Possible Sources of Forecast Errors Generated by the Global/Regional Assimilation and Prediction System for Landfalling Tropical Cyclones. Part I: Initial Uncertainties

Tropical cyclones (TCs), especially those that make landfall, have long been recognized as a cause of considerable damage and disruption in coastal areas (Webster et al., 2005; Roy and Kovordanyi, 2012). Forecasting the tracks of these storms has been a longstanding concern for meteorologists. Over recent decades, ongoing research has led to significant improvements in the forecasting of TC tracks by NWP models (Rappaport et al., 2009; Heming and Goerss, 2010). However, large track-forecast errors still occur (Ryan et al., 2015), and there have been instances where the forecast errors have exceeded 1000 km over 3 days (Yamaguchi et al., 2012). The impact of such large errors is exacerbated if they are associated with landfalling TCs. To improve disaster prevention and mitigation measures, more accurate TC track forecasts are required, especially for landfalling TCs. As the current theory of TC motion is mature and widely accepted (Chan, 2010), error diagnosis has become an important aspect of attempts to further improve TC forecasts (Galarneau and Davis, 2013).

Forecast errors derive from both uncertainties in the state of the atmosphere at the beginning of the forecast period and model imperfections, and many studies have examined the impacts of these factors on TC forecasts (Yamaguchi and Majumdar, 2010; Brennan and Majumdar, 2011; Kepert, 2011; Green and Zhang, 2014; Galarneau and Hamill, 2015; Ryan et al., 2015). Generally, both the initial and model uncertainties have significant impacts on TC forecasts. However, in a real forecast the forecast errors usually result simultaneously from both the model and initial uncertainties. It is important to consider whether we can identify the forecast errors derived from initial errors, and those derived from the model errors. If this is possible, then we can reduce the uncertainty associated with the largest sources of error.

Given the above, it is important to consider how best to distinguish the contributions of the initial errors from those of the model errors. This task is generally difficult, despite previous efforts. For example, (Orrell et al., 2001) and (Orrell, 2003) compared the evolution of real forecast errors with the evolution of an approximate accumulated model error to determine the impact of the model errors. They pointed out that at the start of the forecast the evolution of the model errors is similar to the evolution of the forecast errors, and therefore the forecast errors result mainly from the model errors. However, as the simulation proceeds, the approximate accumulated model error begins to underestimate the forecast error. A similar experiment was carried out by (Liu et al., 2010) to evaluate error development in the Lorenz-96 system. (Nicolis et al., 2009) classified the leading timescales for initial errors and model errors by comparing their development rates, and one of the key points in their approach was determining the crossover time at which the contribution of the initial condition errors matches that of the model errors under the assumptions of small uncorrelated initial errors and small parameter errors. (Yu et al., 2012) used the conditional nonlinear optimal perturbations of initials (CNOP-I) and parameters (CNOP-P) to identify the source of the spring predictability barrier (SPB) for El Niño events. They found that the CNOP-P errors do not cause a significant SPB for El Niño events, but that the CNOP-I errors generate a significant SPB. On the basis that model systems contain many parameters, and that it is difficult to consider all of these parameters, (Duan and Zhou, 2013) proposed a nonlinear forcing singular vector (NFSV) approach to evaluate the effect of tendency errors on predications, and the NFSV was found to be another important source of the El Niño prediction error (Duan and Zhao, 2015). A similar method was proposed by (Xu et al., 2004), who applied a nonlinear optimization method to identify whether the initial errors or the model errors are dominant. However, this method requires an adjoint model and an optimization algorithm, and is limited by the development of the adjoint model. (Yamaguchi et al., 2012) distinguished TC track prediction errors attributable to the initial conditions ("initials", hereafter) and those attributable to the NWP model by replacing the initials of the JMA Global Spectral Model with those of the ECMWF. They found that by doing so, the average position error was reduced by about 10%. In some cases the predictions were significantly improved, whereas in others the replacement of the initials did not improve the prediction even when the ECMWF forecast was accurate, thereby indicating the need to deal with the uncertainties associated with the NWP model. However, (Yamaguchi et al., 2012) used only a single NWP model, and they did not further reveal the source of the initial errors, or the source of the model errors.

