Warm Season Mesoscale Superensemble Precipitation Forecasts in the Southeastern United States

T. J. Cartwright West Virginia State University, and West Virginia State Community and Technical College, Institute, West Virginia

Search for other papers by T. J. Cartwright in
Current site
Google Scholar
PubMed
Close
and
T. N. Krishnamurti The Florida State University, Tallahassee, Florida

Search for other papers by T. N. Krishnamurti in
Current site
Google Scholar
PubMed
Close
Full access

Abstract

With current computational limitations, the accuracy of high-resolution precipitation forecasts has limited temporal and spatial resolutions. However, with the recent development of the superensemble technique, the potential to improve precipitation forecasts at the regional resolution exists. The purpose of this study is to apply the superensemble technique to regional precipitation forecasts to generate more accurate forecasts pinpointing exact locations and intensities of strong precipitation systems. This study will determine the skill of a regional superensemble forecast out to 60 h by examining its equitable threat score and its false alarm ratio. The regional superensemble consists of 12–60-h daily quantitative precipitation forecasts from six models. Five are independent operational models, and one comes from the physically initialized Florida State University regional spectral model. The superensemble forecasts are verified during the summer 2003 season over the southeastern United States using merged River Forecast Center stage-IV radar–gauge and satellite analyses. Precipitation forecasts were skillful in outperforming the operational models at all model times. Precipitation results were stratified by time of day to allow detections of the diurnal cycle. As expected, warm season daytime precipitation is commonly forced by convection, which is difficult to accurately model. Major synoptic regimes, including subtropical highs, midlatitude troughs/fronts, and tropical cyclones, were examined to determine the skill of the superensemble under various synoptic conditions.

* Current affiliation: MU-ADVANCE Program, Marshall University, Huntington, West Virginia

Corresponding author address: Dr. Tina J. Cartwright, MU-ADVANCE Program, Marshall University, 241J Byrd Biotech Center, Huntington, WV 25755. Email: johnson516@marshall.edu

Abstract

With current computational limitations, the accuracy of high-resolution precipitation forecasts has limited temporal and spatial resolutions. However, with the recent development of the superensemble technique, the potential to improve precipitation forecasts at the regional resolution exists. The purpose of this study is to apply the superensemble technique to regional precipitation forecasts to generate more accurate forecasts pinpointing exact locations and intensities of strong precipitation systems. This study will determine the skill of a regional superensemble forecast out to 60 h by examining its equitable threat score and its false alarm ratio. The regional superensemble consists of 12–60-h daily quantitative precipitation forecasts from six models. Five are independent operational models, and one comes from the physically initialized Florida State University regional spectral model. The superensemble forecasts are verified during the summer 2003 season over the southeastern United States using merged River Forecast Center stage-IV radar–gauge and satellite analyses. Precipitation forecasts were skillful in outperforming the operational models at all model times. Precipitation results were stratified by time of day to allow detections of the diurnal cycle. As expected, warm season daytime precipitation is commonly forced by convection, which is difficult to accurately model. Major synoptic regimes, including subtropical highs, midlatitude troughs/fronts, and tropical cyclones, were examined to determine the skill of the superensemble under various synoptic conditions.

* Current affiliation: MU-ADVANCE Program, Marshall University, Huntington, West Virginia

Corresponding author address: Dr. Tina J. Cartwright, MU-ADVANCE Program, Marshall University, 241J Byrd Biotech Center, Huntington, WV 25755. Email: johnson516@marshall.edu

1. Introduction and background

In recent years, quantitative precipitation forecasts (QPFs; see Table 1 for a list of acronym used in this paper) have improved their skills because of increases in resolution (e.g., Black 1994). Summertime QPFs remain a challenge because convective rainfall is poorly predicted (e.g., Stensrud et al. 2000). Numerical models have problems simulating summertime convection in the correct location and at the correct time because of the small-scale nature of the features that act to initiate convection (Kain and Fritsch 1992; Stensrud and Fritsch 1994a, b; Toth et al. 1998; Buizza et al. 1999a). The resulting reliance on various parameterization schemes introduces error into the forecast for convective rainfall. These parameterization schemes include convection, boundary layer processes, radiation, evapotranspiration, and cloud microphysical processes. However, with the recent development of the superensemble technique, summertime tropical QPFs have dramatically improved at the global model resolution (Krishnamurti et al. 2000a, b). The purpose of this study is to apply the superensemble technique to regional precipitation forecasts in order to generate more accurate forecasts pinpointing the exact locations of strong precipitation systems.

Recently, studies have noted the need for independent model forecasts and have utilized sets of independent numerical weather prediction (NWP) model forecasts from independent centers. Ebert (2001) acquired seven operational NWP models to examine the skill of 24- and 48-h daily QPFs over Australia, the poor man’s ensemble technique. As was noted in Ebert’s study, the use of independent models captures the uncertainties in both initial conditions and model formulation through the differences in the input data and forecast parameterizations of its member models.

