions of complex terrain. Mon. Wea. Rev., in an ensemble Kalman filter. Mon. Wea. Rev., 129, 2776–2790. 139, 2008–2024. Hong, S.-Y., J. Dudhia, and S.-H. Chen, 2004: A revised approach Anderson, J. L., 2001: An ensemble adjustment Kalman filter for to ice microphysical processes for the bulk parameterization of data assimilation. Mon. Wea. Rev., 129, 2884–2903. clouds and precipitation. Mon. Wea. Rev., 132, 103–120. ——, 2009: Spatially and temporally varying adaptive covariance ——, Y. Noh, and J. Dudhia, 2006: A new vertical diffusion inflation for ensemble filters. Tellus, 61A, 72–83. package with an explicit treatment of entrainment processes. ——, T. Hoar, K. Raeder, H. Liu, N. Collins, R. Torn, and A. Mon. Wea. Rev., 134, 2318–2341. Avellano, 2009: The Data Assimilation Research Testbed: A Houtekamer, P. L., and H. L. Mitchell, 2001: A sequential en- community facility. Bull. Amer. Meteor. Soc., 90, 1283–1296. semble Kalman filter for atmospheric data assimilation. Mon. Benjamin, S. G., and Coauthors, 2004: An hourly assimilation– Wea. Rev., 129, 123–137. forecast cycle: The RUC. Mon. Wea. Rev., 132, 495–518. Hu, X.-M., J. W. Nielsen-Gammon, and F. Zhang, 2010: Evalua- Berner, J., S.-Y. Ha, J. P. Hacker, A. Fournier, and C. Snyder, 2011: tion of three planetary boundary layer schemes in the WRF Model uncertainty in a mesoscale ensemble prediction system: model. J. Appl. Meteor. Climatol., 49, 1831–1844. Stochastic versus multiphysics representations. Mon. Wea. Huang, X. Y., and Coauthors, 2009: Four-dimensional variational Rev., 139, 1972–1995. data assimilation for WRF: Formulation and preliminary re- Cavallo, S. M., J. Dudhia, and C. Snyder, 2011: A multilayer upper- sults. Mon. Wea. Rev., 137, 299–314. boundary condition for long wave radiative flux to correct Janjic, Z. I., 1994: The step-mountain eta coordinate model: Fur- temperature biases in a mesoscale model. Mon. Wea. Rev., ther developments of the convection, viscous sublayer, and 139, 1952–1959. turbulence closure schemes. Mon. Wea. Rev., 122, 927–945. Chen, F., and J. Dudhia, 2001: Coupling an advanced land surface– Kain, J., 2004: The Kain–Fritsch convective parameterization: An hydrology model with the Penn State–NCAR MM5 modeling update. J. Appl. Meteor., 43, 821–825. system. Part I: Model implementation and sensitivity. Mon. Kleist, D. T., D. F. Parrish, J. C. Derber, R. Treadon, W.-S. Wu, and Wea. Rev., 129, 569–585. S. Lord, 2009: Introduction of the GSI into the NCEP global Compo, G. P., J. S. Whitaker, and P. D. Sardeshmukh, 2006: Fea- data assimilation system. Wea. Forecasting, 24, 1691–1705. sibility of a 100-year reanalysis using only surface pressure Lin, X., and K. G. Hubbard, 2004: Uncertainties of derived dew- data. Bull. Amer. Meteor. Soc., 87, 175–190. point temperature and relative humidity. J. Appl. Meteor., 43, de Rosnay, P., M. Drusch, D. Vasiljevic, G. Balsamo, C. Albergel, 170–181. and L. Isaksen, 2013: A simplified extended Kalman filter for Mahfouf, J.