A new generation of the national operational NWP system, known as the Global/Regional Assimilation and Prediction System (GRAPES) Global Forecast System (GRAPES_ GFS), was developed by the China Meteorological Administration (CMA) to enhance meteorological services in China in the 21st century (Zhang and Shen, 2008). The GRAPES_GFS model has developed rapidly over recent years (Huang et al., 2013, 2014), and now generates reliable forecasts for the subtropical high (Chen et al., 2008). In addition, the anomaly correlation coefficients calculated based on the 500-hPa geopotential height between GRAPES_GFS hindcasts and NCEP analysis were found to be high (Zhang and Shen, 2008, Fig. 2), and it was shown that GRAPES_GFS has an effective forecast time of about 6 days. This indicates that GRAPES_GFS may be capable of forecasting the tracks of TCs well.

In this paper, we examine the forecast ability of the GRAPES_GFS model (hereafter, GRAPES) for landfalling TCs, and identify the source of the forecast errors associated with these TCs. To isolate the two main factors that determine the forecast errors (i.e., the initial uncertainty and the model uncertainty), the present study is divided into two parts. In part I, we distinguish between the forecast errors attributable to the initial errors and those attributable to the model errors by using a similar approach to (Yamaguchi et al., 2012). Then, we focus on the TCs for which the forecast errors were caused mainly by the initial uncertainty, and determine the source of the initial errors. In part II (to be presented in a separate paper), we focus on those forecast errors that resulted mainly from model uncertainty, and we determine the source of the model errors.

The remainder of this paper is organized as follows. Section 2 introduces the methodology, including the model designs and the approach used to distinguish the forecast errors attributable to the initial errors from those attributable to the model errors. Section 3 classifies the TC cases according to the contributions of the initial errors and model errors. Section 4 reveals the source of the initial errors by analyzing the initial uncertainties and the results of sensitivity experiments conducted to identify the important variables and areas in the initials for accurate forecasts of landfalling TCs using the GRAPES model. Finally, discussion and conclusions are presented in section 5.

The GRAPES model has a horizontal resolution of 1^°× 1^° and 36 vertical levels. The physical parameterizations used in this analysis were as follows: the WSM6 scheme for microphysics (Hong and Lim, 2006); a modification of RRTM (Rapid Radiative Transfer Model) scheme named RRTMG for long- and shortwave radiation (Iacono et al., 2008), the common land model scheme for land surface processes (Dai et al., 2003); the MRF scheme for the PBL (Hong and Pan, 1996); and the simple Arakawa-Schubert scheme for cumulus clouds (Han and Pan, 2011). The default initials, and boundary conditions, were supplied by the NCEP FNL Operational Global Analysis (1^°× 1^°) interpolated into the GRAPES forecast system. Herein, the forecasts made by the GRAPES model with its default initials are denoted as GRAPES-FNL.

According to (Yamaguchi et al., 2012), among nine forecast centers, the TC tracks over the period 2008-10 were forecast most accurately by the ECMWF for the forecast period of 1 to 5 days (see their Fig. 1). Consequently, we assumed that the model and initials used by ECMWF were both accurate, and therefore used the ECMWF initials as the candidate initials for GRAPES. The initials used by the ECMWF model can be found in the Year of Tropical Convection (YOTC) dataset. YOTC was a program launched in May 2008 by the WCRP and the World Weather Research Programme/THORPEX. Its purpose is to organize coordinated observations, modeling and forecasting of organized tropical convection, with the goal to better understand its multiscale structure and interactions (see http://www.wmo.int/pages/prog/arep/wwrp/new/yotc.html for further details of the project). The source of the ECMWF analysis data can be found from YOTC (Yamaguchi et al., 2012). We downloaded the YOTC data with a horizontal resolution of 1^°, equivalent to that of GRAPES. Herein, the forecasts made by GRAPES using the ECMWF initials are denoted as GRAPES-EI. For comparison, the forecasts of the ECMWF model with its initials (denoted as ECMWF-EI) were also downloaded from the YOTC dataset with a resolution of 1^°. The track and intensity observations used to test the accuracy of the model forecasts were derived from the TC best-track dataset of the Shanghai Typhoon Institute of the CMA and were downloaded from the typhoon website of the Chinese government (www.typhoon.gov.cn; Ying et al., 2014).