Ebert’s study utilized 28 months of 24- and 48-h model QPFs for rainfall over Australia at 1° resolution with rain gauge observations for verification. Probabilistic forecasts were evaluated using the Brier skill score and the relative operating characteristic. The 24-h forecasts showed skill for rainfall up to and exceeding 50 mm day−1, but the 48-h forecasts showed skill for only low rain rates and did not perform better than climatology (Ebert 2001). Averaging the forecasts increased the spatial extent and reduced the rain intensity as compared to the member models (Ebert 2001). Deterministic forecasts were evaluated using the RMS error, bias score, and equitable threat score. The mean RMS error was 20% lower than the mean values for the member models’ QPFs and was usually better than the best-performing model for the given day (Ebert 2001). The equitable threat scores were also improved by the ensemble mean because the improved rain detection outweighed the increase in the number of false alarms.

The Ebert (2001) study revealed several conclusions important for this study. The number of member models that capture a particular rain event was shown to be a good predictor of how well the ensemble will capture that event (Ebert 2001). The relative cost for independent operational centers QPFs is low and can be done with limited computer resources. As upgrades are made at various centers, the ensemble QPF will also reflect those improvements. The poor man’s ensemble samples uncertainties in both the initial conditions and the model formulation, which strengthens the overall diversification of the model forecasts and their QPFs. It is therefore less prone to systematic biases and errors.

Applying the superensemble technique to a set of higher-resolution models is computationally expensive. Generating a superensemble forecast requires multiple forecasts over many days. High-resolution modeling is in and of itself very computationally intensive and developing a robust superensemble further amplifies this need. This study reveals that the superensemble technique will provide a means to enhance warm season rainfall prediction with currently available research and operational forecasts. Section 2 outlines the precipitation data used for verification and superensemble calculation. Section 3 contains an overview of the member model data. Section 4 outlines the methodology used to create the superensemble precipitation forecast. Section 5 discusses the results created, and section 6 contains an overview and summary of this study.

2. Precipitation data

To generate the coefficients for the superensemble forecast, observations of precipitation are needed to verify the performance of the member models. During a training period, comparisons are made between each of the member models and the precipitation observations at each grid point. The accuracy of the precipitation estimates is paramount. In this study, the River Forecast Center (RFC) stage-IV quantitative precipitation estimates (QPEs) were used wherever possible over the study domain.

The calculations used to generate the stage-IV QPEs are produced from a multitiered process. During the stage-I process, National Weather Service WSR-88D radars generate hourly digital precipitation (HDP) products. Many offices utilize the standard ZR relationship (300R1.4), but some offices that are located at lower latitudes utilize a more tropical ZR relationship (300R1.2). The chosen ZR relationship may also depend on the season or weather system. This initial HDP product is mapped onto a 4 km × 4 km polar stereographic grid for the particular region. The stage-II dataset consists of this initial stage-I data, which has undergone bias adjustment from a regional rain gauge network. The stage-III data product is a compilation of the stage-II data from regional networks composited onto a 4-km grid that has been adjusted for radar inconsistencies. The final product, stage IV, is a national composite of the regional stage-III datasets. The stage-IV data comes from the regional hourly and 6-hourly multisensor precipitation analyses produced by the 12 RFCs. Sources of error for these precipitation estimates may result from partial beam filling, beam overshooting, beam blockage, an uncertain ZR relationship, hail, ground returns from anomalous propagation, ground clutter, and ground clutter suppression. Manual quality control is performed at the RFCs on the final product.

To calculate the superensemble coefficients over the Gulf of Mexico and the Atlantic Ocean where the ground-based estimates are not available, this study utilizes the experimental TRMM multisatellite precipitation analyses (MPA-RT) dataset. The satellite estimate is based on merged microwave and calibrated infrared (IR) estimates of precipitation (Huffman et al. 2003). It combines the following products: NOAA/NWS/CPC merged 4-km geostationary satellite IR brightness temperature data, U.S. Navy/FNMOC satellite data records (SDRs), SSM/I brightness temperatures, and TRMM real-time 2A12RT estimates of precipitation based on GPROF computations on TMI data. The combined precipitation estimate is then mapped onto a global 0.25° × 0.25° resolution map. A recent study (Katsanos et al. 2004) outlined the verification results for the MPA-RT precipitation estimates over the central and eastern Mediterranean. The satellite estimates were verified against 73 gauge observations. Their results show unbiased results for the low and medium precipitation thresholds. Overall, the error characteristics of the MPA-RT precipitation estimates are lower than the expected error of the superensemble forecast.

3. Model data

Beginning in June 2003, forecasts from nine mesoscale models were accumulated for this investigation. Table 2 outlines the model type, parameterization schemes, and resolutions for the six independent member models. Four of the potential member models are from the same FSUNRSM but with independent initial conditions. The superensemble technique requires that all models be interpolated to a uniform grid and resolution. The FSUNRSM member models have an original resolution of 43 km. Since the process of altering a model’s resolution can introduce errors, we chose to interpolate the other member models to a common grid of 43-km resolution. Six member models were available up to the 48-h forecasts, and only three where available for the 60-h forecast.

