Journal of the Meteorological Society of Japan https://doi.org/10.1007/s44394-026-00040-5 (2026) 104:38 ARTICLE A long-term dataset of tropical cyclone radius of maximum wind in the Western North Pacific Based on a Bayesian consensus of multiple estimates Udai Shimada1,2 · Taiga Tsukada3 Received: 2 June 2026 / Accepted: 18 September 2026 © The Author(s) 2026 Abstract The radius of maximum wind (RMW) is one of the key parameters for characterizing the structure of tropical cyclones (TCs). Although multiple methods have been developed to estimate the RMW, their estimates can sometimes differ substantially and are not always available, making it difficult to construct a comprehensive long-term RMW dataset. This study develops a Bayesian consensus method to derive a single RMW estimate from multiple RMW estimates. In this framework, the posterior probability distribution of the RMW is estimated using likelihoods constructed from methodspecific error distributions combined with an outlier distribution, and the final RMW is determined as the maximum a posteriori (MAP) estimate. By explicitly accounting for an outlier distribution for each estimation method, the proposed method can determine a reasonable RMW close to one of the clusters even when the estimates split into two clusters. Using this consensus method and, as a primary source, estimates from satellite-based global precipitation data and from infrared and microwave satellite data for TCs with an eye, a long-term RMW dataset is constructed at 3-h intervals for 631 TCs in the western North Pacific during 2000–2025. Verification of the consensus RMW estimates shows that errors decrease with increasing TC intensity and posterior probability, and that posterior probability itself serves as an indicator of estimate reliability. This new RMW dataset enables a realistic representation of the inner-core structure, particularly small RMWs, which were difficult to estimate before the advent of synthetic aperture radar (SAR)-derived winds and in the absence of aircraft observations. Keywords Tropical cyclones · Radius of maximum wind · Long-term dataset · Satellite observation 1 Introduction The radius of maximum wind (RMW) is a key parameter for characterizing the wind structure of tropical cyclones (TCs), together with the maximum wind speed (Vmax) and outer size, as these parameters are physically related (Emanuel and Rotunno 2011; Chavas and Lin 2016). The magnitude Udai Shimada ushimada@mri-jma.go.jp 1 Meteorological Research Institute, Japan Meteorological Agency, Tsukuba, Japan 2 Typhoon Science and Technology Research Center, Yokohama National University, Yokohama, Japan 3 Cooperative Institute for Research in the Atmosphere, Colorado State University, Fort Collins, CO, USA of the RMW is associated with subsequent changes in TC intensity (Carrasco et al. 2014). TC rapid intensification often occurs after rapid RMW contraction (e.g., Stern et al. 2015; Wu and Ruan 2021; Li et al. 2022). The RMW can rapidly jump from the inner eyewall to the outer eyewall during an eyewall replacement cycle (e.g., Willoughby et al. 1982; Black and Willoughby 1992; Sitkowski et al. 2011). RMW information is used for wind structure construction (e.g., Chen et al. 2025; Liu et al. 2025; Tao et al. 2026), which is also utilized for forecasting ocean waves and storm surge (e.g., Aijaz et al. 2019). While recently developed artificial intelligence (AI)-based weather models have demonstrated significant improvements in TC track forecasts, they have shown negative biases and poor representation of TC intensity and structure (e.g., Yamaguchi et al. 2025; DeMaria et al. 2025). Such weaknesses can be mitigated if a reliable RMW dataset for improved model training 38 Page 2 of 19 is available. Given the importance of the RMW, a World Meteorological Organization (WMO workshop has called for the development of methods to construct more accurate RMW datasets (e.g., Rogers 2023; Ito et al. 2025). Recently, new RMW estimation methods have been developed, and multiple RMW estimates have become available at the same analysis time. One reason for this progress is the availability of accurate RMWs from spaceborne synthetic aperture radar (SAR) observations, which provide detailed TC wind fields with a high spatial resolution (~ 500 m to 3 km), although the observation frequency with full TC coverage is still low (Mouche et al. 2017, 2019, 2026; Combot et al. 2020; Vinour et al. 2021; Shimada et al. 2024; Tsukada et al. 2025; Avenas et al. 2026). Methods developed using SAR-derived RMWs as ground truth include those based on outer wind radii and eye size derived from infrared and passive microwave satellite observations (Avenas