-F., 1991: Analysis of soil moisture from near-surface the global operational soil moisture analysis at ECMWF. parameters: A feasibility study. J. Appl. Meteor., 30, 1534–1547. Quart. J. Roy. Meteor. Soc., 139, 1199–1213, doi:10.1002/ Meng, Z., and F. Zhang, 2008: Tests of an ensemble Kalman filter qj.2023. for mesoscale and regional-scale data assimilation. Part III: Dee, D. P., 1995: On-line estimation of error covariance parameters for Comparison with 3DVAR in a real-data case study. Mon. atmospheric data assimilation. Mon. Wea. Rev., 123, 1128–1145. Wea. Rev., 136, 522–540. Derber, J., and F. Bouttier, 1999: A reformulation of the back- Mitchell, H. L., and P. L. Houtekamer, 2000: An adaptive ensemble ground error covariance in the ECMWF global data assimi- Kalman filter. Mon. Wea. Rev., 128, 416–433. lation system. Tellus, 51A, 195–221. Mlawer, E. J., S. J. Taubman, P. D. Brown, M. J. Iacono, and S. A. Desroziers, G., and S. Ivanov, 2001: Diagnosis and adaptive tuning Clough, 1997: Radiative transfer for inhomogeneous atmo- of observation-error parameters in a variational assimilation. spheres: RRTM, a validated correlated-k model for the Quart. J. Roy. Meteor. Soc., 127, 1433–1452. longwave. J. Geophys. Res., 102 (D14), 16 663–16 682. ——, L. Berre, B. Chapnik, and P. Poli, 2005: Diagnosis of obser- Roberts, N. M., and H. W. Lean, 2008: Scale-selective verification vation, background and analysis-error statistics in observation of rainfall accumulations from high-resolution forecasts of space. Quart. J. Roy. Meteor. Soc., 131, 3385–3396. convective events. Mon. Wea. Rev., 136, 78–97. 1508 MONTHLY WEATHER REVIEW VOLUME 142

Romine, G., C. Schwartz, C. Snyder, J. Anderson, and M. Weisman, Tippett, M. K., J. L. Anderson, C. H. Bishop, T. M. Hamill, and J. S. 2013: Model bias in a continuously cycled assimilation Whitaker, 2003: Ensemble square root filters. Mon. Wea. Rev., system and its influence on convection-permitting fore- 131, 1485–1490. casts. Mon. Wea. Rev., 141, 1263–1284. Torn, R. D., and G. J. Hakim, 2008: Performance characteristics of Schwartz, C. S., and Coauthors, 2009: Next-day convection- a pseudo-operational ensemble Kalman filter. Mon. Wea. allowing WRF model guidance: A second look at 2-km versus Rev., 136, 3947–3963. 4-km grid spacing. Mon. Wea. Rev., 137, 3351–3372. ——, and C. A. Davis, 2012: The influence of shallow convection on Shin, H., and S.-Y. Hong, 2011: Intercomparison of planetary tropical cyclone track forecasts. Mon. Wea. Rev., 140, 2188–2197. boundary-layer parameterizations in the WRF model for a single ——, G. J. Hakim, and C. Snyder, 2006: Boundary conditions for day from CASES-99. Bound.-Layer Meteor., 139, 261–281. limited-area ensemble Kalman filers. Mon. Wea. Rev., 134, Skamarock, W. C., and Coauthors, 2005: A description of the ad- 2490–2502. vanced research WRF version 3. Tech. Note NCAR/TN- Wang, X., and C. H. Bishop, 2003: A comparison of breeding and 4751STR, NCAR, 113 pp. ensemble transform Kalman filter ensemble forecast schemes. Stauffer, D. R., N. L. Seaman, and F. S. Binkowski, 1991: Use of J. Atmos. Sci., 60, 1140–1158. four-dimensional data assimilation in a limited-area me- Zhang, D.-L., and R. A. Anthes, 1982: A high-resolution model of soscale model. Part II: Effects of data assimilation within the planetary boundary layer—Sensitivity tests and compari- the planetary boundary layer. Mon. Wea. Rev., 119, 734–754. sons with SESAME-79 data. J. Appl. Meteor., 21, 1594–1609.