Accurately determining the rotational center of a tropical cyclone (TC) is of critical importance. The convergence and upward movement of warm and moist air in the central region result in the release of latent heat, which acts as a vital energy source for the development of TCs. Pinpointing the center of both symmetric and asymmetric TC clouds is crucial for analyzing and predicting the movement and changes in TC intensity. Integrating a center position derived from various observations into numerical models has been demonstrated to significantly enhance the skill of TC forecasts (e.g., Kurihara et al. 1993; Chen and Snyder 2007; Hendricks et al. 2011; Kleist 2011; Cha and Wang 2013; Heming 2016; Christophersen et al. 2022). It is noteworthy that previous studies have assumed the typical center error as 20 km in model initialization (e.g., Wu et al. 2010; Kunii 2015), despite the lack of widely accepted consensus in optimal accuracy for the position center. Hsiao et al. (2010) showed a 34% reduction in 72-h TC track forecast errors by correcting the initial TC center with observational data. Furthermore, improved initialization of the TC center led to better forecasts in the TC structure, as demonstrated by Kunii (2015).
Currently, one of the most precise systems for determining the position of the TC center is Automated Rotational Center Hurricane Eye Retrieval (ARCHER), which operates in real time at the Cooperative Institute of Meteorological Satellite Studies (CIMSS) (Olander and Velden 2019). ARCHER utilizes spatial images from both geostationary and polar-orbiting satellites, employing spiral-centering and ring-fitting methods to establish the central location of the TC. The accuracy of this method varies considerably depending on the intensity of the TC, which can range from tropical depressions (TD) to tropical storms (TS), typhoons (TY), and super typhoons (ST) (Wimmers and Velden 2010, 2016). For strong TCs, such as TY and ST, where the eye is discernible and cloud structures around the eyewall are well defined, the error is less than 10 km. Conversely, for weak TCs, such as TD and TS, where an eye is not typically present and cloud organization is lacking, the error increases to over 40 km (Wimmers and Velden 2016). Notably, the largest errors are typically observed in the TD phase, before the system fully develops into a TC.
Furthermore, the error magnitude in determining the TC center depends on the type of satellite observation used. When utilizing polar-orbiting satellite-based microwave imagers, the error ranges from 11 to 35 km, whereas for geostationary satellites, the error ranges from 21 to 56 km (Wimmers and Velden 2016). Additionally, excluding the visible channel and relying solely on infrared (IR) channel(s) from geostationary satellites nearly doubles the error when determining the TC center (Wimmers and Velden 2016). The use of microwave imagers mounted on polar-orbiting satellites, which detect changes in precipitating ice and liquid hydrometeors, enables the observation of the strengthening or weakening of convective activity (Alliss et al. 1992; Lee et al. 2002). However, visible-channel images from geostationary satellites offer an advantage in identifying cloud features in the lower troposphere. Most visible solar radiation fluxes are reflected by thick clouds, making it easier to infer cloud characteristics. In contrast, the IR brightness temperature, which is inversely related to the cloud-top temperature, presents challenges in locating the TC center when the eye is obscured by high clouds or when the center position is vertically displaced by vertical wind shear (Velden et al. 2006; Olander and Velden 2007).
The Korea Meteorological Administration (KMA) currently faces a significant TC center error up to 100 km (Cha et al. 2016; Kim et al. 2022) because polar-orbiting satellite observations are not regularly employed in its operating system. Consequently, there is a pressing need to establish an objective method to reduce such large errors. Therefore, this study aims to develop an algorithm that exclusively utilizes geostationary satellites that are accessible at all times, and addresses the operational constraints faced by the KMA. To achieve this goal, an artificial intelligence (AI) approach was implemented. This AI algorithm was designed to comprehend and analyze nonlinear variations in spatiotemporal domains related to TCs.
In the next section, we introduce the training, validation, and testing data used to develop the AI models. The following sections provide the framework of the AI models and compare the errors obtained from the AI models and satellite channels with those derived from the ARCHER products. The final section presents the conclusions of this study.
Data
JTWC best track.
To train and validate the models, latitude and longitude information of the TC center was obtained from the Joint Typhoon Warning Center (JTWC) best track data. This dataset covers the period from 2016 to 2021 and can be accessed at www.metoc.navy.mil/jtwc/jtwc.html?western-pacific. The JTWC best track data provide reliable, consistent, and authoritative information for the TC center. As the JTWC best track dataset is available at specific times (0000, 0600, 1200, and 1800 UTC), model training and validation were conducted when the corresponding best track data were accessible. This ensured consistency and alignment between the model evaluation and best track data. To evaluate the performance of the models in relation to the TC intensity, four intensity grades were defined based on the 1-min maximum wind speed (υmax) obtained from the JTWC best track data. The intensity grades were TD (υmax < 17 m s−1), TS (17 ≤ υmax < 32 m s−1), TY (32 ≤ υmax < 65 m s−1), and ST (υmax ≥ 65 m s−1). In addition, TDs that did not develop into TSs were included in this study, as they were distinct from other tropical disturbances and waves (Landsea and Franklin 2013).