et al. 2023; Tsukada and Horinouchi 2023; Shimada 2025). In addition to these new approaches, previously developed methods (Willoughby and Rahn 2004; Mueller et al. 2006; Kossin et al. 2007; Vickery and Wadhera 2008; Lajoie and Walsh 2008; Knaff et al. 2015; Chavas and Knaff 2022), as well as recent machine-learning-based methods (e.g., Zhuo and Tan 2021, 2023; Lu et al. 2022), are also applicable for estimating the RMW. However, RMW estimates can differ substantially among methods, their time series may fluctuate irregularly, and some methods do not always provide estimates due to limited data availability. Although many RMW estimation methods have been developed, to our knowledge, a long-term RMW dataset extending back several decades, with homogeneous quality throughout the entire period, has not yet been constructed. To achieve this, a consensus method is required to derive a single representative RMW from multiple RMW estimates. Such a consensus RMW cannot be obtained by a simple arithmetic average of the estimates. For example, an arithmetic average is inappropriate in cases of TCs with concentric eyewalls, where one method estimates the RMW near the inner eyewall while another estimates it near the outer eyewall. Additionally, ideal RMW estimates intended for diverse applications should (i) evolve smoothly over time to represent the scale of the TC inner core, (ii) capture variations associated with eyewall replacement cycles, (iii) be continuously available throughout the TC life cycle, and (iv) maintain homogeneous data quality across the entire period. The purpose of this study is to develop a consensus method of RMW estimation based on a Bayesian framework and to construct a long-term RMW dataset. Specifically, the posterior probability distribution of the RMW is estimated using likelihoods constructed from method-specific error distributions combined with an outlier distribution, and the final RMW is determined as the maximum a 13 Journal of the Meteorological Society of Japan (2026) 104:38 posteriori (MAP) estimate. By explicitly accounting for an outlier distribution for each method, the posterior distribution can exhibit multimodality. This approach can determine a reasonable RMW close to one of the clusters even when the estimates split into two distinct clusters, rather than yielding an intermediate value between them. For the construction of the long-term RMW dataset, this study targets 631 TCs that occurred in the western North Pacific from 2000 to 2025, with a temporal resolution of 3 h. To ensure homogeneous quality throughout the period, RMWs are estimated using only observations that have been continuously available since 2000. Additionally, new estimation methods are developed using global satellitederived precipitation data (Global Satellite Mapping of Precipitation, GSMaP; Kubota et al. 2007; Kubota et al. 2009) and atmospheric reanalysis data [the European Centre for Medium-Range Weather Forecasts (ECMWF) fifth-generation reanalysis product, ERA5; Hersbach et al. 2020] to provide RMW estimates at all analysis times and to suppress irregular temporal fluctuations. The proposed consensus method selectively determines an RMW value closest to the most accurate estimate among the available methods. In this study, the RMW is defined as the radius of maximum azimuthally averaged wind, rather than the radius of the local maximum wind speed (hereafter, the local RMW), following previous studies (e.g., Vinour et al. 2021; Avenas et al. 2023; Tsukada and Horinouchi 2023; Shimada 2025). This definition allows the RMW to represent the horizontal scale of the inner-core region of large positive vorticity in the Northern Hemisphere, as the positive vorticity in a modified Rankine vortex decreases rapidly outside the RMW. Additionally, this definition reduces sensitivity to local maxima caused by instantaneous gusty wind speeds and is therefore expected to yield more stable RMWs when satellite observations are used as ground truth. One caveat is that a relatively large difference may exist between the local RMW and the RMW defined in this study for TCs with elliptical eyewalls or concentric eyewalls. The rest of the paper is organized as follows. Section 2 describes the data used in this study. Section 3 presents the RMW estimation methods used in the consensus. Section 4 describes the consensus method. Section 5 presents the verification results of the consensus, including case studies. Section 6 presents the long-term RMW dataset. Section 7 provides a summary of the study. 