The operational numerical weather prediction (NWP) systems at the Japan Meteorological Agency (JMA) indicated that the typhoon track forecasts made by the control member of the ensemble prediction system (EPS) tended to be worse than those made by the high-resolution global NWP. The control forecast of the EPS with horizontal triangular truncation at 319 wavenumbers and 60 vertical levels (T319/L60 resolution) was initialized by eliminating the higher-wavenumber components of the global analysis at T959/L60 resolution. When the data assimilation cycle was performed at the lower T319/L60 resolution, the forecast gave typhoon track forecasts closer to the high-resolution global NWP. Therefore, it stands to reason that the resolution transform of the initial condition must be responsible for the degradation of the typhoon track forecasts at least to considerable extent. To improve the low-resolution forecast, two approaches are tested in this study: 1) applying a smoother spectral truncation for the resolution transform and 2) performing noncycled lower-resolution data assimilation during preprocessing. Results from the single case study of Typhoon Nuri (2008) indicate almost no impact from the former approach, but a significant positive impact when using the latter approach. The results of this study illuminate the importance of considering a model’s resolving capability during data assimilation. Namely, if the initial conditions contain features caused by unresolved scales, degraded forecasts may result.
The Japan Meteorological Agency (JMA) began to operate its global NWP system at a T959/L60 resolution in November 2007 (Nakagawa 2009) and its typhoon ensemble prediction system (TEPS) in 2008 (Yamaguchi and Komori 2009; Yamaguchi et al. 2009). Typically, the EPS has a lower resolution than the global NWP system, which is true for the TEPS at a T319/L60 resolution. The initial conditions for the control forecast of TEPS are generated by simply eliminating higher-wavenumber components of the high-resolution global analyses. Five ensemble perturbations are generated by the singular vector (SV) method (e.g., Buizza 1994), and the plus–minus pairs are added to the control initial conditions. Thus, the ensemble mean of the 11-member initial conditions equals the set of control initial conditions.
Operations of the high-resolution global NWP and the lower-resolution TEPS in 2008 yielded an interesting result in that the typhoon track forecasts by the TEPS control member tended to have larger errors than the high-resolution global forecasts. Moreover, the 11-member ensemble mean of TEPS ensemble forecast tracks, which was generally better than the single control member, was still worse than the single high-resolution forecasts. The initial conditions were essentially identical; the higher-wavenumber components were simply eliminated. The forecast degradation may have been due to the differences among the forecast models. Namely, the T959/L60 model may have been more accurate than the T319/L60 model. However, the preoperational investigations in 2007 indicated essentially identical performance between the high- and low-resolution models in the case where the initial conditions are generated by each independent data assimilation cycle at each resolution. Therefore, we expect that the forecast degradation was caused by the resolution transformation of the initial conditions. This idea agrees with a commonly accepted hypothesis that the initial conditions are suitable to the model with which the data assimilation cycle is performed.
This was our first experience at JMA with clear forecast degradation due to the resolution transform of the initial conditions using high-wavenumber elimination. C. Reynolds (2010, personal communication) had a similar experience with the Navy Operational Global Atmospheric Prediction System (NOGAPS) at the Naval Research Laboratory (NRL), and R. Buizza (2010, personal communication) mentioned the experience of the European Centre for Medium-Range Weather Forecasts (ECMWF) with their EPS in terms of the importance of resolution change from the high-resolution analysis. However, to the best of the authors’ knowledge, no study published thus far has brought to light the problem of the degradation of tropical cyclone (TC) track forecasts. Although it is not easy to identify the exact mechanism for track degradation, it would be beneficial to report the facts observed in the operational systems, which may motivate further scientific investigations in the wider research community.
Many studies have been published on the initialization of TCs, including those discussing JMA’s prior developments (e.g., Iwasaki et al. 1987; Ueno 1989; Ueno and Onogi 1991; Ueno 1995). Most of the previous studies focused mainly on TC bogusing and repositioning methods to better initialize TC vortices that are compatible with a model’s resolution when only relatively poor analyses are available due to the general lack of observations around TCs over the ocean. For example, Kurihara et al. (1993) proposed a method of initializing a high-resolution TC forecast using a low-resolution analysis by replacing a crudely resolved TC in the large-scale analysis with a high-resolution vortex. Many operational centers have developed and investigated TC bogusing and repositioning methods: for example, the National Centers for Environmental Prediction (NCEP; Lord 1991; Liu et al. 2000), NRL (Goerss and Jeffries 1994), the Met Office (Heming et al. 1995), and ECMWF [never used in operations; Serrano and Unden (1994)]. More recent studies include that of Wu et al. (2006), who assimilated bogus data into a mesoscale model and found that wind data played an important role in TC prediction due to the geostrophic adjustment of smaller scales.