The JTWC best track data are the final outcomes of a retrospective analysis conducted to identify and document comprehensive information related to a TC throughout its lifespan (Chu et al. 2002). This analysis encompasses TCs that have been officially identified and those that may have gone unnoticed during the operational period (Ying et al. 2014). A combination of various observational sources was used to determine the center location of the TC in this dataset. These sources included geostationary satellite observations, ARCHER operational products derived from geostationary and polar-orbiting satellites, radar data, airborne observations, and land- and ocean-surface observations. We note that the determination of the TC center within the JTWC best track dataset is a subjective process performed by experienced forecasters. These forecasters consider multiple factors, such as the movement of the TC, and assign appropriate weights to different observations based on their reliability (Guard et al. 1992).
Geostationary satellite observation: Himawari-8.
To acquire the necessary imagery for this study, Japanese Himawari-8 satellite images were obtained at 10-min intervals for the period from 2016 to 2021. The images were accessed using the P-Tree system operated by the Earth Observation Research Center of the Japan Aerospace Exploration Agency. Detailed information on the P-Tree system can be found at www.eorc.jaxa.jp/ptree/userguide.html. The Himawari-8 geostationary satellite is positioned at the equator (140.7°E) and provides coverage over a vast area, ranging from 60°S to 60°N and from 80°E to 160°W. This wide coverage makes it suitable for capturing the entire life cycle of TCs over the western North Pacific (WNP). The TC images were obtained using gridded brightness temperatures from four IR channels centered at 8.6, 10.4, 11.2, and 12.4 μm and one shortwave IR (SWIR) channel centered at 3.9 μm, along with gridded reflectivity data from the visible channel centered at 0.64 μm. The spatial resolution of the raw data are 0.5 km × 0.5 km for the visible channel and 2 km × 2 km for the IR and SWIR channels (Bessho et al. 2016). The resolution of each channel was interpolated to 4 km × 4 km to reduce the computational burden during the training of the AI models. The result changed little when different spatial resolutions, such as 2 km × 2 km, were used in the calculations. Brightness temperatures from a single channel and the differences between two channels were obtained as inputs to the AI models, considering their representation of cloud properties (Shimizu 2020). Observations from the 10.4 and 0.64-μm channels were used to calculate cloud top temperature and reflected solar radiation from the cloud and surface, respectively, following the approach suggested by Dvorak (1984). The brightness temperature differences between the 10.4- and 12.4-μm channels (i.e., 10.4 minus 12.4 μm), the 8.6- and 11.2-μm channels (i.e., 8.6 minus 11.2 μm), and the 10.4- and 3.9-μm channels (i.e., 10.4 minus 3.9 μm) were used to calculate cloud optical thickness (Inoue 1987), cloud phase (Ackerman et al. 1990), and cloud type (Lee et al. 1997), respectively. The cloud properties derived from these brightness temperature differences were chosen because they have been used in previous studies analyzing TC positions (Lee 2000; Kim and Hong 2019; Chen et al. 2022).
TC center location from ECMWF 6- and 12-h forecasts.
The predicted location of the TC center from the European Centre for Medium-Range Weather Forecasts (ECMWF) global atmospheric high-resolution model was used as the initial position for the AI model. The ECMWF model has a horizontal resolution of approximately 9 km and comprises 137 vertical levels. The center position forecasts were provided by ECMWF in terms of latitude and longitude at 12-h intervals (0000 and 1200 UTC) based on model predictions (ECMWF 2019). To create the test sets for the AI models, the predicted TC centers from the ECMWF model were obtained for the 6- and 12-h forecasts initialized at 0000 or 1200 UTC. Specifically, the predicted centers were selected for periods corresponding to 0600 and 1200 UTC for 0000 UTC initialization and 1800 and 0000 UTC (on the following day) for 1200 UTC initialization. These predicted TC centers were then assigned as initial guesses when preparing the test sets for the AI models. By incorporating the ECMWF model’s predicted center positions as an initial estimate, the AI models were provided with an informed starting point for their predictions and subsequent evaluations.
Positioning of the TC center from ARCHER operation.
To evaluate the performance of the newly developed AI models, they were compared with the ARCHER system. The operational products of the TC center provided by ARCHER can be accessed online through the CIMSS website at http://tropic.ssec.wisc.edu/real-time/archerOnline/cyclones/. We note that ARCHER operates at 3-h intervals, but it provides TC center information 30 min earlier than the time recorded in the JTWC best track. Consequently, the time points of the ARCHER products that corresponded to the JTWC best track were as follows: 2330 (on the previous day), 0530, 1130, and 1730 UTC. To align the timing of the ARCHER positions with those of the JTWC, linear interpolation was applied, compensating for the 30-min difference. This ensured that the ARCHER products and JTWC best track data were synchronized for comparison and evaluation. When linear interpolation was applied to the JTWC best track instead of the ARCHER, the results scarcely changed (not shown). Both approaches yielded consistent results, indicating the robustness of the evaluation process.
ARCHER continuously updates its operational TC center positions to enhance the accuracy of the center location. Additional polar-orbiting satellite data that may not have been available during the initial operational analysis are often obtained several hours later. ARCHER utilizes this additional data to update the center position by incorporating all available satellite information. In the ARCHER data accessible online, the operational estimate is replaced when the center location is updated. As a result, the ARCHER data used in this study may consist of both operational and updated positions, depending on the specific period analyzed.