Five of the available mesoscale models originate from independent operational and research facilities throughout the United States, such as the Eta from NCEP, the experimental Eta from NSSL, the RUC from FSL, the MM5 from NCAR, and COAMPS from NRL. Four additional forecasts were created with the FSUNRSM. Three were generated by the use of different rainfall rates in the physical initialization (PI) process and the fourth FSUNRSM member was generated without any PI as a control run. The four FSUNRSM simulations can be summarized by the following:

  1. CALVAL SSM/I by Ferraro and Marks (1995)

  2. OLSON SSM/I by Olson et al. (1990)

  3. COMBINED GPROF–SSM/I–TRMM 2A12 by Kummerow et al. (1998)

  4. No PI

4. Methodology

The computation of the superensemble forecast requires a training and a forecast phase with all member models and observations available. Figure 1 describes an outline for this procedure. This study will utilize approximately 360 different 60-h forecasts from the multimodel forecasts. Using the 360 multimodel forecasts, which include the selected NWP models for 60 forecast days, and the best estimate of the respective observed rainfall estimations, a simple linear multiple regression is computed to determine the statistical weights. Each of these weights describes the model biases at each geographical location for each participating model.

The availability of the real-time operational regional models was extremely limited. Therefore, only 60 days were available for both phases (training and forecast) within the summer season of 2003. This resulted in the need to utilize cross validation to create the superensemble forecast. For any particular forecast day, other days in the member model dataset were used to generate the statistical weights needed for the superensemble forecast (Wilks 1995). It should be noted that this would not be necessary for a real-time implementation of a regional superensemble where the previous 45 days forecasts would be available. To create a relatively larger dataset of superensemble forecasts, it was necessary to utilize cross validation to create 60 days of superensemble forecasts.

An important dimension of the superensemble forecast is the length of the training period. Forty-five days of training was found to be optimal. Figure 2 demonstrates the ETSs for the 24-h forecast at various rainfall thresholds for 30, 45, and 59 days of training. Clearly, the ETS for the 24-h forecast was maximized at the various thresholds by utilizing a training period of 45 days. Lengthy training periods such as 59 days have been shown to be longer than the typical summer signal of precipitation, and shorter periods such as 30 days are not long enough to capture the best precipitation signal for summer convection (Krishnamurti et al. 2000b).

The superensemble technique uses a regular multiple regression method to obtain the regression coefficients for each ensemble forecast at each grid point for each forecast time. It is a method that combines individual forecasts from a group of models to produce an optimal ensemble forecast. The following equation describes how the superensemble prediction is created at a given grid point:
i1520-0434-22-4-873-e1
where S(t) is a superensemble prediction for day t; O is a time mean of the observed state; ai is a weight for model i; i is the model index; the summation is over N, the number of models; Fi is a time mean of the prediction by model i; and Fi(t) is a prediction by model i. The weights are computed at each grid point by minimizing the following function:
i1520-0434-22-4-873-e2
where O(t) is an observed state, t is time, and the summation is over the length of the training period. The observed state and the minimization are only available during the training phase. Due to the possibility of a singularity in the covariance matrix, constraints were placed on the possible values of the weights. The absolute value of the weights could not be greater than 3.0, which suggests that no single model could be worth more weight than one-half of the number of models in the study. Weights outside the boundary of the constraint were most often observed during the 0000 UTC and the 60-h forecast times because of the fewer number of available member models, which increased the instability in the covariance matrix. As mentioned in section 2, the verification dataset, consisting of the stage-IV and the MPA-RT precipitation estimates, can contain errors in their observations. However, the magnitudes of these errors are much lower than the magnitudes of the expected errors of the precipitation forecasts from any of the member models or the superensemble.

In an effort to evaluate the skill of the superensemble forecast as compared to simpler forecasts, several other compilation forecasts were created with the member models and observations. These include the following: simple ensemble mean, bias-removed ensemble mean, persistence, and seasonal mean.

The simple ensemble mean is the straightforward ensemble of the member models without removing their biases. The formula for computing the simple ensemble mean is given by the following:
i1520-0434-22-4-873-e3
Part of the skill of the superensemble technique can be attributed to the removal of the model bias. Therefore, a multimodel bias-removed ensemble mean was created to evaluate the skill of this simpler ensemble mean to that of the superensemble. A multimodel bias-removed ensemble is given by the following:
i1520-0434-22-4-873-e4
By comparing Eqs. (3) and (1), the superensemble scales the individual model forecast contributions according to their relative performance during the training period through the assignment of mathematical weights to them. The superensemble prediction can be rewritten in a form that resembles an ensemble of modified unbiased forecasts Fsi as
i1520-0434-22-4-873-e5
where Fsi = Nai(FiFi) + O.

5. Results

a. Member model selection

As mentioned in the first section, recent ensemble studies have shown the need for a diversity of ensemble members for mesoscale ensemble studies (Ebert 2001). Likewise, one would expect that the superensemble forecasting technique is highly dependent on the member model selection. A comparison study was conducted to see the effect of choosing different sets of member models on the quality of the superensemble forecast.

To examine the impact of the differences in the member models, two sets of member models were utilized to create two different superensemble forecasts. The list below summarizes the differences between the superensemble forecasts:

  1. six diverse members—Eta, experimental Eta, MM5, RUC, COAMPS, and FSUNRSM with TRMM PI; and

  2. four similar members—FSUNRSM with no PI, FSUNRSM with Ferraro PI, FSUNRSM with Olson PI, and FSUNRSM with TRMM PI.

