Geophysical Research Letters

Tropical cyclone track forecasts using JMA model with ECMWF and JMA initial conditions

Authors

  • Munehiko Yamaguchi,

    Corresponding author
    1. Meteorological Research Institute of the Japan Meteorological Agency, Tsukuba, Japan
      Corresponding Author: M. Yamaguchi, Meteorological Research Institute of the Japan Meteorological Agency, 1-1 Nagamine, Tsukuba, Ibaraki 305-0052, Japan. (myamagu@mri-jma.go.jp)
    Search for more papers by this author
  • Tetsuo Nakazawa,

    1. Meteorological Research Institute of the Japan Meteorological Agency, Tsukuba, Japan
    2. World Meteorological Organization, Geneva, Switzerland
    Search for more papers by this author
  • Kazumasa Aonashi

    1. Meteorological Research Institute of the Japan Meteorological Agency, Tsukuba, Japan
    Search for more papers by this author

Corresponding Author: M. Yamaguchi, Meteorological Research Institute of the Japan Meteorological Agency, 1-1 Nagamine, Tsukuba, Ibaraki 305-0052, Japan. (myamagu@mri-jma.go.jp)

Abstract

[1] The JMA's Global Spectral Model (JMA/GSM) was run from the initial conditions of ECMWF, which are available in the YOTC data set, to distinguish between TC track prediction errors attributable to the initial conditions and those attributable to the NWP model. The average position error was reduced by about 10% by replacing the initial conditions, and in some cases, the predictions were significantly improved. In these cases, the low wavenumber component of the ECMWF analysis was found to account for most of the improvement. In addition, the observed tracks were captured by the JMA Typhoon Ensemble Prediction System (TEPS), which deals with initial condition uncertainties. In some cases, however, the replacement of the initial conditions did not improve the prediction even when the ECMWF forecast was accurate. In these cases, TEPS could not capture the observed track either, implying the need for dealing with uncertainties associated with the NWP model.

1. Introduction

[2] The accuracy of tropical cyclone (TC) track prediction has improved greatly over the last few decades, in large part because of improvements in numerical weather prediction (NWP) models, data assimilation schemes, and enhanced observations obtained by satellites and aircraft [e.g., Elsberry, 2007; Heming and Goerss, 2010]. According to Chan [2010], since Chan et al. [2002]paper, research on the physics of general TC motion has been almost non-existent, which suggests that most scientists are quite content with the current theories of TC motion. In reality, however, significant prediction errors still exist and there are prediction cases where the position error can exceed 1000 km over 3 days. Although the theories of TC motion might have reached a satisfactory level, our knowledge on the source of prediction errors is still poor [Carr and Elsberry, 2000a, 2000b].

[3] In general, it is difficult to distinguish between prediction errors attributable to the initial conditions from those attributable to the NWP model, although recent research projects such as The Observing System Research and Predictability Experiment (THORPEX) have made it possible to separate these two causes to some extent. This separation is achieved by running one NWP model from the initial conditions of another NWP system with higher prediction performance. The initial conditions are thought to be essential for accurate predictions in cases where the prediction is significantly improved by replacement of the original initial conditions. Meanwhile, we may gain insight into modifications that will improve an NWP model by analyzing those cases where replacing the initial conditions with those of another NWP system yielding accurate predictions does not reduce the prediction error in the assessed model.

[4] Figure 1 shows the verification results of TC track predictions in the western North Pacific from 2008 to 2010 using the track predictions of the unperturbed control member of the ensemble prediction system (EPS) in the THORPEX Interactive Grand Global Ensemble (TIGGE) [Bougeault et al., 2010] data set. The position errors of the European Centre for Medium-range Weather Forecasts (ECMWF) are the smallest among the nine major NWP centers for prediction periods of 1 to 5 days. For example, the 5-day position error of ECMWF is comparable to the 4-day position errors of the U.K. Met Office (UKMO) and the Japan Meteorological Agency (JMA). This raises the following questions: would it be possible to reduce the position errors of the NWP models by running them from the ECMWF initial conditions and by how much? Although the answer could prove very valuable for improving TC track predictability, these questions have yet to be addressed. Besides the separation of the two error sources, the initial conditions and the NWP model, is helpful from a standpoint of developing the NWP system since developers of initial conditions can avoid spending much of their time in trying to improve prediction cases where the NWP model is considered as a major error source, and vice versa.

Figure 1.

Position errors of 1- to 5-day TC track predictions of the unperturbed EPS control member from the Bureau of Meteorology in Australia (BOM), the Chinese Meteorological Administration (CMA), the Canadian Meteorological Centre (CMC), the Center for Weather Forecasts and Climate Studies in Brazil (CPTEC), the European Centre for Medium-Range Weather Forecasts (ECMWF), the Japan Meteorological Agency (JMA), the Korean Meteorological Administration (KMA), the National Center for Environmental Prediction in the U.S. (NCEP), and the U.K. Met Office (UKMO). The analysis verified 58 TCs that occurred in the western North Pacific between 2008 and 2010.