ARCHER is primarily designed to position the center of a TC during its operational period, specifically based on the TC classification during that time. However, there may be instances in which a particular event was not initially classified as a TC during the operational period, but was later identified as a TC in the JTWC best track retrospective analysis. Given that the model developed in this study is based on the JTWC TC classification, it can produce center positions for times that may not be available in ARCHER. All analyses conducted in this study were subject to the constraints of the simultaneous availability of JTWC, ARCHER, and our model results. The dataset utilized consisted of a total of 3,647 TC samples with 6-h intervals, encompassing 184 TCs observed from 2016 to 2021.
Development of artificial intelligence models
Two independent AI algorithms were adopted by training them separately using distinct training, validation, and testing datasets. First, a convolutional neural network (CNN), which is a machine learning algorithm designed specifically for image analysis (LeCun et al. 2015), was used. The CNN architecture comprises three main layers: convolutional, pooling, and fully connected (represented by the black, red, and blue boxes, respectively, in Fig. 1). Second, a convolutional long short-term memory (ConvLSTM), another type of machine learning algorithm that combines the principles of both CNN and LSTM networks (Shi et al. 2015), was used. LSTM models excel in processing sequential data. The configuration of the ConvLSTM used is shown in Fig. 2.
CNN configuration, referred to as VGGNet, employed in this study. The algorithm takes an image as input, processes the image through five sets of two convolutional layers (black boxes), one max pooling layer (red boxes), three fully connected layers (blue boxes), and returns the output. The output consists of two components representing the latitudinal and longitudinal differences between the center of the input image and the estimated TC center (blue dots). The ReLU activation function is applied after the convolutional and first two fully connected layers. This function introduces nonlinearity and enhances the network’s capability to learn intricate patterns. The last fully connected layer is followed by the hyperbolic tangent activation function, which further processes the output. The numbers enclosed in parentheses indicate the number of channels and the height and width dimensions of the input image in each layer.
Citation: Bulletin of the American Meteorological Society 105, 3; 10.1175/BAMS-D-23-0130.1
Configuration of the (a) ConvLSTM model, (b) convolutional block, and (c) ConvLSTM cell. (a) The ConvLSTM model takes six sequential images at 10-min intervals (t−5, t−4, t−3, t−2, t−1, and t0) and processes them through the convolutional block (orange boxes), ConvLSTM layer (blue boxes), and fully connected layers (light blue boxes). The output consists of two components representing the latitudinal and longitudinal differences between the center of the input image and the estimated TC center (blue dots). (b) The architecture of the convolutional block in this study is identical to VGGNet, but without the fully connected layers. (c) In the ConvLSTM layer, six ConvLSTM cells are embedded. Each ConvLSTM cell takes an input image of a certain time step (Xt), and hidden and cell states of the previous time step (ht−1 and Ct−1, respectively). Those are processed through forget, input, and output gates to calculate the hidden and cell states of the corresponding time step (ht and Ct, respectively).
Citation: Bulletin of the American Meteorological Society 105, 3; 10.1175/BAMS-D-23-0130.1
The AI models were trained using separate training, validation, and testing datasets. The testing dataset consisted of TC samples collected in 2021. For the training and validation datasets, samples from 2016 to 2020 were randomly divided at a 4:1 ratio. The division of the dataset was based on the TCs, ensuring that multiple time steps for a single TC were assigned to the same dataset. To account for the potential bias in the performance scores obtained from a single calendar year, the same procedure was repeated using the 2020 samples as the testing dataset for sensitivity experiments. In this case, the training and validation datasets included TC samples from the years 2016–19 and 2021. Following this methodology, the AI models were trained, validated, and evaluated using different datasets to comprehensively assess their performance. It is noted that the ratios of TC grades were imbalanced—40%–43%, 35%–36%, 20%–21%, and 2%–3% for TD, TS, TY, and ST, respectively—within the training datasets and TC activity in the WNP was below normal during the two independent testing years, 2020/21. The details of the two AI algorithms and the training procedures using brightness temperature and reflectance images are described in the supplemental material.
A complete flowchart of the AI ensemble models for positioning the developed TC center is shown in Fig. 3. Three ensemble models were developed: a CNN-ensemble model, a ConvLSTM-ensemble model, and a combined model that incorporates both CNNs and ConvLSTMs. Each ensemble model consisted of six CNNs and six ConvLSTMs that were trained individually and then combined to form an ensemble.
Flowchart of AI models positioning the TC center with an example of TC Lupit (0000 UTC 8 Aug 2021). (a) Four CNNs are trained using IR images of the ambient time step as input and four ConvLSTMs are trained using IR images of six sequential time steps at 10-min intervals. The difference across the models and the determination of first-guess position (gray dot) are described in detail in the text. (b) Eight estimated TC centers (four from CNNs and four from ConvLSTMs) are obtained. The TC centers of the weighted average using the cosine of latitude are calculated among four CNN estimations and four ConvLSTM estimations. (c) The daytime and nighttime CNNs and ConvLSTMs are trained using visible and SWIR images. (d) The outputs are obtained from either daytime or nighttime models. (e) Finally, a CNN ensemble model (black plus sign) combines the five estimated TC centers (i.e., four IR-image-based CNNs, and either the daytime or nighttime CNN) with a weighted average using the cosine of latitude. Similarly, a ConvLSTM ensemble model (white plus sign) combines 5 estimations from ConvLSTMs. The combined model (red plus sign) merges 10 estimations from both the CNNs and ConvLSTMs. Red boxes show a close-up view centered on the first-guess position for each TC imagery.