2 Data In this study, best track Vmax data from the Japan Meteorological Agency (JMA) are used, except in Sect. 6. Because JMA Vmax is not provided during the tropical depression Journal of the Meteorological Society of Japan (2026) 104:38 (TD) stage, it is extrapolated assuming a linear inverse relationship with the central pressure available during the TD stage. Consequently, the extrapolated Vmax may occasionally reach tropical storm (TS) strength (18 m s–1) during the TD stage. In Sect. 6, Vmax and RMW data from the Joint Typhoon Warning Center (JTWC; Howell et al. 2025) are used for comparison. The following datasets are used for RMW estimation: hourly GSMaP data with a spatial resolution of 0.1°, ERA5 10-m wind data with a spatial resolution of 0.25° at 3-h intervals, intercalibrated passive microwave (PMW) satellite data at 85–91 GHz, and hourly geostationary infrared (IR) satellite observations. GSMaP is a multi-satellite global precipitation product that uses observations from various satellite sensors, including the 85–91-GHz, 37-GHz, and other channels of PMW imagers, PMW sounders, and geostationary IR imagers (JAXA 2026). The GSMaP data are mapped onto TC-centered polar coordinates and averaged at 3-h intervals using seven hourly values centered on each analysis time (from −3 to +3 h) to mitigate differences in precipitation estimates among satellite sensors, following Shimada et al. (2017). The TC center position at each time is derived from hourly interpolated best track data provided by the JMA. Details of GSMaP data processing and its applicability to TCs are described in Shimada et al. (2017, 2018). The TC center in the ERA5 data is identified as the location of the minimum in the smoothed ERA5 mean sea level pressure field, following Aizawa et al. (2024). The ERA5 10-m wind fields are then interpolated onto TC-centered polar coordinates. The PMW satellite data are obtained from the TC PRIMED dataset (Tropical Cyclone Precipitation, Infrared, Microwave, and Environmental Dataset; Razin et al. 2023b). This dataset includes observations from the Special Sensor Microwave Imager (SSM/I), the Special Sensor Microwave Imager/Sounder (SSMIS), the Tropical Rainfall Measuring Mission (TRMM) Microwave Imager (TMI), the Advanced Microwave Scanning Radiometer for the Earth Observing System (AMSR-E), the Advanced Microwave Scanning Radiometer 2 (AMSR2), and the Global Precipitation Measurement (GPM) Microwave Imager (GMI). Geostationary IR satellite observations are from the JMA (e.g., Bessho et al. 2016). To ensure consistency in horizontal resolution among successive satellites, the IR data are resampled to a spatial resolution of 5 km in this study. This study uses SAR wind data and Advanced Scatterometer (ASCAT) wind data as reference datasets for the development of RMW estimation methods and for the verification of RMW estimates. Unless otherwise noted, SAR/ ASCAT data acquired within ± 1.5 h of each 3-hourly analysis time are used for these purposes. The SAR wind data are used for TCs at TS intensity or higher because only a few SAR samples are available for TD cases. Accordingly, Page 3 of 19 38 the ASCAT data are used for TD cases and TS cases with Vmax = 18 m s–1. The SAR wind data, including TC center position information, are obtained from the National Oceanic and Atmospheric Administration’s (NOAA) Center for Satellite Applications and Research (STAR) Tropical Storm Cyclone Winds page (STAR 2026) for the period 2015–2025, with a horizontal resolution of 3 km1. Data from 2015 to 2024 are used as the reference dataset (unless otherwise noted), whereas data from 2025 are used for verification. Cases with two or more distinct peaks in the radial profile of the azimuthally averaged SAR wind speed are excluded from the reference dataset to ensure the derivation of a single RMW estimate. The ASCAT wind data are obtained from the National Aeronautics and Space Administration (NASA) Earthdata Search (NASA 2026) for the period 2012–2025, with a horizontal resolution of 25 km. The TC center is defined as the location of the maximum streamfunction derived from the vorticity field computed from ASCAT wind vectors. The ASCAT wind fields are then bilinearly interpolated onto TC-centered polar coordinates with a radial spacing of 2 km to derive the RMW. In this process, wind data flagged for rain attenuation are also used to maintain wind-field coverage. Due to the coarse spatial resolution, possible center displacement, and inclusion of flagged wind data, however, the azimuthally averaged ASCAT wind profile yields an unrealistically small RMW (≤ 10 km) in ~10% of the samples. To address this issue, the radius of the local maximum ASCAT wind (i.e., the local RMW) is adopted as the RMW when the azimuthally averaged RMW is less than 150 km, because the two RMW estimates are generally consistent within this range except for such problematic cases. 