The focus of this study is different from those of the previous studies. Global NWP began to be useful for TC forecasts only about a decade ago (e.g., Elsberry 1995), and NCEP started the operation of the first EPS around that time (Toth and Kalnay 1993). Because the global EPS began to increase the level of resolution and became potentially useful for TC forecasts only very recently, the need to initialize TCs in a lower-resolution EPS using a higher-resolution analysis has emerged very recently. This has not been addressed in the literature to the best of the authors’ knowledge.
In this note, we report the diagnostics obtained by the operations of high-resolution global NWP and TEPS in 2008. In addition, we propose and test two approaches to improve the low-resolution forecasts at a minimal additional cost. Section 2 describes the error statistics of the 2008 operational typhoon predictions, in which the skill of the high-resolution global NWP exceeded that of the TEPS control forecast. In section 3 we test two possible approaches to improve the low-resolution forecast using a case study of Typhoon Nuri (the 12th typhoon in 2008). Finally, a summary and discussion are provided in section 4. For simplicity, we define the short names of the operational forecasts and our experiments in Table 1. Details of the experiments will be described later in section 3. Following the Japanese operational convention, in this note the term typhoon includes tropical cyclones in the northwestern Pacific with maximum wind speed of greater than 34 kt (17.5 m s−1).
2. Statistical fact in 2008 typhoon forecast
In daily weather briefings to monitor typhoons in 2008, the developers of TEPS noticed that TEPS_CTL tended to have larger typhoon track prediction errors than T959_OPE. This has been proven correct by a statistical investigation of forecast track errors for all typhoons that occurred in 2008, where T959_OPE has significantly smaller track errors by at least 30 km for forecast lead times longer than or equal to 36 h (Fig. 1).
A possible explanation of this statistical fact is the difference between the forecast models; that is, the T959 model provides a more accurate prediction than the T319 model. However, the preoperational experiments of TEPS in 2007, whose initial conditions were provided by a T319 data assimilation cycle, did not indicate such degradation of typhoon track forecasts. This motivated a T319 data assimilation cycle experiment in 2008, which is denoted as T319_CYC in Table 1. The typhoon track errors of T319_CYC are compared with those of TEPS_CTL and T959_OPE for all typhoons in 2008. The results indicate that T319_CYC significantly outperforms TEPS_CTL, although it does slightly worse than T959_OPE. Therefore, the resolution transform from the T959 initial conditions to T319 must be responsible for the forecast degradation, at least to some extent.
3. Case study of Typhoon Nuri (2008)
The NWP forecast for Typhoon Nuri (2008), initialized at 1200 UTC 19 August 2008, is a typical case indicating that T959_OPE outperforms TEPS_CTL. The best track, analyzed by the Regional Specialized Meteorological Center (RSMC) Tokyo-Typhoon Center, is west-northwestward, which is better captured by T959_OPE (Fig. 2). TEPS_CTL indicates a stronger northward tendency, which would lead to a false alert for Taiwan. The T319_CYC experiment provides a significantly better track forecast than TEPS_CTL, although it is slightly worse than T959_OPE. This agrees with the statistics shown in section 2.
In the present case of Typhoon Nuri analyzed at 1200 UTC 19 August 2008, essentially no observation was available in the vicinity of the typhoon. Therefore, the preprocessing of the operational global data assimilation system, being aware of the existence of the typhoon from the real-time Dvorak analysis of RSMC Tokyo Typhoon Center, generated bogus data for assimilation around the typhoon. The bogus data are generated based on the Dvorak analysis and model’s grid spacing, retaining the asymmetric typhoon structure in the first guess (Ueno 1995). In the present case, no notable analysis increment was found around Typhoon Nuri; that is, the location and strength of the typhoon in the first guess were very close to those of the bogus typhoon.
Although operating an independent data assimilation cycle for T319_CYC would improve the TEPS significantly, we aim to avoid this approach due to the high computational cost. Here, we introduce two alternative methods for improving the T319 forecast at minimal additional cost when the T959 analysis cycle preexists. The first approach, called T319_TS1 in Table 1, is to apply a smooth spectral truncation in transforming the T959 analysis to the T319 resolution. In TEPS_CTL, zero power is forced to the T959 analysis for wavenumbers greater than 319, but all power remains for wavenumbers smaller than 320 (Fig. 3a). A smoother spectral truncation is chosen, in which the power goes to zero at 320 wavenumbers linearly from 106 wavenumbers (Fig. 3b). The results indicate that T319_TS1 has almost no impact on the track forecast (Fig. 2), although the initial conditions become generally smoother in the global domain.
The second approach, called T319_TS2, is to perform a “noncycled” T319 data assimilation, as illustrated in Fig. 4. Here, the T319 data assimilation is not cycled, since it is initiated from the existing T959 analysis for the high-resolution deterministic NWP forecast. The T959 analysis is degraded to the T319 resolution by simply eliminating the high-wavenumber components in the same way as TEPS_CTL. Then, a T319 forecast is made to generate the first guess for the data assimilation at T319 resolution. Without the cycle process, performing data assimilation once is much less expensive since it is computed only when typhoons exist and TEPS is activated. The results indicate that T319_TS2 significantly outperforms TEPS_CTL and is almost identical to T319_CYC for the initial 36 h (Fig. 2).