Citation: Bulletin of the American Meteorological Society 105, 3; 10.1175/BAMS-D-23-0130.1
Six CNNs were trained using different combinations of input images. The four CNNs were trained using the four combinations of IR images (Fig. 3a), which are the IRsets1–4 in Table 1:
- 1)IRset1: only the 10.4-μm image;
- 2)IRset2: both the 10.4-μm and 10.4- minus 12.4-μm images;
- 3)IRset3: the 10.4-μm and 8.6- minus 11.2-μm images; and
- 4)IRset4: the 10.4-μm, 10.4- minus 12.4-μm, and 8.6- minus 11.2-μm images.
A list of input variable combinations for each AI model.
The outputs from each are represented by white and black circles and squares in Fig. 3b. For the training and validation datasets, the reference position (gray dots in Fig. 3a) was obtained by adding a random perturbation following a gamma distribution to the center position provided by the JTWC best track. For data augmentation, random perturbations were applied in four directions (northeast, northwest, southeast, and southwest) at each time step. The reference position for the testing dataset was obtained from the TC center for the 6- and 12-h forecast initialized at 0000 or 1200 UTC using the ECMWF model.
After training the above four CNNs, daytime and nighttime CNN models were developed (Fig. 3c). The former was trained using the 0.64-μm image, while the latter was trained using the 10.4- minus 3.9-μm image. For these models, a randomly perturbed position was used as the reference position for model training and validation. The reference position for the testing dataset was the latitude-weighted average of the outputs obtained from the four models trained using IR images. Similarly, four ConvLSTMs were developed by training IR images, and daytime and nighttime ConvLSTM models were developed.
Consequently, a CNN ensemble model (Fig. 3e) was developed by incorporating the estimations of five CNNs (i.e., four IR image-based CNNs and either the daytime or nighttime CNN). The CNN ensemble model combined the estimations of these five CNNs using a weighted average, i.e., the cosine of the latitude. Similarly, a ConvLSTM ensemble model (Fig. 3e) was developed by averaging the estimates of the five ConvLSTMs in the same manner. Finally, a model (Fig. 3e) was developed by combining the estimations of the five CNNs and five ConvLSTMs. It is noted that the ensemble models are only used for the testing datasets to evaluate the positioning errors.
Positioning tropical cyclone center
Positioning errors according to channel combinations and TC grades.
The errors in the center positioning generated by the two AI models were analyzed for the period from 2020 to 2021, which spanned two years of the independent testing datasets. Here we show the average performance by comprising several categories, whereas the thorough assessment for individual samples can be found in section 4 in the supplemental material. Figure 4 illustrates the specific error distributions, including the 10th, 30th, 50th (i.e., median), 70th, and 90th percentiles of all samples from both the CNN and ConvLSTM algorithms. The analysis is based on (i) various channel combinations and (ii) TC intensity grades. In terms of channel combination, we note that the 10.4 μm is the most crucial among the numerous channel information for accurately locating the TC center (Fig. 4a). While not shown in the figure, the errors calculated from other IR channels, excluding the 10.4 μm and the SWIR channel, are considerably larger. Consequently, many studies have focused solely on using the 10.4-μm channel in center positioning (e.g., Jaiswal and Kishtawal 2010; Wei et al. 2011; Wang et al. 2020; Tan 2021). For different channel combinations with the 10.4-μm channel, the errors range from 24 to 28 km at the median, from 35 to 40 km at the 70th percentile, and from 58 to 64 km at the 90th percentile. By adding channels, the error magnitudes were reduced, and the best results were achieved by averaging all five channel combinations (i.e., channel ensemble). The channel ensemble approach decreased the 90th-percentile error by up to 8 km. Overall, the CNN and ConvLSTM models dem-onstrated similar errors. However, the ConvLSTM model exhibited superior performance when the daytime and nighttime conditions were treated separately, particularly when the channel ensemble approach was applied.
Positioning errors of the TC center obtained from the CNN (blue) and ConvLSTM (red) models as a function of (a) five channel combinations and (b) TC intensity grades. Panel (b) shows the results from the ensemble average of the five channel combinations. The inner box and outer line cover the 30th–70th- and 10th–90th-percentile error values of all samples, respectively. The line inside the inner box indicates the 50th-percentile error value (i.e., a median).