3 RMW estimation methods In this study, six estimation methods are incorporated into the consensus framework to construct a long-term RMW dataset, four of which are newly developed. These new methods are needed because RMW estimates are required at all analysis times, including the TD stage, over a long period extending back to 2000 using a homogeneous dataset, requirements that are not fully satisfied by previously developed methods. In this section, two GSMaP-based + TD methods (𝑅𝑅𝑅𝑅𝑅𝑅 TS GSMaP and 𝑅𝑅𝑅𝑅𝑅𝑅 GSMaP ; see Appendix A for the notation used in this paper), a modified Rankine vortex TD + method (𝑅𝑅𝑅𝑅𝑅𝑅 TS Rankine ), an ERA5-based method (𝑅𝑅𝑅𝑅𝑅𝑅 ERA5 ), an IR-based method (𝑅𝑅𝑅𝑅𝑅𝑅 IR ), and a PMW-based method (𝑅𝑅𝑅𝑅𝑅𝑅 PMW ) are introduced. The notations “TS+” and “TD” 1 The 3-km product is a spatial average of the 500-m instantaneous wind speed to mimic a 1-min sustained wind speed (STAR 2026). 13 38 Page 4 of 19 Journal of the Meteorological Society of Japan indicate that the corresponding methods are applicable to cases with Vmax ≥ 18 m s–1 and Vmax < 20 m s–1, respectively. The intensity ranges overlap to allow the TD methods to be applied to cases that temporarily weaken to TD-like intensity while remaining classified as TS in the best-track data. 3.1 Error distribution The Bayesian framework necessitates error distributions for each constituent estimates. Consequently, two types of estimation error distributions are adopted in the consensus method in this study: a normal distribution: 𝑝𝑝 (𝜀𝜀) = √ 1 2𝜋𝜋𝜋𝜋 exp [− (𝜀𝜀 𝜀 𝜀𝜀) ] 2𝜎𝜎 2 2 and a log-normal distribution: 𝑝𝑝𝑝𝑝𝑝𝑝 𝑝 √ 1 2𝜋𝜋𝜋𝜋(𝑈𝑈𝑈𝑈𝑈) exp [− (ln(𝑈𝑈𝑈𝑈𝑈𝑈𝑈𝑈𝑈) 2𝜎𝜎 2 2 ] (𝑈𝑈 𝑈𝑈𝑈 𝑈 𝑈) where ε denotes the estimation error (defined as the estimate minus the SAR-derived RMW), and μ and σ denote the mean (bias) and standard deviation of the error, respectively. The constant 𝑈𝑈 is an offset introduced for the log-normal distribution and is selected as the value that provides the best fit to the log-normal distribution among multiple candidates near the point at which the frequency rapidly decreases. Specifically, because a log-normal distribution requires ln(𝑈𝑈 𝑈 𝑈𝑈𝑈 to follow a normal distribution, deviations from linearity in a normal quantile–quantile (Q–Q) plot are evaluated using the root-mean-square error (RMSE), and the value of 𝑈𝑈 that minimizes the RMSE is selected. The error distributions of the RMW estimation methods used in the consensus method are summarized in Table 1. In the following evaluation of the error distribution of RMW estimates from each method, it is necessary to identify and exclude outliers, as outliers can cause the distribution to Table 1 Error distribution parameters of the RMW estimation methods used in the consensus framework Estimates Applicable Error μ [km] σ [km] U case distribution [km] + Normal 1.359 10.617 – Vmax ≥ 18 m 𝑅𝑅𝑅𝑅𝑅𝑅 TS GSMaP –1 s TD Normal –5.977 81.914 – max < 20 m 𝑅𝑅𝑅𝑅𝑅𝑅 GSMaP V–1 s + Log-normal 2.981 0.776 14 Vmax ≥ 18 m 𝑅𝑅𝑅𝑅𝑅𝑅 TS Rankine –1 s Normal –3.846 67.849 – Vmax < 20 m 𝑅𝑅𝑅𝑅𝑅𝑅 TD ERA5 –1 s Dvorak eye Log-normal 2.066 0.599 9 𝑅𝑅𝑅𝑅𝑅𝑅 IR pattern Log-normal 2.045 0.518 7 max > 33 m 𝑅𝑅𝑅𝑅𝑅𝑅 PMW V–1 s 13 (2026) 104:38 deviate from a normal or log-normal distribution. For example, in cases of concentric eyewalls, a method may estimate a radius associated with the inner eyewall as the RMW, even though the true RMW is located near the outer eyewall. This results in a large estimation error, which constitutes an outlier far from the true RMW. Additionally, outliers can arise when the precipitation distribution used for RMW estimation is poorly correlated with the actual RMW. This study employs the median absolute deviation (MAD) as a robust estimator of the standard deviation of RMW estimates in the presence of outliers: ̃ MAD = median(|𝜀𝜀𝑖𝑖 − 𝜀𝜀|) where 𝜀𝜀𝑖𝑖 denotes each estimation error and 𝜀𝜀 ̃ denotes the median of the estimation errors. For a standard normal distribution 𝒩𝒩𝒩0,1), the MAD is ~0.6745. Accordingly, assuming that the inlier errors are approximately normally distributed, the MAD computed from the errors can be used to estimate the standard deviation of the underlying inlier distribution as 𝜎𝜎ˆMAD = MAD/0.6745. For simplicity, assuming that 10% of the data consist of outliers, the central 90% interval of a standard normal distribution is given by ±1.645. There𝜎𝜎MAD can be adopted as the criterion for outlier fore, ±1.645ˆ exclusion among errors in this study. + TD 3.2 𝑹𝑹𝑹𝑹𝑹𝑹TS GSMaP and 𝑹𝑹𝑹𝑹𝑹𝑹GSMaP Two GSMaP-based methods are developed to collectively provide RMW estimates at all analysis times, regardless of + the presence or absence of a TC eye: 𝑅𝑅𝑅𝑅𝑅𝑅 TS GSMaP for TS TD strength and above, and 𝑅𝑅𝑅𝑅𝑅𝑅 GSMaP for cases with Vmax + < 20 m s–1 (i.e., including weak TS cases). 