Sixteen TCs that occurred in the Northwest Pacific and made landfall between 2008 and 2009 are studied. In each case the forecast period was 72 h ahead of landfall (Table 1). Some TCs were short-lived, and the initials used in these forecasts were for tropical depressions rather than tropical storm intensity.

Although the ECMWF model usually produces accurate forecasts, it is still necessary to examine its performance with respect to the storms under consideration here and at the chosen resolution. First, the forecasts made by GRAPES-FNL were compared with those made by ECMWF-EI, and the average forecast error of the 16 TCs obtained from ECMWF-EI was the smallest of all (Fig. 1). Figure 2 shows that the 72 h forecasts of the TC tracks from ECMWF-EI are better than those from GRAPES-FNL for all 16 TCs. In most cases, the accumulated track errors generated by ECMWF-EI are significantly smaller than those from GRAPES-FNL. Next, comparisons were made between the GRAPES-EI and GRAPES-FNL runs, which showed that the forecasts for 12 of the 16 TCs were improved in the GRAPES-EI experiment compared with GRAPES-FNL (Fig. 2). As both the ECMWF model and its initials are generally accurate, we concluded that the poor forecasts of the remaining four TCs by GRAPES were caused by model errors. However, additional analysis of the other 12 cases is required before a decision can be made on whether their poor forecasts in GRAPES-FNL were caused by initial uncertainties in the FNL data or by uncertainties within the GRAPES model. To do this, we defined an index to measure the improvement of the forecasts, as follows:

Figure 1.  Average forecast errors (km) for the 16 TCs every 6 h. Blue line: forecast by GRAPES with default initials; red line: forecast by ECMWF with its initials; green line: forecast by GRAPES but with ECMWF initials.

Figure 2.  Accumulated track errors (km; accumulated every 6 h over the analysis period of 72 h) for each TC from the GRAPES model with its default initials (GRAPES-FNL, blue bars) and with the ECMWF initials (GRAPES-EI, green bars), and from the ECMWF model with its initials (ECMWF-EI, red bars). The red box indicates the TCs in the much-improved group; the green box indicates the TCs in the unimproved group; and the remainder are the little-improved group.

\begin{equation} \!{IM}\!=\!\!\dfrac{(\!E_{GRAPES\hbox{-}FNL}\!\!-\!E_{ECMWF\hbox{-}EI}\!)\!-\!(\!E_{GRAPES\hbox{-}EI}\!-\!E_{ECMWF\hbox{-}EI}\!)}{E_{GRAPES\hbox{-}FNL}\!\!-\!E_{ECMWF\hbox{-}EI}}, (1)\end{equation} where E represents the departure of the forecast from the observations (i.e., the 72 h accumulated track errors calculated every 6 hours) and the subscripts indicate the various forecast errors.

If IM is greater than or equal to 0.5, the improvement of the initials is highly effective, as it would make the forecasts much closer to the observations. In this case, we would conclude that the initial uncertainties are the main source of the forecast error; and consequently, the initials should be improved first. If IM is less than 0.5, the improvement of the initials is less effective, and in this case the model uncertainty would be taken as the main source of the forecast uncertainty, and improvement of the model would be necessary.