Citation: Bulletin of the American Meteorological Society 105, 3; 10.1175/BAMS-D-23-0130.1
When considering the TC intensity grades, the center positioning errors decreased as the TCs became stronger (Fig. 4b), which is consistent with previous studies (Wimmers and Velden 2010, 2016). The median errors for both model ensembles were 34–39 km for TD, 29–32 km for TS, and 12–17 km for TY and ST. When the TC is weak (particularly for TD), the clouds are disorganized, making it challenging to correctly locate the center. However, as the TC intensifies (for TY and ST), clouds develop strongly around the center, resulting in easier center positioning (Hu and Zou 2020; Mayers and Ruf 2021). The AI model ensembles were not overfitted to weak TCs despite the larger number of training samples. This suggests that the distribution of training samples does not significantly affect the determination of the TC center location. The 90th-percentile errors exceed 110 km, which can be attributed to the fact that the ECMWF forecast positioning error used for initializing the present model often exceeds 100 km (approximately 20% of the TD grades). The errors in the initial and final positions of the present model show a significant correlation, except for the ST grade (not shown). Overall, the ConvLSTM model demonstrated smaller errors than the CNN model, except for the 90th-percentile error of the TD. The difference between the two models was particularly significant for the TD and TS grades.
CNN + ConvLSTM and comparison with ARCHER products.
When the results from both the AI models and multiple channels were combined, the TC center location became closer to the observed position. It is important to note that by merging multiple results, the explicit errors of the models and/or channels can be reduced (Bishop 2006). The center position errors obtained by combining the two models (CNN + ConvLSTM; the present model) are presented in Figs. 5a–c and are categorized by day and night, TC intensity grade, and translation speed, respectively. In the figure, the results from the present model are consistently compared with those of ARCHER. The difference in errors between the day and night was negligible, with errors less than 3 km for all error percentiles (Fig. 5a). Similarly, the difference was not significant at 95% confidence level; however, the present model showed a smaller error, with an improvement of up to 19 km at the 90th percentile. The improvement was more pronounced for larger errors. Although this study primarily utilized IR channels, the present model performs comparably to ARCHER, which incorporates an additional microwave channel to provide a closer view of the TC center.
Positioning errors of the TC center obtained from ARCHER (blue) and the present model (combined two models; red) according to (a) day and night, (b) the four TC intensity grades, and (c) the translation speed. The box plot is as in Fig. 4.
Citation: Bulletin of the American Meteorological Society 105, 3; 10.1175/BAMS-D-23-0130.1
The errors for each of the four TC grades are shown in Fig. 5b. The median errors for both models were approximately 40 km for TD, 30 km for TS, and 13 km for TY and ST. The present model demonstrated a meaningful reduction in error, up to 25 km at the 90th percentile, compared to ARCHER, specifically for TS. The performance of ARCHER is highly influenced by the satellite observations utilized, particularly for weaker TCs (Wimmers and Velden 2016). When the ARCHER model lacks access to polar-orbiting satellite data and relies solely on geostationary IR and/or visible channels, the median errors can exceed 50 km, especially for weaker TCs, such as TD and TS. In contrast, both models generated similar small magnitudes for all error percentiles when dealing with strong TCs, such as TY and ST.
The center positioning error also varied depending on the translation speed of the TC. A fast-moving TC under atmospheric conditions with strong wind shear in the midlatitudes can exhibit a vertically tilted cloud feature, posing challenges in accurately locating the center (Wimmers and Velden 2010). As depicted in Fig. 5c, the error magnitude is minimal for TC translation speeds below 30 km h−1. The median error was ≤25 km and the 70th percentile did not exceed 40 km. However, when the speed exceeded 30 km h−1, the error increased significantly, with the median reaching 34–36 km and the 70th percentile reaching 66 km. When comparing the errors between the present model and ARCHER, there was no notable difference at slower speeds of ≤30 km h−1. In contrast, the present model exhibited a smaller error than ARCHER for faster-moving TCs with speeds of ≥30 km h−1. This improvement can be attributed to the significant error reduction in the TS. Approximately 70% of these fast-moving TCs are active north of 25°N, indicating substantial improvements in the midlatitudes.
Figure 6 displays the spatial distributions of the center errors of the two models. Figures 6a–c represent the median error values, and Figs. 6d–f represent the 90th-percentile error values. The contour line indicates the number of TC samples that occurred within a 5° × 5° moving-window grid from 2020 to 2021. Most regions in the WNP have ≥10 TC samples while the Philippine Sea to the East China Sea has ≥50 TC samples. The median errors of the present model were approximately 25 km in the northeastern region relative to 15°N, 120°E, and they ranged from 30 to 40 km in the rest of the region, including the South China Sea and southern Philippine Sea (Fig. 6a). In contrast, ARCHER exhibits larger errors in the midlatitudes around 35°N (>45 km) and smaller errors in the southern Philippine Sea (<25 km) compared to the present model (Figs. 6b,c). In the remaining regions, the difference between the two methods was <10 km (Fig. 6c).
Spatial distribution of median errors (shading; km) of the TC center positioning obtained from (a) the present model and (b) ARCHER. (c) Difference between (a) and (b). (d)–(f) As in (a)–(c), but for the 90th-percentile error. The contour indicates the number of TC samples that occurred in a 5° × 5° moving-window grid every 6 h for 2020/21. The contour lines are 10, 30, and 50. Red dots in (c) and (f) indicate the area of significant difference at the 95% confidence level in the Wilcoxon signed-rank tests.