𝑅𝑅𝑅𝑅𝑅𝑅 TS GSMaP exploits a structural relationship in which the radius of maximum precipitation (Rmp) is located outside the surface RMW within a sloped eyewall (e.g., Jorgensen 1984). The GSMaP-based methods are developed based on the following concepts: i) They provide RMW estimates at all analysis times. ii) They provide error distributions based on the precipitation distribution, rather than simply assuming normal or log-normal distributions (see Sect. 4.2). iii) When other estimates are available, the consensus framework can reduce errors from the GSMaP-based methods by incorporating the more accurate IR/PMWbased estimates for TCs with an eye and the Rankine/ ERA5-based estimates for those without an eye. Specifically, Rmp is derived from GSMaP data using the following procedures: Journal of the Meteorological Society of Japan (2026) 104:38 i) Apply a 12-h running mean to the 3-hourly azimuthally averaged precipitation to reduce temporal discontinuities in the precipitation distribution. ii) Detect Rmp as the radius of maximum azimuthally averaged precipitation, and identify the first, second, and third peaks of the azimuthally averaged precipitation in order of increasing radius beyond a radius of 10 km (Rmp1, Rmp2, and Rmp3, when present). Note that Rmp is identical to Rmp1 when Rmp > 10 km. iii) When Rmp < 10 km, set Rmp = Rmp1 if Rmp1 is available to account for possible TC center displacement. iv) When Rmp≥ 400 km, redefine Rmp as the candidate with the maximum precipitation among Rmp1, Rmp2, and Rmp3. v) If Rmp differs by 10 km or more from both its values at the previous and subsequent times (± 3 h), and the differences have opposite signs, redefine 𝑅𝑅mp as the candidate among Rmp1, Rmp2, and Rmp3 (in this order) whose difference from the previous value is less than 10 km. If no such candidate exists, replace 𝑅𝑅mp with the average of the previous and subsequent values. The regression relating 𝑅𝑅mp to the SAR-derived RMW is derived using samples with 15 km < 𝑅𝑅mp < 150 km and maximum azimuthally averaged precipitation rates exceeding 5 mm h–1 at each time (Fig. 1a): + 𝑅𝑅𝑅𝑅𝑅𝑅 TS GSMaP = 0.59𝑅𝑅mp + 4.68 (1) The correlation is weak for samples with 𝑅𝑅mp < 16 km, which is attributed to the coarse spatial resolution of GSMaP (0.1°), as small displacements in the TC center position can lead to relatively large errors in cases with small RMWs. Fig. 1 a Scatter plot of Rmp (km) versus SAR-derived RMW (km). Black circles denote samples used in the regression (red line), satisfying 15 km < 𝑅𝑅mp < 150 km and maximum azimuthally averaged precipitation rates exceeding 5 mm h–1; white circles denote the remain+ ing samples. b Distribution of 𝑅𝑅𝑅𝑅𝑅𝑅 TS GSMaP errors (km), with Gaussian fits for all samples (gray) and for samples excluding outliers deter- Page 5 of 19 38 + Therefore, when 𝑅𝑅mp < 16 km, set 𝑅𝑅𝑅𝑅𝑅𝑅 TS GSMaP = 19 km, which corresponds to the median value of these samples; + otherwise, set 𝑅𝑅𝑅𝑅𝑅𝑅 TS GSMaP using Eq. (1). Figure 1b shows + the estimation error distribution of 𝑅𝑅𝑅𝑅𝑅𝑅 TS GSMaP relative to the SAR-derived RMW. The mean absolute error (MAE) is 14.1 km. After applying the outlier exclusion criterion (±28.8 km), the error distribution fits a normal distribution reasonably well, with a standard deviation of 10.6 km. + The current version of the 𝑅𝑅𝑅𝑅𝑅𝑅 TS GSMaP method does not explicitly account for the poor relationship between 𝑅𝑅mp and the RMW in weak, unorganized, or sheared TCs. Developing a more robust RMW estimation method for such TCs remains an important future task. TD cases tend to have a local wind maximum within regions of strong precipitation. Accordingly, 𝑅𝑅𝑅𝑅𝑅𝑅 TD GSMaP is defined as follows: Set 𝑅𝑅𝑅𝑅𝑅𝑅 TD GSMaP = 𝑅𝑅mp within 300 km, and identify the radius of maximum local precipitation, 𝑅𝑅mp−local . ii) When 𝑅𝑅mp < 10 km, set 𝑅𝑅𝑅𝑅𝑅𝑅 TD GSMaP = 𝑅𝑅mp−local if 𝑅𝑅mp−local ≥ 10 km; otherwise, set 𝑅𝑅𝑅𝑅𝑅𝑅 TD GSMaP = 10 km. iii) When 𝑅𝑅mp = 300 km and 𝑅𝑅mp−local < 300 km, set 𝑅𝑅𝑅𝑅𝑅𝑅 TD GSMaP = 𝑅𝑅mp−local . i) Figure 2 shows a scatter plot of 𝑅𝑅𝑅𝑅𝑅𝑅 TD GSMaP versus ASCATderived RMWs, along with the corresponding error distribution. The estimation errors are relatively large, reflecting not only the weak correlation between 𝑅𝑅𝑅𝑅𝑅𝑅 TD GSMaP and the RMW but also the uncertainties in the ASCAT-derived RMWs as reference data. This level of accuracy is considered unavoidable for TD cases. mined by a threshold value of 28.8 km (red). Sample size (N), median 𝜎𝜎 absolute deviation (MAD), criterion for outlier exclusion (±1.645ˆ MAD), mean absolute error (MAE), and correlation coefficient (Corr) are shown. MAE and Corr