[5] In this study, the JMA's global spectral model (GSM) was run from the ECMWF initial conditions to answer the questions above, and individual prediction cases were analyzed to separate TC track prediction errors associated with the initial conditions from those inherent in the NWP model. For this purpose, we considered 16 TCs that evolved in the western North Pacific basin from August to November 2009. The source of the ECMWF's analysis data is the Year of Tropical Convection (YOTC) data set.

[6] This paper is organized as follows. Section 2 describes the data used in this study and the method used to create the initial conditions for JMA/GSM from the ECMWF analysis data. Section 3 describes the results of the numerical experiments, and section 4 discusses the results. Section 5 presents the conclusions of this study.

2. Data and Methodology

[7] In this study, we used the JMA's global forecasting system, which consists of the 4-dimensional variational (4DVAR) data assimilation scheme and the JMA/GSM [Japan Meteorological Agency, 2007]. The resolution of the JMA/GSM used in this study was TL319L60, although that of the operational system is TL959L60. First, the JMA/GSM was run from the initial conditions created during a 6-hourly data assimilation cycle of the 4DVAR to obtain the reference track predictions (hereafter, the JMA model and JMA initial conditions: JM-JI). The predictions were initiated at 1200 UTC only. Second, the JMA/GSM was run from the ECMWF initial conditions (hereafter, the JMA model and ECMWF initial conditions: JM-EI). We used the following method to create initial conditions for the JMA/GSM from the ECMWF analysis data. We downloaded the ECMWF analysis data with a horizontal resolution of 0.5625° (equivalent to TL319) from the YOTC database. JMA/GSM uses 60 vertical layers up to 0.1 hPa, but only 25 vertical levels from 1000 hPa to 1 hPa are available in the YOTC data set. Therefore, we adopted a linear interpolation technique in the vertical direction to create the initial conditions from the YOTC data set. The predictions of the ECMWF model and its initial conditions (hereafter, EM-EI) were obtained from the TIGGE database, which provides the forecast fields of the ECMWF high-resolution deterministic forecasting system as well as those of the EPS.

[8] We verified 16 TCs that occurred in the western North Pacific basin from August to November 2009 using the best-track data analyzed by the Regional Specialized Meteorological Center (RSMC) Tokyo–Typhoon Center. Only TCs that were of tropical storm or stronger intensity at the initial time were selected for verification. TCs of tropical depression intensity at the initial time were not included in the verification, but if the TC intensity was reduced to tropical depression status during the prediction period (up to 5 days), the prediction was verified, including the times when the TC was categorized as a tropical depression.

3. Results

[9] The verification results of the TC track position errors from 0 to 5 days by JM-JI, EM-EI, and JM-EI (Figure 2) show that the difference between JM-JI and EM-EI is similar to that seen in the verification of TCs over 3 years (Figure 1); that is, EM-EI position errors were smaller than the JM-JI errors with a lead time of one day. The position errors of JMA/GSM were reduced by replacing the JMA/GSM initial conditions with those of ECMWF (JM-EI inFigure 2). The improvement rate of JM-EI with respect to JM-JI was 5%, 11%, 9%, 11%, and 15% on days 1 to 5, respectively. Thus, the replacement of the initial conditions accounts for 20%, 29%, 29%, 38%, and 68% of the difference between JM-JI and EM-EI on days 1 to 5, respectively. The difference between JM-JI and JM-EI was statistically significant at the 90% level on days 4 and 5.

Figure 2.

Position errors (left vertical axis) from 0 to 5 days of JM-JI, EM-EI, and JM-EI, and the number of samples (right vertical axis).

[10] Comparison of observed TC tracks and the tracks predicted by ECMWF and JMA/GSM showed that TC track prediction was significantly improved by the initial condition replacement (Figures 3a and 3b). For Typhoon Dujuan (Figure 3a), the recurvature of the JM-JI track occurred earlier than the observed track or the EM-EI track recurvature, and the TC's subsequent movement with the westerly jet stream was faster, resulting in a position error of 595 km on day 3. The replacement of the initial conditions improved both the timing of the recurvature and the subsequent movement speed, reducing the error to 122 km. For Typhoon Lupit (Figure 3b), JM-JI failed to predict the recurvature and predicted that Lupit would make landfall in the Philippines, whereas EM-EI successfully predicted the recurvature, though the track showed a slow bias after the recurvature. After replacement of the initial conditions in JMA/GSM, the model was able to predict the recurvature of Lupit, thus reducing the position error from 720 km to 280 km.