Citation: Bulletin of the American Meteorological Society 105, 3; 10.1175/BAMS-D-23-0130.1
For the 90th-percentile error, the magnitude was approximately twice as large as the median for both models; however, it was relatively larger for ARCHER (Figs. 6d,e). The differences are particularly noticeable in the South China Sea and midlatitude regions, including the East China Sea, Korean Peninsula, and southern Japan. In the South China Sea, the present model had an error of over 60 km, whereas in the rest of the region, it was within 50 km (Fig. 6d). However, ARCHER showed errors of over 60 km in the South China Sea, over 50 km in the East China Sea and surrounding areas, and within 40 km in the remaining region (Fig. 6e). As a result, the difference between the models was more pronounced in the midlatitude regions (Fig. 6f). The present model reduced the 90th-percentile error by more than 20 km in these regions. This result is consistent with the error characteristics shown in Fig. 5. There was a relatively high proportion of weak and fast-moving TCs in the midlatitude regions compared to the rest of WNP; thus, a large improvement in the 90th percentile was observed over these regions. The number of samples at approximately 35°N, where the error decreases significantly, is ∼10 or less, which is a relatively small number compared to the other regions. This indicates that skill improvements can be achieved regardless of the number of samples.
Conclusions
Accurate estimation of the center position of a TC is a crucial step in TC forecasting operations. Center identification serves as the foundation for correctly determining key variables, such as the central pressure, maximum wind speed, and radius of maximum wind. When the center position is properly estimated, these variables are closely aligned with the observed values. In addition, the center position and key variables are implemented as initial conditions for the numerical and statistical models for TC forecasting. Therefore, relatively precise TC position estimation plays a pivotal role in improving the overall forecast accuracy. Recognizing its significance, numerous meteorological organizations have dedicated efforts to developing methods and techniques for the precise estimation of TC centers.
In this study, the brightness temperature of four IR channels and one SWIR channel and the reflectivity of one visible channel of geostationary satellite imagery served as the inputs to two AI models, CNN and ConvLSTM, to develop a model to estimate the center location of a TC. Combining all channel information and utilizing the two AI models produced the best performance when compared to the JTWC best track. Compared to CIMSS ARCHER, the present model performed similarly or even slightly better. The difference in the error magnitude between the daytime and nighttime was negligible. This is a positive indication for operational use, as it implies that the model can maintain its accuracy even during the night when a visible channel is not available. Similar to ARCHER, the error magnitude of the present model decreased as the TC intensity increased. The median (i.e., 50th-percentile) error was less than 15 km for TY and ST, 30 km for TS, and 35 km for TD. We note that for the TS grade, the present model demonstrated a 90th-percentile error of 60 km, which is 25 km lower than the 85-km error observed in ARCHER. This improvement is particularly evident in the fast-moving midlatitude TCs with speeds over 30 km h−1.
The present model results were compared with the JTWC best track at 0000, 0600, 1200, and 1800 UTC. The ARCHER model, which is used operationally, provides center location estimates 30 min earlier than these time points to provide timely information to forecasters. To ensure a fair comparison with ARCHER, the center location estimates from ARCHER were interpolated to a value of 30 min later, considering the time difference between the present model results and ARCHER’s operational estimates. In addition, ARCHER updates the center location with additional geostationary and/or polar-orbiting satellite information that becomes available after operational time.
The model developed in this study, despite not using polar-orbiting satellite information to estimate the TC center location, performed with an accuracy similar to that of ARCHER, thus demonstrating its potential for operational use. The present model can be applied to meteorological organizations (e.g., the KMA) that have limitations in employing real-time polar-orbiting satellite data in their operating system. Operational organizations can use this model in conjunction with ARCHER to improve the reliability of TC center estimates when polar-orbiting satellite data are unavailable.
Acknowledgments.
This work was funded by the Korea Meteorological Administration Research and Development Program under Grant RS-2023-00236880.
Data availability statement.
The TC best track data of the JTWC are available at www.metoc.navy.mil/jtwc/jtwc.html?best-tracks, the real-time TC center location from ARCHER of CIMSS at http://tropic.ssec.wisc.edu/real-time/archerOnline/cyclones/, and Himawari-8 data of the JAXA at www.eorc.jaxa.jp/ptree/userguide.html.
References
Ackerman, S. A., W. L. Smith, H. E. Revercomb, and J. D. Spinhirne, 1990: The 27–28 October 1986 FIRE IFO cirrus case study: Spectral properties of cirrus clouds in the 8–12 μm window. Mon. Wea. Rev., 118, 2377–2388, https://doi.org/10.1175/1520-0493(1990)118<2377:TOFICC>2.0.CO;2.
Alliss, R. J., S. Raman, and S. W. Chang, 1992: Special Sensor Microwave/Imager (SSM/I) observations of Hurricane Hugo (1989). Mon. Wea. Rev., 120, 2723–2737, https://doi.org/10.1175/1520-0493(1992)120<2723:SSMOOH>2.0.CO;2.
Bessho, K., and Coauthors, 2016: An introduction to Himawari-8/9—Japan’s new-generation geostationary meteorological satellites. J. Meteor. Soc. Japan, 94, 151–183, https://doi.org/10.2151/jmsj.2016-009.
Bishop, C. M., 2006: Pattern Recognition and Machine Learning. Springer, 738 pp.