are calculated using all samples, including outliers 13 38 Page 6 of 19 Journal of the Meteorological Society of Japan Fig. 2 a Scatter plot of 𝑅𝑅𝑅𝑅𝑅𝑅 TD GSMaP (km) versus ASCAT-derived RMW (km). The blue shading indicates kernel density estimation (×10−5 km–2). b Distribution of 𝑅𝑅𝑅𝑅𝑅𝑅 TD GSMaP errors (km), with Gaussian fits for all samples (gray) and for samples excluding outliers determined by a threshold value of 202.5 km (red). Sample size (N), median 3.3 TD + 𝑹𝑹𝑹𝑹𝑹𝑹TS Rankine and 𝑹𝑹𝑹𝑹𝑹𝑹ERA𝟓𝟓 In RMW estimation methods based on precipitation distributions, center displacement and the weak correspondence between asymmetric precipitation structures and the lowlevel wind structure, particularly in cases exhibiting the shear pattern (Dvorak 1984), can be major sources of error. Therefore, this study develops a method to estimate the RMW from a hypothetical radial wind profile constructed using bias-corrected ERA5 10-m wind fields, JMA Vmax, and + a modified Rankine vortex (𝑅𝑅𝑅𝑅𝑅𝑅 TS Rankine ) for TS strength and above. In addition, a different method (𝑅𝑅𝑅𝑅𝑅𝑅 TD ERA5 ) is developed for cases with Vmax < 20 m s–1 based on ERA5 10-m wind fields. + For 𝑅𝑅𝑅𝑅𝑅𝑅 TS Rankine , the radial profile of the azimuthally averaged wind speed is assumed, based on the theoretical framework of previous studies (Emanuel and Rotunno 2011; Chavas and Lin 2016), to be characterized by three parameters: the RMW, Vmax, and the outer size. Here, the effect of the Coriolis parameter is neglected. Furthermore, the radial profile approximately follows a modified Rankine vortex with an appropriate decay exponent 𝛼𝛼 (Mallen et al. 2005): 𝑉𝑉 (𝑟𝑟) = { 𝑉𝑉max 𝑉𝑉max ( 𝑟𝑟 , 𝑟𝑟 𝑟 𝑟𝑟𝑟𝑟𝑟𝑟 𝑅𝑅𝑅𝑅𝑅𝑅 𝑅𝑅𝑅𝑅𝑅𝑅 𝛼𝛼 𝑟𝑟 ) , 𝑟𝑟 ≥ 𝑅𝑅𝑅𝑅𝑅𝑅 where 𝑉𝑉 (𝑟𝑟) denotes the axisymmetric wind speed at radius TS+ 𝑟𝑟. Based on these assumptions, 𝑅𝑅𝑅𝑅𝑅𝑅 Rankine is derived as follows: 13 (2026) 104:38 𝜎𝜎 absolute deviation (MAD), criterion for outlier exclusion (±1.645ˆ MAD), mean absolute error (MAE), and correlation coefficient (Corr) are shown. MAE and Corr are calculated using all samples, including outliers i) JMA Vmax is used because it is in good agreement with SAR Vmax derived from the azimuthally averaged wind speed (Fig. 3a). ii) Because ERA5 wind speeds (𝑉𝑉ERA5 ) exceeding 12 m s–1 sampled outside the RMW in the ERA5 wind field (𝑅𝑅mERA5 ) have a negative bias (Fig. 3b), the bias is corrected linearly based on the comparison with SAR wind speeds (𝑉𝑉𝑉ERA5 , m s–1) as follows2: 𝑉𝑉𝑉ERA5 = 1.6𝑉𝑉ERA5 − 7.3, (𝑉𝑉ERA5 ≥ 12.2) .(2) The decay exponent α is then determined as: 𝛼𝛼 𝛼 log (𝑉𝑉mERA5 /12) , log (𝑅𝑅12 /𝑅𝑅mERA5 ) where 𝑅𝑅12 denotes the innermost radius outside the RMW at which the azimuthally averaged 𝑉𝑉𝑉ERA5 decreases to 12 m s–1, and 𝑉𝑉mERA5 denotes Vmax derived from the azimuthally averaged 𝑉𝑉𝑉ERA5 . iii) Wind speeds inside 𝑅𝑅mERA5 are extrapolated inward using the modified Rankine vortex profile with α determined in step (ii). The radius at which the extrapolated wind speed reaches JMA Vmax is identified as a candidate RMW (Rcand; Fig. 3c). Rcand implies the following: • Case A: for TCs with a single eyewall, Rcand is expected to approximately correspond to the true RMW; and • Case B: for TCs with concentric eyewalls, Rcand may be located substantially outside the inner eyewall. 2 As of the 2026 season, JMA does not use SAR wind observations to estimate Vmax. Journal of the Meteorological Society of Japan (2026) 104:38 Fig. 3 a Scatter plot of JMA Vmax (m s–1) versus SAR-based Vmax (m s–1) derived from the azimuthally averaged wind speed. The dashed line denotes the 1:1 line. b Two-dimensional histogram of 𝑉𝑉ERA5 and SAR wind speeds sampled outside 𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅. The black line indicates the regression given by Eq. (3). c Example of radial profiles of the azimuthally averaged SAR wind speed (red), 𝑉𝑉ERA5 (green), 𝑉𝑉𝑉ERA5 (blue), and the modified Rankine vortex (black) constructed using α determined from R12, 𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅, and 𝑉𝑉mERA5 . d Scatter plot of + 𝑅𝑅𝑅𝑅𝑅𝑅 TS Rankine (km) versus SAR-derived RMW (km). e Scatter plot of TS+ + the mean of 𝑅𝑅𝑅𝑅𝑅𝑅 TS Rankine and 𝑅𝑅𝑅𝑅𝑅𝑅 GSMaP (km) versus SAR-derived iv) To account for the possibility of Case B, Rcand is accepted + as the RMW only when it is located inside 𝑅𝑅𝑅𝑅𝑅𝑅 TS GSMaP TS+ TS+ . Thus, 𝑅𝑅𝑅𝑅𝑅𝑅 Rankine = 𝑅𝑅cand when 𝑅𝑅cand < 𝑅𝑅𝑅𝑅𝑅𝑅 GSMaP + ; otherwise, 𝑅𝑅𝑅𝑅𝑅𝑅 TS is not defined. If Rankine TS+ + 𝑅𝑅𝑅𝑅𝑅𝑅 Rankine < 16 km, reset 𝑅𝑅𝑅𝑅𝑅𝑅 TS Rankine = 15 km. + Although 𝑅𝑅𝑅𝑅𝑅𝑅 TS Rankine alone does not achieve high accu+ racy (MAE = 16.3 km; Fig. 3d), nor does 𝑅𝑅𝑅𝑅𝑅𝑅 TS GSMaP (MAE TS+ + = 16.5 km), the mean of 𝑅𝑅𝑅𝑅𝑅𝑅 TS Rankine and 𝑅𝑅𝑅𝑅𝑅𝑅 GSMaP yields Page 7 of 19 38 