Figure 3.

Track predictions by JM-JI, EM-EI, JM-EI, and JM-EI2 and the observed track for (a) Typhoon Dujuan, initiated at 1200 UTC on 5 September 2009, and (b) Typhoon Lupit, initiated at 1200 UTC on 21 October 2009. (c and d) The TEPS track predictions for Dujuan and Lupit, respectively, for the same initial times. The triangles are plotted along each track every 24 hours.

[11] Additional experiments were conducted in which the initial conditions were created by blending the low-wavenumber component (≤T42, ∼300 km) of the ECMWF analysis with the higher one of the JMA analysis (JM-EI2). The horizontal resolution of T42 was equal to that of the ensemble initial perturbations in the ECMWF EPS. Even this change in initial conditions improved the track (Figures 3a and 3b, orange lines), which indicates that the representation of the steering flow formed by the synoptic environment around the TC is important for accurate TC track prediction, as demonstrated in previous studies [e.g., Chan and Gray, 1982]. It also shows the validity of adopting low-resolution singular vectors to create the ensemble initial perturbations in the ECMWF and JMA EPSs, which have horizontal resolutions of T42 and T63, respectively.

[12] Figures 3c and 3d show the ensemble TC track predictions by the JMA Typhoon EPS (TEPS) [Yamaguchi et al., 2009] for Dujuan and Lupit. TEPS has 11 members, and the ensemble initial perturbations were created by the singular vector method. The uncertainties in the NWP model were not considered in TEPS. The uncertainties in the timing of the recurvature of Dujuan were well represented by the ensemble members, and TEPS succeeded in predicting that the representation of the westerly jet might cause the position error to become large in the along-track direction. In the case of Lupit, some ensemble members were successful in predicting the recurvature of Lupit though the unperturbed control member was not able to predict the recurvature like the JM-JI prediction. These two ensemble cases imply that TEPS can successfully express the uncertainties of the TC track predictions when they are sensitive to initial conditions.

[13] Figures 4a and 4bshow two examples where replacement of the initial conditions did not improve TC track prediction even though the EM-EI position error was small. Typhoon Morakot (Figure 4a) made landfall in Taiwan, where it caused torrential rainfall and catastrophic damage. Typhoon Parma (Figure 4b) made landfall in the Philippines and added to the damage caused by Typhoon Ketsana, which had struck the previous week. In both cases JM-JI showed a northward bias and failed to predict the landfalls. Moreover, the northward bias remained even after replacement of the initial conditions with the ECMWF initial conditions. The insensitivity of these predictions to changes in the initial conditions indicates that modifications of JMA/GSM would be needed to predict the observed tracks more accurately.

Figure 4.

Track predictions by JM-JI, EM-EI, and JM-EI and the observed track for (a) Typhoon Morakot, initiated at 1200 UTC on 4 August 2009, and for (b) Typhoon Parma, initiated at 1200 UTC on 30 September 2009. (c and d) The TEPS track predictions by for Morakot and Parma, respectively, for the same initial times. The triangles are plotted along each track every 24 hours.

[14] In the TEPS predictions of the tracks of Morakot and Parma (Figures 4c and 4d, respectively), all ensemble members showed a northward bias and failed to predict the observed landfall. In addition, the ensemble spread was relatively small, which might lead the user to mistakenly infer a small prediction error. For these two TCs, addressing only the initial condition uncertainties was not sufficient to capture the observed track, and for better probabilistic predictions, a method for addressing uncertainties associated with the NWP model would be needed.

[15] The northward bias is not unique to TEPS; the same bias is seen in ensemble predictions of other major NWP centers (Figures S1 and S2 in the auxiliary material). Moreover, the northward bias tends to appear to the east of the Philippines. To reduce TC track prediction errors and to improve the accuracy of probabilistic forecasting, the cause of this bias must be identified and the NWP systems, including the EPSs, must be modified to correct the bias.

4. Discussions

[16] Although both ECMWF and JMA use 4DVAR for their data assimilation scheme, the differences of their initial conditions are large enough to yield prediction cases as have been shown in Figure 3. One of the causes of the differences may be observational data and the way the data are treated in the 4DVAR. Enhanced use of all-sky microwave observations can be considered superiority of ECMWF over JMA [Bauer et al., 2010, 2011]. In addition, the NWP models are thought to be a cause of the differences of the initial conditions since they are created by blending observations and the best-estimate of the atmosphere named the first guess, which is a short-range, say six-hour, forecasts by the NWP models. Given that TCs are in the data-sparse oceanic environment, the differences of the NWP models between ECMWF and JMA can be one of the major sources producing the differences of the initial conditions.