The first model suite contains the maximum amount of independent model predictions. The second model suite uses only FSUNRSM simulations with variations in the initial conditions to make up the suite of four member models. The FSUNRSM runs with multiple PI simulations generate very similar forecasts. As suggested by Ebert (2001), an ensemble with member models that share their parameterization schemes and model physics would have an inferior forecast than would an ensemble with diverse member models. This is because ensemble suites with a variety of member models samples uncertainties in both the initial conditions and the model formulation, which strengthens the overall diversification of the model forecasts and their QPFs. In Fig. 3, the equitable threat scores (ETSs) and true skill statistics (TSSs) are shown for the 12-h superensemble forecasts for the two different member model suites. The model suite that contains the similar FSUNRSM simulations scores lower than those model suites that contain diversity in their precipitation forecasts.

b. Skill scores

In an effort to determine the skill of the superensemble forecast, several different measures of skill are examined to quantify the ability of the superensemble forecast compared to the member models. The equitable threat score (ETS) is a popular measure of the fraction of forecast events that were correctly predicted, accounting for hits associated with random chance. The values of the ETS range from 0 to 1, where a value of 0 indicates no skill. The second skill score to be considered is the false alarm ratio (FAR), which answers the question of how many predicted “yes” events actually did not occur (i.e., were false alarms). The values of FAR range from 0 to 1, where a perfect scoring forecast has a FAR of 0.

In Fig. 4, the ETSs are shown for the 12-, 24-, 36-, and 48-h forecasts of the superensemble forecast (SUP), the simple ensemble mean (EMN), the bias-corrected mean (BCE), and the three best member models (MEM1–3). At 12 h, the superensemble performs better than the member models and the simple ensembles at all precipitation thresholds except at 0.2 and 10 mm. At 0.2 mm, the member 3 model forecast performs slightly better. However, the skill levels for member 3 drop quickly and approach zero at precipitation thresholds greater than 10 mm. At 10 mm, member 2 outperforms the others, but all other thresholds have lower skills than the superensemble forecast. At 24 h, the ETS magnitudes are slightly lower than at 12 h, but the superensemble forecast outperforms all of the member and simple ensemble forecasts. At the lowest thresholds, again member 3 is the second-best-skilled model. However, this model’s scores are zero at all thresholds greater than 10 mm. At 36 and 48 h, the superensemble outperforms all other forecasts. Again, the member 3 forecast performs well at the low precipitation thresholds but notably underperforms at the significant precipitation thresholds. It should also be noted that when qualitatively comparing the forecasts valid at 0000 UTC (Fig. 4, left panels; 12- and 36-h forecasts) to those at 1200 UTC (Fig. 4, right panels; 24- and 48-h forecasts), the 1200 UTC forecasts behave as a function that decreases linearly as a function of threshold while the 0000 UTC forecasts have a more notable decrease in skill at the medium threshold range. This may suggest the relative difficulty in the superensemble capturing the medium and upper ranges of the precipitation values at the 0000 UTC hour. In the afternoon (0000 UTC), more convective precipitation develops. Light precipitation may cover a relatively large area, which the models may accurately forecast. This would increase the POD and, in turn, raise the ETS. However, small pinpoint locations of heavy precipitation are invariably underpredicted or forecasted in the wrong location. These missed forecasts lower the ETS for the medium to upper precipitation thresholds.

In Fig. 5, the FARs are shown for the 12-, 24-, 36-, and 48-h forecasts of the superensemble forecast (SUP), the simple ensemble mean (EMN), the bias-corrected mean (BCE), and the three best member models (MEM 1–3). These skill scores capture the strength of the superensemble technique. Most of the member models tend to overestimate the light to medium levels of precipitation. The superensemble technique removes these model biases and applies a higher relative weight to the more accurate member models. The technique removes many false alarms, which causes the superensemble to have a dramatically low FAR. As seen with the ETS, member 3 competes with the superensemble forecast at the lowest threshold values. At the precipitation threshold of 25 mm and higher, the FAR for member 3 cannot be computed because this member model does not forecast heavy precipitation events. When qualitatively comparing the scores between the 0000 UTC (Figs. 5a,c, ; 12- and 36-h forecasts) to the 1200 UTC (Figs. 5b,d; 24- and 48-h forecasts), the superensemble’s FAR starts to notably increase at the higher thresholds for the 1200 UTC forecast. Again, this indicates the influence of the diurnal cycle of precipitation and model prediction. At the afternoon time (both Figs. 5a and 5c are valid at 0000 UTC), a greater amount of isolated heavy precipitation is occurring so the opportunity for the models, including the superensemble, to properly make the forecast is better. This would cause the FAR to remain low even at the higher thresholds. However, at the morning time (1200 UTC; Fig. 5, right panel), less isolated heavy precipitation is occurring so the forecasts tend to overestimate precipitation amounts, which, in turn, increases the FAR.