RMW (km). f Distribution of 𝑅𝑅𝑅𝑅𝑅𝑅 TS Rankine errors (km), with Gaussian and log-normal fits. Gaussian fits are derived from all samples (gray) and from samples excluding outliers determined by a threshold value of 31.3 km (red). The log-normal fit (blue) is derived from samples with 𝜀𝜀 < 14.0 km (N = 106). Sample size (N), median absolute devia𝜎𝜎MAD), mean absotion (MAD), criterion for outlier exclusion (±1.645ˆ lute error (MAE), and correlation coefficient (Corr) are shown. MAE and Corr are calculated using all samples, including outliers. In (a–b), the samples are taken from the period 2015–2023, whereas the samples in (d–f) are taken from the period 2015–2024 improved accuracy (MAE = 14.6 km; Fig. 3e). Accordingly, + the inclusion of 𝑅𝑅𝑅𝑅𝑅𝑅 TS Rankine in the consensus framework is + expected to compensate for the weaknesses of 𝑅𝑅𝑅𝑅𝑅𝑅 TS GSMaP . + The estimation error 𝜀𝜀 of 𝑅𝑅𝑅𝑅𝑅𝑅 TS Rankine exhibits a mode on the positive side and a negatively skewed tail in its frequency distribution (Fig. 3f). Therefore, this study adopts a log-normal distribution. The optimal value of 𝑈𝑈 for 13 38 Page 8 of 19 + 𝑅𝑅𝑅𝑅𝑅𝑅 TS Rankine is set to 14 km. Outliers appear on both sides of the error 𝜀𝜀 with approximately uniform frequency. In contrast to TS cases, TD cases generally do not have a modified Rankine vortex structure; therefore, a different approach is required. We found that the ERA5-derived RMWs tend, on average, to be located close to ASCATderived RMWs (Fig. 4a), likely because ERA5 assimilates ASCAT wind observations. In this study, the ERA5-derived RMW, 𝑅𝑅𝑅𝑅𝑅𝑅 TD ERA5 , is defined as follows: i) Identify 𝑅𝑅mERA5 within 300 km of the TC center. ii) When 𝑅𝑅mERA5 ≤ 50 km or 𝑅𝑅mERA5 = 300 km, replace it with the radius of maximum local ERA5 wind speed. If the resulting radius is still less than 50 km, set 𝑅𝑅mERA5 = 50 km. iii) Finally, 𝑅𝑅𝑅𝑅𝑅𝑅 TD ERA5 is obtained by applying a bias correction: 𝑅𝑅𝑅𝑅𝑅𝑅 TD ERA5 = 𝑅𝑅mERA5 − 34 The MAE of 𝑅𝑅𝑅𝑅𝑅𝑅 TD ERA5 is 64.6 km (Fig. 4b), which is relatively large and comparable to that of 𝑅𝑅𝑅𝑅𝑅𝑅 TD GSMaP (Fig. 2). TD However, the consensus of 𝑅𝑅𝑅𝑅𝑅𝑅 TD GSMaP and 𝑅𝑅𝑅𝑅𝑅𝑅 ERA5 reduces the MAE to 58.5 km. 3.4 RMWIR The IR-based method developed by Tsukada and Horinouchi (2023) is adopted with some modifications. In this method, 𝑅𝑅𝑅𝑅𝑅𝑅 IR is estimated from 𝑅𝑅eye , which is defined as the radius of the minimum enclosing circle of the –45°C IR brightness temperature contour. To increase the availability of estimates, this study computes 𝑅𝑅𝑅𝑅𝑅𝑅 IR whenever Fig. 4 a Scatter plot of 𝑅𝑅𝑅𝑅𝑅𝑅 TD ERA5 (km) versus the ASCAT-derived RMW (km). The blue shading indicates kernel density estimation (×10−5 km–2). b Distribution of 𝑅𝑅𝑅𝑅𝑅𝑅 TD ERA5 errors (km), with Gaussian fits for all samples (gray) and for samples excluding outliers determined by a threshold value of 180.8 km (red). Sample size (N), median 13 Journal of the Meteorological Society of Japan (2026) 104:38 the JMA Dvorak technique (Nishimura et al. 2023) indicates an “eye pattern,” rather than restricting the estimation to cases classified as having a “clear eye,” as in Tsukada and Horinouchi (2023). With these modifications, a new regression relationship is derived in this study using SAR-derived RMWs within ± 0.5 h of hourly IR data: 𝑅𝑅𝑅𝑅𝑅𝑅 IR = 0.61𝑅𝑅eye + 7.85 The MAE of 𝑅𝑅𝑅𝑅𝑅𝑅 IR is 4.6 km (Fig. 5a), which is larger than that reported by Tsukada and Horinouchi (2023). The error distribution of 𝑅𝑅𝑅𝑅𝑅𝑅 IR is well represented by a log-normal distribution, similar to that for 𝑅𝑅𝑅𝑅𝑅𝑅 TS Rankine . The optimal value of 𝑈𝑈 for the log-normal distribution of 𝑅𝑅𝑅𝑅𝑅𝑅 IR is 9 km. 3.5 RMWPMW The PMW-based method developed by Shimada (2025) is adopted using the following regression: 𝑅𝑅𝑅𝑅𝑅𝑅 PMW = 0.69𝑅𝑅TB40 + 6.11, where 𝑅𝑅TB40 denotes the radius corresponding to the −40°C azimuthally averaged 85−89-GHz brightness temperature. Although Shimada (2025) proposed an additional estimation method, this study adopts only the above method for cases with JMA best track Vmax > 33 m s–1, as the alterna+ tive method exhibits poorer accuracy than 𝑅𝑅𝑅𝑅𝑅𝑅 TS GSMaP . The MAE of 𝑅𝑅𝑅𝑅𝑅𝑅 PMW is 4.2 km (Fig. 5b). The error distribution of 𝑅𝑅𝑅𝑅𝑅𝑅 PMW is well represented by a log-normal distribution, with 𝑈𝑈 = 7 km. 