Cha, D.-H., and Y. Wang, 2013: A dynamical initialization scheme for real-time forecasts of tropical cyclones using the WRF model. Mon. Wea. Rev., 141, 964–986, https://doi.org/10.1175/MWR-D-12-00077.1.
Cha, Y., J. Kim, K. Kang, N.-Y. Kang, and W.-T. Yun, 2016: A study on the determination of Korea affecting tropical cyclone center for best tracking. Adv. Meteor., 2016, 1–13, https://doi.org/10.1155/2016/5946342.
Chen, L., X. Zhuge, X. Tang, J. Song, and Y. Wang, 2022: A new type of red-green-blue composite and its application in tropical cyclone center positioning. Remote Sens., 14, 539, https://doi.org/10.3390/rs14030539.
Chen, Y., and C. Snyder, 2007: Assimilating vortex position with an ensemble Kalman filter. Mon. Wea. Rev., 135, 1828–1845, https://doi.org/10.1175/MWR3351.1.
Christophersen, H., J. Sippel, A. Aksoy, and N. L. Baker, 2022: Recent advancements for tropical cyclone data assimilation. Ann. N. Y. Acad. Sci., 1517, 25–43, https://doi.org/10.1111/nyas.14873.
Chu, J. H., C. R. Sampson, A. S. Levine, and E. Fukada, 2002: The Joint Typhoon Warning Center tropical cyclone best-tracks, 1945–2000. Joint Typhoon Warning Center Rep. NRL/MR/7540-02-16, 22 pp., https://www.metoc.navy.mil/jtwc/products/best-tracks/tc-bt-report.html.
Dvorak, V. F., 1984: Tropical cyclone intensity analysis using satellite data. NOAA Tech. Rep. NESDIS 11, 47 pp., http://satepsanone.nesdis.noaa.gov/pub/Publications/Tropical/Dvorak_1984.pdf.
ECMWF, 2019: IFS documentation CY46R1. ECMWF Tech. Rep., 82 pp., https://www.ecmwf.int/en/elibrary/81138-ifs-documentation-cy46r1-part-i-observations.
Guard, C. P., L. E. Carr, F. H. Wells, R. A. Jeffries, N. D. Gural, and D. K. Edson, 1992: Joint Typhoon Warning Center and the challenges of multibasin tropical cyclone forecasting. Wea. Forecasting, 7, 328–352, https://doi.org/10.1175/1520-0434(1992)007<0328:JTWCAT>2.0.CO;2.
Heming, J. T., 2016: Met Office Unified Model tropical cyclone performance following major changes to the initialization scheme and a model upgrade. Wea. Forecasting, 31, 1433–1449, https://doi.org/10.1175/WAF-D-16-0040.1.
Hendricks, E. A., M. S. Peng, X. Ge, and T. Li, 2011: Performance of a dynamic initialization scheme in the Coupled Ocean–Atmosphere Mesoscale Prediction System for Tropical Cyclones (COAMPS-TC). Wea. Forecasting, 26, 650–663, https://doi.org/10.1175/WAF-D-10-05051.1.
Hsiao, L.-F., C.-S. Liou, T.-C. Yeh, Y.-R. Guo, D.-S. Chen, K.-N. Huang, C.-T. Terng, and J.-H. Chen, 2010: A vortex relocation scheme for tropical cyclone initialization in Advanced Research WRF. Mon. Wea. Rev., 138, 3298–3315, https://doi.org/10.1175/2010MWR3275.1.
Hu, Y., and X. Zou, 2020: Comparison of tropical cyclone center positions determined from satellite observations at infrared and microwave frequencies. J. Atmos. Oceanic Technol., 37, 2101–2115, https://doi.org/10.1175/JTECH-D-20-0049.1.
Inoue, T., 1987: A cloud type classification with NOAA 7 split-window measurements. J. Geophys. Res., 92, 3991–4000, https://doi.org/10.1029/JD092iD04p03991.
Jaiswal, N., and C. M. Kishtawal, 2010: Automatic determination of center of tropical cyclone in satellite-generated IR images. IEEE Geosci. Remote Sens. Lett., 8, 460–463, https://doi.org/10.1109/LGRS.2010.2085418.
Kim, J., S.-O. Hwang, S.-S. Kim, I. Oh, and D.-J. Ham, 2022: Algorithms for determining Korea Meteorological Administration (KMA)’s official typhoon best tracks in the National Typhoon Center. Atmosphere, 32, 381–394, https://doi.org/10.14191/Atmos.2022.32.4.381.
Kim, Y., and S. Hong, 2019: Convective cloud RGB product and its application to tropical cyclone analysis using geostationary satellite observation. J. Korean Earth Sci. Soc., 40, 406–413, https://doi.org/10.5467/JKESS.2019.40.4.406.
Kleist, D. T., 2011: Assimilation of tropical cyclone advisory minimum sea level pressure in the NCEP Global Data Assimilation System. Wea. Forecasting, 26, 1085–1091, https://doi.org/10.1175/WAF-D-11-00045.1.
Kunii, M., 2015: Assimilation of tropical cyclone track and wind radius data with an ensemble Kalman filter. Wea. Forecasting, 30, 1050–1063, https://doi.org/10.1175/WAF-D-14-00088.1.