In Fig. 6, the ETSs and FARs are shown for the 60-h forecasts from the superensemble, simple ensemble means, and member forecasts. Only three member models were available for the 60-h forecast time. This lack of member model diversity greatly affects the ability of the superensemble to enhance the precipitation forecast. There is relatively little improvement at the light precipitation thresholds by the superensemble technique. Overall, however, there is marginal skill at the precipitation thresholds less than 5 mm. At greater thresholds, neither the superensemble nor the other forecasts show adequate skill at predicting the precipitation location and intensity. The FAR values rapidly increase at the higher rainfall thresholds. The superensemble technique still consistently decreases the FAR rate from the simple ensemble and member forecasts at all precipitation thresholds. It should be noted that the 72-h forecasts were also examined, but only two member models were available. The superensemble technique cannot properly be applied with only two member models; therefore, the study will only examine up to the 60-h forecast time.

c. Synoptic regimes

To determine general weather regimes, where the superensemble forecast performs better or worse, upper-level and surface maps were examined over the forecast season. Three broad categories where created to group significant synoptic conditions together. The days with multiple consecutive occurrences of a single dominant synoptic feature were chosen in each category. Three categories were identified to dominate the study area as subtropical ridges, midlatitude cyclones, and tropical cyclones. Figure 7 contains a chart depicting the average 10-mm ETS values for the three different synoptic regimes.

Ten days were found to demonstrate a significant subtropical ridge in the model domain, which strongly influenced the isolated precipitation pattern. As shown in Fig. 7, the ETS for this subcategory is 0.045 for the 10-mm threshold. This low value demonstrates the difficulty in the superensemble forecast in capturing the precipitation in the model domain while under the synoptic forcing of a subtropical ridge where rainfall is dominated by isolated convection, which is not well captured by the models. Fourteen days were found to have significant synoptic forcing due to the presence of a midlatitude cyclone. The skill scores are relatively inconsistent with widely ranging ETS values. The average 10-mm ETS value is roughly equivalent to the entire forecast period average.

The most impacted days were those that demonstrated a tropical cyclone forcing. Days were selected when a storm designated as a tropical depression, tropical storm, or hurricane was located in the model domain. However, a short forecast period limits the availability of tropical cyclone case days. The skill scores are notably higher for the TC forcing days than the previous two synoptic subcategories. Despite the fact that there are relatively few available forecasts that may affect the reliability of this comparison, the ability of the superensemble technique to enhance the member models forecasts is impressive.

An improved forecast was made for 30 June 2003, when Tropical Storm Bill was at its peak intensity and made landfall over Louisiana. To aid in the visual inspection of the characteristics of these forecasts, Fig. 8 contains the observations and the 24-h forecasts from the superensemble and all six member models. The forecast captures the general structure of Tropical Storm Bill, especially the asymmetrical region of the heaviest region of precipitation. However, the forecast does underestimate the intense precipitation and overestimates the aerial coverage of the light precipitation. The member models that forecast a realistic amount of heavy precipitation also significantly overestimate the lightest precipitation threshold. These same member models significantly underestimate the heaviest precipitation. None of the member models outperforms the superensemble. The superensemble tends toward improving the member model forecasts by reducing the overestimation of light rain and adding a significant rain forecast that most of the member models miss. A single member model forecast tends to either overestimate (i.e., too wet) or underestimate (i.e., too dry) both the light and heavy rain. The superensemble technique combines the most accurate components of each member model into a single improved forecast.

Over 2 and 3 July 2003, Tropical Storm Bill weakened and became an extratropical low over the United States. In Fig. 9 the comparison between the observations (left panels) and the superensemble forecast (right panels) for the 12–48-h forecast is shown. The superensemble tends to overestimate the light to midlevels of the precipitation thresholds. This is evident by noting the locations of the small-scale precipitation convective cells in the 12-h observation panel and the more homogeneous rainfall field of the superensemble. However, the superensemble does pinpoint the intense precipitation over North Carolina. These relative skills are surprisingly evident in Fig. 10, which shows the average ETS for the 12- and 24-h forecasts from the superensemble, simple ensembles, and member models for all tropical cyclone days. The ETSs are highest at the lightest and also at the heaviest rainfall thresholds. The lowest ETSs for the superensemble are surprisingly found at the midrange of the precipitation thresholds of 10–25 mm. The superensemble seemingly pinpoints the locations of the most intense precipitation, which improves the ETS values.

6. Summary and conclusions

This study has evaluated the short-range mesoscale superensemble precipitation forecasts for the southeast United States over the warm season. Real-time operational and research mesoscale model forecasts were utilized to create the first mesoscale superensemble forecast in an effort to better predict summertime rainfall. Five independent operational and research mesoscale models were gathered from the top NWP facilities across the United States: NCEP, NCAR, NRL, FSL, and NSSL. These multimodels serve as the foundation for the needed variations in initial conditions and model physics in an effort to capture the spread in possible precipitation forecasts. The study also began with the generation of four additional precipitation forecasts from various rain-rate retrieval algorithms used in the PI process of the FSUNRSM.

Different compilations of member models were assessed to determine the best set of member models for the most accurate superensemble forecast. A regional mesoscale superensemble forecast was then created through the compilation of the mesoscale model forecasts. After evaluation of the ETS with different member model sets, six member models were found to create the best superensemble forecast: operational Eta, Eta with KF, NCAR MM5, FSL RUC, NRL COAMPS, and FSUNRSM with TRMM PI. Four member models that included only variations of the FSUNRSM were found to have very low skill scores. This suggests that a spectrum of model physics is needed for a robust superensemble forecast. Various training period lengths were examined to find the optimal number of training days for the most accurate superensemble forecast. Sensitivity tests reveal that 45 days of training maximizes the superensemble forecast performance.