Table 1 shows that only four TCs had large IM values, whereas eight had small IM values. According to the above definition, the four TCs with large IMs are those for which forecast errors resulted from initial uncertainty, and we classified these as TCs for which the initials should be improved first (the initial-first group). The remaining eight TCs with small IMs were defined as TCs for which the model should be improved first (the model-first group). Therefore, there are a total of twelve TCs for which the model is unable to forecast the TC landfalls, and we should therefore improve the model first. It is noted that the IM value for Morakot (2009) is 79.9, which is two orders of magnitude larger than the other eleven IM values. This is due to the large improvement for Morakot (2009) when using ECMWF initials. The forecast of GRAPES-EI is now better than the forecast of ECMWF-EI in the Morakot (2009) case. As we focus on the impact of the initial uncertainty associated with the TC landfall forecasts made by the GRAPES model, we concentrate on the initial-first TCs in the remainder of this paper. However, the initials of the model-first TCs will also be analyzed for comparison. To facilitate these comparisons, the initial-first TCs were renamed as the "much-improved" group, and the model-first TCs were divided into two subgroups as follows: The eight TCs with small IMs were classified as the "little-improved" group, and the other four, which showed no improvements when using the ECMWF initials, were defined as the "unimproved" group.

Figure 3.  Initial position errors when using the NCEP FNL dataset (the default initials of GRAPES) versus using the ECMWF YOTC dataset. The red box indicates the TCs in the much-improved group, and the remainder are the little-improved group.

Figure 4.  Differences in initial values between the NCEP FNL dataset (the default initials of GRAPES) and the ECMWF YOTC dataset. The first row shows the differences in SLP (hPa); the second row shows the differences in geopotential height at 500 hPa (m); the third row shows the differences in the zonal wind (m s$^-1$) at 850 hPa; the fourth row shows the differences in specific humidity (g g$^-1$) at 850 hPa; and the fifth row shows the differences in temperature (K) at 700 hPa. The first column shows data for TC Hagupit (2008) (an example of the much-improved group); the second column shows data for TC Nuri (2008) (little-improved group); and the third column shows data for TC Fengshen (2008) (unimproved group).

In this section, the difference between the ECMWF initials and the default initials is examined for the three TC groups defined in section 3 as we try to answer the following questions: First, what will be improved in the initials when using ECMWF initials? Second, what are the differences among the initials in the three TC groups? In other words, what type of TC forecasts can be significantly improved by improving the initials? Analysis was carried out from the perspectives of the positions, intensity, and environmental features of the TCs.

First, for those TCs whose forecasts have been improved when using ECMWF initials, we analyzed what has been improved in the initials. As it was pointed out by (Yamaguchi et al., 2012) that the initial positions of TCs may have an important impact on their track forecasts, we first examined the positions of the TCs at the initial time. Figure 3 shows that the initial position errors for three cases are larger using ECMWF initials than the default initials, while for the other cases they are comparable between the two types of initials. This indicates that the improvements in the TC track forecasts are not necessarily related to the improvement in the initial position of the TC.

The initial intensity of the TC is another factor to which the track forecast is sensitive (Galarneau and Hamill, 2015). Thus, we focus next on the intensity of the TCs. The differences in SLP between the default initials and ECMWF initials are obvious at the TC center (Fig. 4), and the SLP from the default initials is higher than that from the ECMWF initials for all TCs, regardless of the degree of improvement. Consequently, the initial TCs simulated using the GRAPES default settings have lower intensities than those simulated by the ECMWF model. The average of the differences between the central pressures based on the default initials and those based on the ECMWF initials is positive for all three TC groups (Fig. 5). As the central pressure generated by the ECMWF initials is closer to the observations than that based on the default initials of GRAPES (Table 2), the initial intensity of the TCs simulated using GRAPES-FNL is lower than observed. Consequently, replacing the default initials with the ECMWF initials strengthens the TC intensity and improves the accuracy of TC intensity at the initial time. These conclusions are supported by our analysis of the geopotential height at 500 hPa and the zonal wind at 850 hPa (Fig. 4; second and third columns, respectively). In this case the differences between the default and the ECMWF scenarios is positive in the geopotential heights at the TC center. It is positive in the zonal wind on the north side and negative in the zonal wind on the south side of the cyclone for each TC.