The major difference in the initial conditions between TEPS_CTL and T319_CYC is found near the typhoon center. The central pressure is deepest for T959_OPE and shallowest for T319_CYC (Fig. 5). TEPS_CTL and T319_TS1 are relatively similar to T959_OPE, whereas T319_TS2 is closer to T319_CYC. We find basically the same tendency in the wind field. Obviously, the T959 model generated the sharper typhoon structure than the T319 model. After the resolution degradation, TEPS_CTL captured the sharp structure near the typhoon center, as did T319_TS1 with the smoother spectral truncation. Namely, the T959 model generated the typhoon core structure that could be represented well by up to about 150 wavenumbers, so that the smoother spectral truncation did not change it significantly. However, T319_CYC generated the shallower typhoon structure, which was a natural consequence of the T319 model. We expect that the T319 model does not have the sharp typhoon structure on its solution attractor, even though it can technically be represented by the T319 resolution. Since T319_TS2’s first guess was generated by the T319 model, the typhoon structure was spun up to some extent to provide a shallower structure, closer to that of T319_CYC. Hence, it was essential to generate the first guess using the same model as the forecast model. This single case study indicates the potential of the two approaches, although the level of statistical performance over many cases is not yet clear.
4. Summary and discussion
The T319/L60 control forecasts of the operational TEPS in 2008 indicated larger typhoon track errors than those of the operational T959/L60 global forecasts. The T319/L60 initial conditions were generated by simply eliminating higher-wavenumber components of the operational T959/L60 analysis. Since the independent T319/L60 data assimilation cycle improved the T319/L60 typhoon track forecasts significantly, the spectral truncation of the resolution change must be responsible for the forecast degradation, at least to some extent.
Two approaches for improving the T319/L60 forecasts without keeping the independent data assimilation cycle were proposed in this study: 1) applying a smoother spectral truncation for the resolution transform and 2) performing noncycled lower-resolution data assimilation during preprocessing. According to the case study of Typhoon Nuri (2008), the former approach had essentially no impact, whereas the latter approach improved the track forecast significantly.
In general, it is very difficult to identify what differences in the initial conditions would have caused the forecast track difference. Although we saw in Fig. 5 that the shallower typhoon existed on the solution attractor of the T319 model and was preferable for lower-resolution forecasts, the exact source of the forecast improvement remains unknown.
Applying the smoother spectral truncation in this study apparently made the global fields smoother in general, but it did not affect Typhoon Nuri’s core structure significantly. The resolution degradation could be optimized to yield a shallower typhoon, which might help improve the forecasts. However, such a strong smoothing would affect the global fields significantly, and the optimal resolution degradation may depend on each typhoon case. In a practical global EPS, it would be prohibitive to optimize the resolution degradation for many kinds of phenomena adaptively. Studying such sophisticated resolution degradation methods is beyond the scope of this study.
The results of this study illuminate the importance of considering a model’s resolving capability in data assimilation. Namely, if the initial conditions contain features caused by unresolved scales, they may degrade the forecast. For example, the T319 resolution (not the model processes) could technically resolve the deep typhoon structure in Fig. 5 (b), but this deep structure was generated by the T959 model. This suggests that the best initial conditions for a given model should be generated by data assimilation using the same model at the same resolution as the forecast system.
Since many operational NWP centers operate EPSs with a lower-resolution model initialized by a higher-resolution analysis, the issue raised by this study may be important for improving operational EPSs in general. Although this study focused on only typhoon track forecasts, the conclusion may also be valid for other high-impact weather phenomena that are sensitive to a model’s resolving capability. When the deterministic NWP system adopts an even higher-resolution model and the resolution difference from the EPS becomes larger, which was actually the case at JMA, the issue raised by this study may become more important.
In many studies, it is common to use operational analyses to initialize research models for real case studies. However, it is worthwhile to keep in mind that the forecasts could be significantly degraded simply because the analysis is generated by data assimilation at a different resolution, sometimes even with a different model. This suggests the possible utility of data assimilation methods without using operational analysis.
The authors thank three anonymous reviewers for their constructive comments, which improved the paper. The authors are grateful to W. Hogsett of the National Hurricane Center, C. Reynolds of NRL, R. Buizza of ECMWF, and D. Kleist of NCEP for their useful comments. This research was partly supported by the Office of Naval Research (ONR Grant N000141010149) under the National Oceanographic Partnership Program (NOPP).
Corresponding author address: Takemasa Miyoshi, Dept. of Atmospheric and Oceanic Science, University of Maryland, College Park, College Park, MD 20742. Email: email@example.com