Precipitation forecasts were evaluated through deterministic verification techniques including the equitable threat score and false alarm ratio. The superensemble forecast has the highest ETS over all member model and simple ensemble forecasts. The superensemble forecast removes excess precipitation in its forecast, which dramatically improves the FAR. In general, member models tend to overestimate the lowest precipitation thresholds and underestimate the highest thresholds. The superensemble forecast minimizes these error trends.

To evaluate the skill of the superensemble forecast under different weather regimes, the forecast period was separated into three different categories. In general, forecast skills were lowest under the influence of a subtropical ridge. Midlatitude trough/front forcings slightly improved skill scores above the season average. The most improved synoptic regime was the tropical cyclone. When all the member models agree on a center of convective activity (i.e., a tropical cyclone), the superensemble maximizes its skill scores by removing excess precipitation and pinpointing the convective center.

In general, the member model forecasts are unable to outperform the superensemble. In fact, the superensemble tends toward improving the member model forecasts by significantly reducing the overestimation of light rain and by pinpointing the significant rain forecast. A single member model forecast tends to either overestimate (i.e., too wet) or underestimate (i.e., too dry) the precipitation. The superensemble technique combines the most accurate components of each member model into a single improved forecast. If no member model predicts rainfall at a grid point, then the superensemble is unable to predict any precipitation at that grid point.

Several possibilities exist for future extensions of this work. The mesoscale regional superensemble should be run in an operational real-time application over the United States to evaluate its performance over multiple seasons. Other variables, including surface temperature and winds, should be evaluated to determine the benefits of the superensemble technique. Other superensemble techniques should be utilized. The current technique has the disadvantage of poor handling of singularities in the error covariance matrices. Other sophisticated methods like singular value decomposition (SVD) and empirical orthogonal functions (EOFs) may improve the superensemble solution and generate a better forecast. Probabilistic scoring should be evaluated to determine the impact of the superensemble technique upon probabilistic precipitation forecasts. Other verification techniques, including object-oriented verification, should be used on the precipitation forecasts to evaluate the nature of the superensemble forecast’s improvement.

Overall, the regional mesoscale superensemble forecast was shown to be the most skillful forecast when compared to the state-of-the-art operational mesoscale models. This technique has the potential of increasing forecast accuracy for a multitude of economic applications that rely on accurate rainfall prediction.

Acknowledgments

The authors of this work wish to thank the following operational centers for their mesoscale model forecasts: NCEP, NSSL, FSL, NRL, and NCAR. Several individuals from various agencies were key in acquiring the necessary model and observation data: Mike Baldwin from NSSL, Jason Nachamkin from NRL, Stephen Weygrandt from FSL, and Jim Bresch at NCAR/MMM. Observation data were provided by Jason Nachamkin from NRL and David Bolvin from GSFC. This work was funded by NSF Grant ATM-0108741 and NASA TRMM Grant NAG5-13563.

REFERENCES

  • Black, T. L., 1994: The new NMC mesoscale Eta Model: Description and forecast examples. Wea. Forecasting, 9 , 265278.

  • Buizza, R., Hollingsworth A. , Lalaurette F. , and Ghelli A. , 1999a: Probabilistic predictions of precipitation using the ECMWF Ensemble Prediction System. Wea. Forecasting, 14 , 168189.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Ebert, E. E., 2001: Ability of a poor man’s ensemble to predict the probability and distribution of precipitation. Mon. Wea. Rev., 129 , 24612480.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Ferraro, R. R., and Marks G. F. , 1995: The development of SSM/I rain-rate retrieval algorithms using ground-based radar measurements. J. Atmos. Oceanic Technol., 12 , 755770.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Huffman, G. J., Adler R. F. , Bolvin D. T. , and Nelkin E. J. , 2003: Estimating uncertainty in GPCP and TRMM multi-satellite precipitation estimates. IUGG XXIII: State of the Planet: Frontiers and Challenges, Sapporo, Japan, Int. Union of Geodesy and Geophysics, CD-ROM, JSM18.

  • Kain, J. S., and Fritsch J. M. , 1992: The role of the convection “trigger function” in numerical forecasts of mesoscale convective systems. Meteor. Atmos. Phys., 49 , 93106.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Katsanos, D., Lagouvardos K. , Kotroni V. , and Huffmann G. J. , 2004: Statistical evaluation of MPA-RT high-resolution precipitation estimates from satellite platforms over the central and eastern Mediterranean. Geophys. Res. Lett., 31 .L06116, doi:10.1029/2003GL019142.

    • Search Google Scholar
    • Export Citation
  • Krishnamurti, T. N., Kishtawal C. M. , Shin D. W. , and Williford C. E. , 2000a: Improving tropical precipitation forecasts from a multianalysis superensemble. J. Climate, 13 , 42174227.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Krishnamurti, T. N., and Coauthors, 2000b: Real time multianalysis/multimodel superensemble forecasts of precipitation using TRMM and SSM/I products. FSU Rep. 00-8, 57 pp.

  • Kummerow, C., Olson W. S. , and Giglio L. , 1998: A simplified scheme for obtaining precipitation and vertical hydrometeor profiles from passive microwave sensors. IEEE Trans. Geosci. Remote Sens., 34 , 12131232.