The environmental field is without doubt an important factor that influences the track of a TC. For all three TC groups, the geopotential height at 500 hPa with the GRAPES default initials was lower than that with the ECMWF initials around the subtropical high and monsoon trough, but was higher around the South Asia high (Fig. 4; second column). This indicates that, compared with those generated by the ECMWF initials, the subtropical high generated by the GRAPES initials has weaker intensities, while the South Asia high and the monsoon trough have stronger intensities. However, if we take an overall view, the difference in the TC center is the most noticeable. Although the differences in temperature and humidity between the default initials and ECMWF initials are not significant, we still find systematic differences between them. That is, the temperature generated by the GRAPES default initials was higher than that for the ECMWF initials in the ocean area, but lower than that for the ECMWF initials on land (Fig. 4; fifth column). Thus, in the TC center the temperature was slightly higher with the GRAPES default initials. However, the specific humidity was lower in the TC center but higher in the surroundings with the GRAPES default initials compared with the ECMWF initials (Fig. 4; fourth column).

Figure 5.  (a) Average accumulated track forecast errors (km; accumulated every 6 h over the analysis period of 72 h) for each group of TCs from the GRAPES-FNL results. (b) Average initial differences between the central SLP (hPa) from the NCEP FNL data and the ECMWF YOTC data.

Next we consider the nature of the improvements in the initials when using the ECMWF data. Generally, the intensity of the TC at the initial time is improved and strengthened when using the ECMWF dataset, but the initial positions of the TCs are not. The subtropical high is strengthened, but the South Asia high and monsoon trough are weakened when using the ECMWF data. Also, the temperature at the TC center is lower, whereas the specific humidity is higher, when using the ECMWF initials compared with the GRAPES default initials. The total effect of these factors improves the TC landfall forecasts. However, the question remains as to which factor is the most important. This is considered in section 4.2.

It is important to consider which type of TC forecast can be most improved using the ECMWF initials. Figures 2 and 5 show that the much-improved group has relatively small forecast errors compared with the other two groups. Consequently, to achieve a significant improvement, the model should be able to predict the TC; i.e., the model error should not be too large. The differences of TC intensities between the default initials and the ECMWF initials for the much-improved group are larger than those for the other two groups (Fig. 5), and this indicates that the much-improved group has a sizeable uncertainty associated with the TC intensity. Another interesting point is that three of the four much-improved cases were in the intensifying phase and had a weaker initial intensity, whereas three of the four unimproved cases were decaying and were generally strong at the initial time (Fig. 6 and Table 2). This means that the model physics may catch up with the intensification phase of the cyclone, but fail to describe the decaying phase. Therefore, for intensifying TCs the forecasts are greatly improved when the initials are improved, but for decaying TCs the forecasts cannot be improved simply by improving the initials. An investigation of the environmental fields revealed little of interest in this context.

In summary, given that TC forecasts made with the GRAPES model can be greatly improved by improving the initials, the TCs are expected to have a large uncertainty in initial intensity, and the TCs are in the intensifying phase with a relatively weak initial intensity.

In section 4.1 we demonstrated systematic differences between the GRAPES default initials and ECMWF initials, such as the intensities of the TCs, subtropical high, South Asia high, and monsoon trough. A complete replacement of the default initials with the ECMWF initials as the model input variables (wind, geopotential height, temperature, and relative humidity; note that the boundary conditions were kept the same as the default) improves the forecasts for the studied TCs. Next, we consider which variables should be replaced to yield the greatest improvement in forecasts, and which location is most effective (insofar as the replacement of initials there would obtain the greatest improvement). In other words, we assess which system should be improved at the initial time to generate the greatest improvement in TC forecasts. To address these questions we examined the impacts on the forecasts of replacing each variable, as well as the impact of the replacements in different locations. For this purpose, the four cases of the initial-first group are studied.

Figure 6.  Observed SLP (hPa) for the (a) much-improved and (b) unimproved TCs during the study period.