Kurihara, Y., M. A. Bender, and R. J. Ross, 1993: An initialization scheme of hurricane models by vortex specification. Mon. Wea. Rev., 121, 2030–2045, https://doi.org/10.1175/1520-0493(1993)121<2030:AISOHM>2.0.CO;2.
Landsea, C. W., and J. L. Franklin, 2013: Atlantic hurricane database uncertainty and presentation of a new database format. Mon. Wea. Rev., 141, 3576–3592, https://doi.org/10.1175/MWR-D-12-00254.1.
LeCun, Y., Y. Bengio, and G. Hinton, 2015: Deep learning. Nature, 521, 436–444, https://doi.org/10.1038/nature14539.
Lee, T. F., 2000: Nighttime observation of sheared tropical cyclones using GOES 3.9-μm data. Wea. Forecasting, 15, 759–766, https://doi.org/10.1175/1520-0434(2000)015<0759:NOOSTC>2.0.CO;2.
Lee, T. F., F. J. Turk, and K. Richardson, 1997: Stratus and fog products using GOES-8–9 3.9-μm data. Wea. Forecasting, 12, 664–677, https://doi.org/10.1175/1520-0434(1997)012<0664:SAFPUG>2.0.CO;2.
Lee, T. F., F. J. Turk, J. Hawkins, and K. Richardson, 2002: Interpretation of TRMM TMI images of tropical cyclones. Earth Interact., 6, https://doi.org/10.1175/1087-3562(2002)006<0001:IOTTIO>2.0.CO;2.
Mayers, D., and C. Ruf, 2021: MTrack: Improved center fix of tropical cyclones from SMAP wind observations. Bull. Amer. Meteor. Soc., 102, E700–E709, https://doi.org/10.1175/BAMS-D-20-0068.1.
Olander, T. L., and C. S. Velden, 2007: The advanced Dvorak technique: Continued development of an objective scheme to estimate tropical cyclone intensity using geostationary infrared satellite imagery. Wea. Forecasting, 22, 287–298, https://doi.org/10.1175/WAF975.1.
Olander, T. L., and C. S. Velden, 2019: The advanced Dvorak technique (ADT) for estimating tropical cyclone intensity: Update and new capabilities. Wea. Forecasting, 34, 905–922, https://doi.org/10.1175/WAF-D-19-0007.1.
Shi, X., Z. Chen, H. Wang, D.-Y. Yeung, W.-K. Wong, and W.-C. Woo, 2015: Convolutional LSTM network: A machine learning approach for precipitation nowcasting. NIPS'15: Proceedings of the 28th International Conference on Neural Information Processing Systems, Vol. 1, MIT Press, 802–810.
Shimizu, A., 2020: Introduction to Himawari-8 RGB composite imagery. Meteorological Satellite Center Tech. Note 65, 44 pp., https://www.data.jma.go.jp/mscweb/technotes/msctechrep65-1.pdf.
Tan, C., 2021: TCLNet: Learning to locate typhoon center using deep neural network. 2021 IEEE Int. Geoscience and Remote Sensing Symp. IGARSS, Brussels, Belgium, Institute of Electrical and Electronics Engineers, 4600–4603, https://doi.org/10.1109/IGARSS47720.2021.9554524.
Velden, C., and Coauthors, 2006: The Dvorak tropical cyclone intensity estimation technique: A satellite-based method that has endured for over 30 years. Bull. Amer. Meteor. Soc., 87, 1195–1210, https://doi.org/10.1175/BAMS-87-9-1195.
Wang, P., P. Wang, C. Wang, Y. Yuan, and D. Wang, 2020: A center location algorithm for tropical cyclone in satellite infrared images. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens., 13, 2161–2172, https://doi.org/10.1109/JSTARS.2020.2995158.
Wei, K., Z. Jing, Y.-X. Li, and S. Liu, 2011: Spiral band model for locating tropical cyclone centers. Pattern Recognit. Lett., 32, 761–770, https://doi.org/10.1016/j.patrec.2010.12.011.
Wimmers, A. J., and C. S. Velden, 2010: Objectively determining the rotational center of tropical cyclones in passive microwave satellite imagery. J. Appl. Meteor. Climatol., 49, 2013–2034, https://doi.org/10.1175/2010JAMC2490.1.
Wimmers, A. J., and C. S. Velden, 2016: Advancements in objective multisatellite tropical cyclone center fixing. J. Appl. Meteor. Climatol., 55, 197–212, https://doi.org/10.1175/JAMC-D-15-0098.1.
Wu, C.-C., G.-Y. Lien, J.-H. Chen, and F. Zhang, 2010: Assimilation of tropical cyclone track and structure based on ensemble Kalman filter (EnKF). J. Atmos. Sci., 67, 3806–3822, https://doi.org/10.1175/2010JAS3444.1.
Ying, M., W. Zhang, H. Yu, X. Lu, J. Feng, Y. Fan, Y. Zhu, and D. Chen, 2014: An overview of the China Meteorological Administration tropical cyclone database. J. Atmos. Oceanic Technol., 31, 287–301, https://doi.org/10.1175/JTECH-D-12-00119.1.