    • Search Google Scholar
    • Export Citation
  • Olson, W. S., Fontaine F. S. , Smith W. L. , and Achtor R. H. , 1990: Recommended algorithms for the retrieval of rainfall rates in the tropics using SSM/I (DMSP-F8). Manuscript, University of Wisconsin—Madison, 10 pp. [Available from Dept. of Atmospheric and Oceanic Sciences, University of Wisconsin—Madison, 1225 W. Dayton St., Madison, WI 53706.].

  • Stensrud, D. J., and Fritsch J. M. , 1994a: Mesoscale convective systems in weakly forced large-scale environments. Part II: Generation of a mesoscale initial condition. Mon. Wea. Rev., 122 , 20682083.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Stensrud, D. J., and Fritsch J. M. , 1994b: Mesoscale convective systems in weakly forced large-scale environments. Part III: Numerical simulations and implications for operational forecasting. Mon. Wea. Rev., 122 , 20842104.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Stensrud, D. J., Bao J. W. , and Warner T. T. , 2000: Using initial condition and model physics perturbations in short-range ensemble simulations of mesoscale convective systems. Mon. Wea. Rev., 128 , 20772107.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Toth, Z., and Kalnay E. , 1993: Ensemble forecasting at NMC: The generation of perturbations. Bull. Amer. Meteor. Soc., 74 , 23172330.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Wilks, D. S., 1995: Statistical Methods in the Atmospheric Sciences. Academic Press, 467 pp.

Fig. 1.
Fig. 1.

Flowchart diagram describing the process for generating a superensemble forecast. The vertical dotted line denotes the time t = 0, the area to the left denotes the training phase where a large number of forecast experiments are carried out, and the area to the right denotes the forecast phase.

Citation: Weather and Forecasting 22, 4; 10.1175/WAF1023.1

Fig. 2.
Fig. 2.

Comparison of the ETSs for the 24-h forecasts with 30 (30 TRN), 45 (45 TRN), and 59 (59 TRN) days of training.

Citation: Weather and Forecasting 22, 4; 10.1175/WAF1023.1

Fig. 3.
Fig. 3.

The (a) ETSs and (b) TSSs for the 12-h superensemble forecasts from two different suites of member models: six diverse members and four FSUNRSM members (four similar members).

Citation: Weather and Forecasting 22, 4; 10.1175/WAF1023.1

Fig. 4.
Fig. 4.

The ETSs for the (a) 12-, (b) 24-, (c) 36-, and (d) 48-h forecasts from the SUP, EMN, BCE, and MEM1–3.

Citation: Weather and Forecasting 22, 4; 10.1175/WAF1023.1

Fig. 5.
Fig. 5.

The FARs for the (a) 12-, (b) 24-, (c) 36-, and (d) 48-h forecasts from the SUP, EMN, BCE, and MEM1–3.

Citation: Weather and Forecasting 22, 4; 10.1175/WAF1023.1

Fig. 6.
Fig. 6.

The (a) ETSs and (b) FARs for 60-h forecasts from the SUP, EMN, BCE, and MEM1–3.

Citation: Weather and Forecasting 22, 4; 10.1175/WAF1023.1

Fig. 7.
Fig. 7.

Comparison between average ETSs for the 10-mm threshold over three different synoptic regimes: subtropical ridge, midlatitude trough, and tropical cyclone.

Citation: Weather and Forecasting 22, 4; 10.1175/WAF1023.1

Fig. 8.
Fig. 8.

Comparison between the (a) observations, (b) superensemble, (c) operational Eta, (d) Eta-KF, (e) NCAR MM5, (f) FSUNRSM with TRMM PI, (g) RUC, and (h) COAMPS forecasts (mm) for 24-h forecasts valid at 1200 UTC 30 Jun 2003.

Citation: Weather and Forecasting 22, 4; 10.1175/WAF1023.1

Fig. 9.
Fig. 9.

Comparison between the (a),(c),(e),(g) observations and (b),(d),(f),(h) superensemble forecasts (mm) for the (a),(b) 12-, (c),(d) 24-, (e),(f) 36-, and (g),(h) 48-h forecasts for 2 Jul 2003.

Citation: Weather and Forecasting 22, 4; 10.1175/WAF1023.1

Fig. 10.
Fig. 10.

The ETSs for the (a) 12- and (b) 24-h forecasts for the five tropical cyclone days for SUP, EMN, BCE, and MEM1–3.

Citation: Weather and Forecasting 22, 4; 10.1175/WAF1023.1

Table 1.

Acronyms incorporated in this study.

Table 1.
Table 2.

NWP models that were available for superensemble forecasts.

Table 2.
Save
  • Black, T. L., 1994: The new NMC mesoscale Eta Model: Description and forecast examples. Wea. Forecasting, 9 , 265278.

  • Buizza, R., Hollingsworth A. , Lalaurette F. , and Ghelli A. , 1999a: Probabilistic predictions of precipitation using the ECMWF Ensemble Prediction System. Wea. Forecasting, 14 , 168189.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Ebert, E. E., 2001: Ability of a poor man’s ensemble to predict the probability and distribution of precipitation. Mon. Wea. Rev., 129 , 24612480.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Ferraro, R. R., and Marks G. F. , 1995: The development of SSM/I rain-rate retrieval algorithms using ground-based radar measurements. J. Atmos. Oceanic Technol., 12 , 755770.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Huffman, G. J., Adler R. F. , Bolvin D. T. , and Nelkin E. J. , 2003: Estimating uncertainty in GPCP and TRMM multi-satellite precipitation estimates. IUGG XXIII: State of the Planet: Frontiers and Challenges, Sapporo, Japan, Int. Union of Geodesy and Geophysics, CD-ROM, JSM18.