Figure 7.  Difference in geopotential height (m) at the 500-hPa level between the NCEP FNL and ECMWF YOTC datasets. The boxes indicate areas where the data are replaced. The yellow boxes indicate the area in and around the TC center. The four TCs (indicated by the titles above each panel) are from the much-improved group. (a) 2008 Hagupit; (b) 2008 Higos; (c) 2009 Ketsana; (d) 2009 Morakot.

Figure 8.  Accumulated forecast errors (km) for the typhoon tracks predicted by the GRAPES model with its default initials (NCEP FNL data) and the GRAPES model with the ECMWF initials partly replacing the default initials. The labels h-TC, uv-TC, rh-TC, and T-TC indicate that the geopotential height, wind field, relative humidity, and temperature, respectively, of the default initials have been replaced by the ECMWF initials around the TC central area. The labels huv-TC, huv-subhigh, huv-westTC, huv-southTC, and huv-plateau indicate that the geopotential height and wind field were replaced at the center, northeast, west, south, or southwest of the TC, and the plateau area, respectively. The four TCs (indicated by the titles above each panel) are from the much-improved group.

As described in section 4.1, the wind, geopotential height, temperature, and relative humidity at the TC center generated by the GRAPES default initials show systematic departures from those generated by the ECMWF initials. First, we respectively replaced the four variables in the default initials with the corresponding variables from the ECMWF initials in and around the TC center (yellow rectangles in Fig. 7), and ran GRAPES with the newly produced initials that combined both the default and ECMWF data. The track forecast errors of the TCs were then compared to determine which variable around the TC center is the most important in terms of improving the track forecast.

Figure 8 shows that the track forecast errors were significantly reduced by improving the geopotential height and wind fields surrounding the TC. However, the improvements in the relative humidity and temperature surrounding the TC failed to obviously reduce the forecast errors (Fig. 8). This result further demonstrates that an accurate estimation of TC intensity at the initial time is important to improve the forecasts. The importance of the geopotential height seems comparable to that of the wind fields, and both of these factors are reflections of the TC intensity. Therefore, it is necessary to consider whether this is the main factor that drives the improvement in the forecasts.

As indicated previously, the intensities of the subtropical high, South Asia high, and monsoon trough were changed when using the ECMWF initials. Next, we replaced the geopotential height and wind fields of the default initials in the areas where the above systems are located, with those of the ECMWF data in the corresponding areas, and examined whether this replacement improved the forecasts, and to what extent. To this end, another four equal-sized areas were selected for each TC for comparison (Fig. 7). The first location is on the Tibetan Plateau where the South Asia high is located; we refer to this as the plateau area. There are always large differences between the default initials of GRAPES and the ECMWF initials in this region, and other studies have reported that the South Asia high influences TC movements (Chen, 1965; Ding and Wright, 1983). The second area is located in the west side of the subtropical high and to the northeast of the TC. This is also an important region in terms of TC tracks, and with a relatively large difference between the default initials of GRAPES and the ECMWF initials. As the subtropical high influences TCs in this area, we denote this as the subtropical high area. The third area is located on the Indochina Peninsula and to the west of the TCs. Figure 7 shows obvious differences between the GRAPES default initials and the ECMWF initials in this region. This area is usually influenced by the monsoon trough. The fourth area is also influenced by the monsoon and is located to the south or southwest of the TCs, in the path of southwesterly winds. This region is marked by obvious differences between the GRAPES default initials and the ECMWF initials.