[17] For the prediction cases shown in Figure 4, we cannot eliminate the possibility that the initial conditions affect the results because the YOTC data set only has 25 vertical levels, far less than the ECMWF or JMA models. Moreover, there is a possibility that the boundary conditions such as the sea surface temperature may affect the results. Thus it is not entirely clear that our setting of the numerical experiments can accurately separate between TC track prediction errors attributable to the initial conditions and those attributable to the NWP model. Hopefully we will run the ECMWF model with the JMA initial conditions in the future in order to better understand the initial condition uncertainty and the model error.

[18] In what conditions does the initial uncertainty dominate the model error? This is the question naturally arising from this study. Unfortunately we do not have clear answers to this question. However we would guess from the results of the additional experiments shown in Figure 3, that the representation of the steering flow at initial time would be an important factor for accurate predictions. In addition, the initial location of TCs may play an important role in better TC track predictions as studied by Kurihara et al. [1993] and Hsiao et al. [2010] though the impact of the vortex relocation technique was not investigated in this study. Meanwhile, the model error would be dominant for cases where the position errors tend to appear in a particular direction and the replacing the initial conditions does not change the trend of the errors. Figure S3shows the position errors in the along-track and cross-track directions for 3-day predictions. The different numbers shown in the plots mean the different directions of the movement of the observed TC tracks. The number 7 in red is for typhoon Morakot and Parma, which moved toward the west-northwest to the east of Taiwan and the Philippines, respectively. AsFigure S3shows, the errors tend to be positive in the cross-track direction in JM-JI, and the error trend does not change even when the initial conditions were replaced with ECMWF. Thus it can be inferred that the errors come from the model and the error trend is a bias associated with the model.

5. Conclusions

[19] JMA's global spectral model (JMA/GSM) was run using the ECMWF initial conditions to assess the sensitivity of TC track prediction to changes in the initial conditions and to distinguish between prediction errors attributable to the initial conditions from those attributable to the NWP model. Replacing the original initial conditions of JMA/GSM with those of the ECMWF analysis reduced the TC track prediction errors by 5%, 11%, 9%, 11%, and 15% on days 1 to 5, respectively, explaining 20%, 29%, 29%, 38%, and 68% of the difference in the errors on days 1 to 5, respectively, between the JMA and ECMWF results.

[20] For some TCs, replacement of the initial conditions significantly improved track prediction, and the low wavenumber component (≤T42, ∼300 km) of the ECMWF analysis played a major role in the improved predictions. This result shows the importance of the representation of the synoptic environment, which controls the steering flow of TCs. Whereas the original TC track predictions by JMA/GSM with JMA's initial conditions had large prediction errors, TEPS, which uses the singular vector method to address initial condition uncertainties, captured the observed tracks. This result implies that TEPS successfully expressed the uncertainties of TC track predictions that were sensitive to the initial conditions.

[21] 'For other TCs, however, replacement of the initial conditions did not improve track prediction, even though the ECMWF prediction was accurate. In some cases a northward bias caused a large position error and TEPS could not capture the observed track either. It implies that no matter how the initial conditions are perturbed, the observed track cannot be captured when the prediction error is attributed to the NWP model. Through analysis of the TIGGE data set we found that a northward bias, especially east of the Philippines, is a problem not only at JMA but also at other major NWP centers. Identifying the cause of this bias and modifying NWP systems including EPSs will improve the accuracy of TC track prediction from both deterministic and probabilistic perspectives.

[22] NWP models still make significant prediction errors and the causes of these prediction errors are poorly understood. The TIGGE database provides EPS analysis and forecast data from major NWP centers. The YOTC data set provides not only analysis and forecast data from the ECMWF high-resolution NWP system but also physical parameters such as temperature changes associated with both shallow and deep convection. Making the best possible use of such data sets should lead to further improvements in TC track prediction from both deterministic and probabilistic perspectives.

Acknowledgments

[23] The authors thank two anonymous reviewers for thoughtful reviews and constructive comments. The authors thank the people constructing and maintaining the portals of the Year of Tropical Convection (YOTC) data set and The Observing System Research and Predictability Experiment (THORPEX) Interactive Grand Global Ensemble (TIGGE) for making them useful and user-friendly and providing the analysis and forecast data of the operational NWP centers. We also thank Carolyn Reynolds of the Naval Research Laboratory for reviewing an early version of the paper and Hitoshi Yonehara, Masayuki Kyouda, Akira Shimokoube, and Takuya Komori of the Numerical Prediction Division at JMA for their assistance in creating the JMA/GSM initial conditions from the ECMWF analysis.

[24] The Editor thanks the two anonymous reviewers for assisting with the evaluation of this paper.

Ancillary