  • Kain, J. S., and Fritsch J. M. , 1992: The role of the convection “trigger function” in numerical forecasts of mesoscale convective systems. Meteor. Atmos. Phys., 49 , 93106.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Katsanos, D., Lagouvardos K. , Kotroni V. , and Huffmann G. J. , 2004: Statistical evaluation of MPA-RT high-resolution precipitation estimates from satellite platforms over the central and eastern Mediterranean. Geophys. Res. Lett., 31 .L06116, doi:10.1029/2003GL019142.

    • Search Google Scholar
    • Export Citation
  • Krishnamurti, T. N., Kishtawal C. M. , Shin D. W. , and Williford C. E. , 2000a: Improving tropical precipitation forecasts from a multianalysis superensemble. J. Climate, 13 , 42174227.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Krishnamurti, T. N., and Coauthors, 2000b: Real time multianalysis/multimodel superensemble forecasts of precipitation using TRMM and SSM/I products. FSU Rep. 00-8, 57 pp.

  • Kummerow, C., Olson W. S. , and Giglio L. , 1998: A simplified scheme for obtaining precipitation and vertical hydrometeor profiles from passive microwave sensors. IEEE Trans. Geosci. Remote Sens., 34 , 12131232.

    • Search Google Scholar
    • Export Citation
  • Olson, W. S., Fontaine F. S. , Smith W. L. , and Achtor R. H. , 1990: Recommended algorithms for the retrieval of rainfall rates in the tropics using SSM/I (DMSP-F8). Manuscript, University of Wisconsin—Madison, 10 pp. [Available from Dept. of Atmospheric and Oceanic Sciences, University of Wisconsin—Madison, 1225 W. Dayton St., Madison, WI 53706.].

  • Stensrud, D. J., and Fritsch J. M. , 1994a: Mesoscale convective systems in weakly forced large-scale environments. Part II: Generation of a mesoscale initial condition. Mon. Wea. Rev., 122 , 20682083.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Stensrud, D. J., and Fritsch J. M. , 1994b: Mesoscale convective systems in weakly forced large-scale environments. Part III: Numerical simulations and implications for operational forecasting. Mon. Wea. Rev., 122 , 20842104.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Stensrud, D. J., Bao J. W. , and Warner T. T. , 2000: Using initial condition and model physics perturbations in short-range ensemble simulations of mesoscale convective systems. Mon. Wea. Rev., 128 , 20772107.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Toth, Z., and Kalnay E. , 1993: Ensemble forecasting at NMC: The generation of perturbations. Bull. Amer. Meteor. Soc., 74 , 23172330.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Wilks, D. S., 1995: Statistical Methods in the Atmospheric Sciences. Academic Press, 467 pp.

  • Fig. 1.

    Flowchart diagram describing the process for generating a superensemble forecast. The vertical dotted line denotes the time t = 0, the area to the left denotes the training phase where a large number of forecast experiments are carried out, and the area to the right denotes the forecast phase.

  • Fig. 2.

    Comparison of the ETSs for the 24-h forecasts with 30 (30 TRN), 45 (45 TRN), and 59 (59 TRN) days of training.

  • Fig. 3.

    The (a) ETSs and (b) TSSs for the 12-h superensemble forecasts from two different suites of member models: six diverse members and four FSUNRSM members (four similar members).

  • Fig. 4.

    The ETSs for the (a) 12-, (b) 24-, (c) 36-, and (d) 48-h forecasts from the SUP, EMN, BCE, and MEM1–3.

  • Fig. 5.

    The FARs for the (a) 12-, (b) 24-, (c) 36-, and (d) 48-h forecasts from the SUP, EMN, BCE, and MEM1–3.

  • Fig. 6.

    The (a) ETSs and (b) FARs for 60-h forecasts from the SUP, EMN, BCE, and MEM1–3.

  • Fig. 7.

    Comparison between average ETSs for the 10-mm threshold over three different synoptic regimes: subtropical ridge, midlatitude trough, and tropical cyclone.

  • Fig. 8.

    Comparison between the (a) observations, (b) superensemble, (c) operational Eta, (d) Eta-KF, (e) NCAR MM5, (f) FSUNRSM with TRMM PI, (g) RUC, and (h) COAMPS forecasts (mm) for 24-h forecasts valid at 1200 UTC 30 Jun 2003.

  • Fig. 9.

    Comparison between the (a),(c),(e),(g) observations and (b),(d),(f),(h) superensemble forecasts (mm) for the (a),(b) 12-, (c),(d) 24-, (e),(f) 36-, and (g),(h) 48-h forecasts for 2 Jul 2003.

  • Fig. 10.

    The ETSs for the (a) 12- and (b) 24-h forecasts for the five tropical cyclone days for SUP, EMN, BCE, and MEM1–3.

All Time Past Year Past 30 Days
Abstract Views 0 0 0
Full Text Views 331 97 9
PDF Downloads 83 24 2