Sensitivity experiments were then carried out by respectively replacing the default GRAPES initials with the ECMWF initials in the plateau area, the center of the TCs, and the areas to the northeast, west, and south of the TCs. As mentioned above, the greatest forecast improvement is achieved when both the geopotential height and wind fields are replaced, so these replacements were made simultaneously while keeping the other variables the same as in the default scenario. The results for all cases were similar; that is, the replacements at the centers of the TCs generated the largest forecast improvements (Fig. 8). However, replacing only the default initials with the ECMWF initials in the other locations had little or no effect (Fig. 4). The sensitivity experiments in these areas further confirmed that the improvement in TC intensity at the initial time is important for improvements in the TC track forecasts. Of note is that when the geopotential height and wind fields in and around the TCs are improved, not only the intensity of the TC is improved, but also the steering flows and the structures of the TCs. Many studies have pointed out that steering flows and TC structure have an important influence on TC motion (McTaggart-Cowan et al., 2006; Brennan and Majumdar, 2011; Galarneau and Davis, 2013). This may explain why an improvement in and around the TC center results in an improvement in track forecasts.

Generally, all of the experiments showed that a realistic estimation of TC intensity at the beginning of the forecast is important if the GRAPES model is to generate accurate forecasts of TC landfall tracks.

This paper (part I) and the following paper (part II) focus on determining the source of the forecast errors for TC landfalls made with the GRAPES model. First, the method proposed by (Yamaguchi et al., 2012) was used to distinguish the forecast errors that are attributable to initial errors from those attributable to model errors. Next, attention was paid to cases in which the forecast errors were attributable to the initial errors. An assessment of the differences in the initials between the default forecasts and the improved forecasts revealed systematic differences in the initials. Further comparisons were made between the initials of the TCs whose forecast errors were attributable to the initial errors, and those whose forecast errors were attributable to the model errors, in order to determine the best way to significantly improve the forecasts by improving only the initials. Then, sensitivity experiments were carried out to identify the most important variables and areas at the beginning of the forecasts.

Of the sixteen TCs studied here, all of which made landfall during 2008 and 2009 in the Northwest Pacific, four were identified as TCs whose forecast errors were attributable to the initial errors, and twelve as TCs whose forecast errors were attributable to the model errors. Further examination showed obvious systematic differences between the default initials of the GRAPES model and the ECMWF initials for all the cases. That is, both the TCs and the subtropical high based on the default initials were weaker than the observations and weaker than those based on the ECMWF initials. However, the intensity of the South Asia high and the monsoon trough based on the default initials were stronger than those based on the ECMWF initials. Compared with the ECMWF initials, the default initials in GRAPES had a higher temperature and lower specific humidity at the TC center. The TCs whose forecast errors were attributable to the initial errors had larger initial uncertainties in intensity compared with those attributable to the model errors, and they were usually in the intensifying phase with a relatively weak intensity at the initial time. Sensitivity experiments demonstrated that the accurate estimation of TC intensity at the beginning of the forecast is most important if the GRAPES model is to generate reliable forecasts of the tracks of landfalling TCs.

This study has also confirmed previous findings that the initial values of the wind and pressure fields associated with a TC are important for its track forecast, and not only the subtropical high and the midlatitude trough. In addition, we have demonstrated that when the initial uncertainty (comparing the initials among several NWP centers) is large, the improvement in the initial values may be more beneficial to the forecast than when the initial uncertainty is small. This result indicates that data assimilation may be more useful when the initial uncertainty is large. Generally, the above results indicate when and where the data assimilation is needed and what should be assimilated for good forecasts of landfalling TCs made using the GRAPES model.

Finally, as noted above, twelve cases were identified as TCs whose forecast errors were attributable to model errors. Thus, the GRAPES model has defects in simulating the motion of TCs. In our companion paper (part II), we seek to determine the source of the model errors and suggest how future versions of the model may be improved.

The GRAPES model used in this paper has a very coarse resolution of 1^°× 1^°. As reported by (Fiorino, 2009), based on the ECMWF IFS, simply increasing the model resolution does not lead to significant improvement in the forecast accuracy; rather, large improvement is related more to upgrading the model, including its physics. Thus, we hypothesize that the current conclusion that the model itself plays an important role in improving the forecasts will remain valid even with a high-resolution model of GRAPES. In addition, it is important to note that TC intensity forecasting will continue to be a big challenge for TC research and forecasting communities, even if it is possible to use high-resolution NWP systems with a horizontal resolution of around 2 km.