𝜎𝜎 absolute deviation (MAD), criterion for outlier exclusion (±1.645ˆ MAD), mean absolute error (MAE), and correlation coefficient (Corr) are shown. MAE and Corr are calculated using all samples, including outliers Journal of the Meteorological Society of Japan (2026) 104:38 Page 9 of 19 38 Fig. 5 a As in Fig. 3f, but for 𝑅𝑅𝑅𝑅𝑅𝑅 IR (km). The log-normal fit (blue) is derived from samples with 𝜀𝜀 < 9.0 km (N = 67). b As in Fig. 3f, but for 𝑅𝑅𝑅𝑅𝑅𝑅 PMW (km). The log-normal fit (blue) is derived from samples with 𝜀𝜀 < 7.0 km (N = 123) 4 Consensus estimation 4.1 Bayesian consensus method Bayes' theorem is applied to the consensus framework as follows: 𝑝𝑝 (𝑥𝑥1 , 𝑥𝑥2 𝑥𝑥3 , … , 𝑥𝑥𝑇𝑇 |𝑋𝑋) 𝑝𝑝 (𝑋𝑋) , 𝑝𝑝 (𝑋𝑋𝑋𝑋𝑋1 , 𝑥𝑥2 𝑥𝑥3 , … , 𝑥𝑥𝑇𝑇 ) = 𝑝𝑝 (𝑥𝑥1 , 𝑥𝑥2 𝑥𝑥3 , … , 𝑥𝑥𝑇𝑇 ) where the left-hand side represents the posterior probability density function (PDF) of the RMW, which is denoted by 𝑋𝑋 , given the set of RMW estimates {𝑥𝑥1 , 𝑥𝑥2 , 𝑥𝑥3 , … , 𝑥𝑥𝑇𝑇 }, with 𝑇𝑇 indicating the number of estimates. On the right-hand side, 𝑝𝑝 (𝑥𝑥1 , 𝑥𝑥2 𝑥𝑥3 , … , 𝑥𝑥𝑇𝑇 |𝑋𝑋) denotes the likelihood of the estimates given 𝑋𝑋 , and 𝑝𝑝 (𝑋𝑋) denotes the prior PDF of 𝑋𝑋 . The denominator is a normalization constant. Error PDFs intrinsic to each estimation method are used to compute the likelihood from the multiple estimates. Assuming statistical independence among the estimates, the likelihood can be written as 𝑇𝑇 𝑝𝑝 (𝑥𝑥1 , 𝑥𝑥2 , … , 𝑥𝑥𝑇𝑇 |𝑋𝑋) = ∏ 𝑝𝑝 (𝑥𝑥𝑖𝑖 |𝑋𝑋) . 𝑖𝑖 In this study, the prior PDF of 𝑋𝑋 is assumed to be uniform over 0 < 𝑋𝑋 ≤ 300 km and zero outside this range. Under this assumption, the posterior PDF is proportional to the likelihood: 𝑝𝑝 (𝑋𝑋𝑋𝑋𝑋1 , 𝑥𝑥2 𝑥𝑥3 , … , 𝑥𝑥𝑇𝑇 ) ∝ ∏𝑖𝑖 𝑝𝑝 (𝑥𝑥𝑖𝑖 |𝑋𝑋) , 0 < 𝑋𝑋 ≤ 300km.(3) 𝑇𝑇 Furthermore, in this study, each likelihood 𝑝𝑝 (𝑥𝑥𝑖𝑖 |𝑋𝑋) is assumed to be the sum of an error PDF and an outlier PDF, with the specific formulation given in the next subsection. This formulation allows the likelihood to retain peaks near one or more of the individual estimates (see the next subsection). The final RMW is obtained by evaluating the righthand side of Eq. (3) and is determined as the MAP estimate. 4.2 Likelihoods The posterior PDF of the RMW is estimated with a radial resolution of 1 km by evaluating the right-hand side of TD + Eq. (3). Except for 𝑅𝑅𝑅𝑅𝑅𝑅 TS GSMaP and 𝑅𝑅𝑅𝑅𝑅𝑅 GSMaP , the likelihoods of the RMW estimates are modeled as the sum of an error distribution represented by either a log-normal or a normal distribution with a weight of 90%, and an outlier distribution, assumed to be uniform over the range 0–300 km + with a weight of 10%. The likelihoods of 𝑅𝑅𝑅𝑅𝑅𝑅 TS GSMaP and TD 𝑅𝑅𝑅𝑅𝑅𝑅 GSMaP are described later. For each method, it is not easy to quantitatively estimate the outlier fraction because of the limited sample size. Nevertheless, all methods are expected to include outliers at some rate, as shown in Figs. 1, 2, 3, 4, 5. For an estimate 𝑥𝑥𝑖𝑖 (from 𝑅𝑅𝑅𝑅𝑅𝑅 TS Rankine , 𝑅𝑅𝑅𝑅𝑅𝑅 IR , and 𝑅𝑅𝑅𝑅𝑅𝑅 PMW ) whose errors are modeled by a log-normal distribution, the likelihood 𝑝𝑝 (𝑥𝑥𝑖𝑖 |𝑋𝑋), with method-specific parameters 𝜎𝜎 𝑖𝑖 , 𝜇𝜇𝑖𝑖 , and 𝑈𝑈 𝑖𝑖 (Table 1), is expressed as 𝑝𝑝𝑝𝑝𝑝𝑖𝑖 |𝑋𝑋) = √ 0.9 2𝜋𝜋𝜋𝜋𝑖𝑖 𝑥𝑥𝑥𝑖𝑖 𝑒𝑒𝑒𝑒𝑒𝑒 [− (ln(𝑥𝑥𝑥𝑖𝑖 )−𝜇𝜇𝑖𝑖 ) 2𝜎𝜎𝑖𝑖2 2 ] 1[0,∞] (𝑥𝑥𝑥𝑖𝑖 ) + 0.1 300 1[0,300] (𝑥𝑥𝑖𝑖 ) (4) where 𝑥𝑥𝑥𝑖𝑖 = 𝑈𝑈 𝑖𝑖 − (𝑥𝑥𝑖𝑖 − 𝑋𝑋𝑋, 1 denotes the indicator function, and all radii are expressed in km. For an estimate 𝑥𝑥𝑖𝑖 (from 𝑅𝑅𝑅𝑅𝑅𝑅 TD ERA5 ) whose errors are modeled by a normal distribution, the likelihood 𝑝𝑝 (𝑥𝑥𝑖𝑖 |𝑋𝑋), with a method-specific parameter 𝜎𝜎 𝑖𝑖 (see Table 1), is expressed as 13 38 Page 10 of 19 𝑝𝑝 (𝑥𝑥𝑖𝑖 |𝑋𝑋) = √ 0.9 2𝜋𝜋𝜋𝜋𝑖𝑖 𝑒𝑒𝑒𝑒𝑒𝑒 [− Journal of the Meteorological Society of Japan (𝑥𝑥𝑖𝑖 −𝑋𝑋) 2𝜎𝜎𝑖𝑖2 2 ]+ 0.1 300 1[0,300] (𝑥𝑥𝑖𝑖 ) (5) where all radii are expressed in km. The bias is assumed to be zero for simplicity. By explicitly assuming an outlier distribution, the posterior distribution can be multimodal, which helps avoid selecting a final RMW that is not supported by any individual method. To demonstrate this advantage, Fig. 6 shows an example of two posterior probability distributions obtained with and without an outlier distribution. Here, we suppose that there are four RMW estimates (22, 24, 58, and 62 km) forming two distinct clusters, with standard deviations of 5, 5, 10, and 10 km, respectively, for their error distributions. For simplicity, the error distributions are assumed to be normal distributions. Without an outlier distribution in the likelihood, the estimated posterior distribution is also a normal distribution, and the MAP estimate (30.4 km) lies between the two distinct clusters. In contrast, with an outlier distribution, the posterior distribution can be multimodal, with the MAP estimate (24.7 km) located close to one of the clusters. TD + For the likelihoods of 𝑅𝑅𝑅𝑅𝑅𝑅 TS GSMaP and 𝑅𝑅𝑅𝑅𝑅𝑅 GSMaP , additional modifications are introduced. In the estima+ TD tion of 𝑅𝑅𝑅𝑅𝑅𝑅 TS GSMaP and 𝑅𝑅𝑅𝑅𝑅𝑅 GSMaP , a single radius must (2026) 104:38 be selected as 𝑅𝑅mp [Eq. (1)]. However, in cases with concentric eyewalls, pronounced asymmetries, or broad radial wind profiles, potential RMWs may be distributed over a wider range of radii than in single-eyewall cases. In such + situations, the error distribution term of the 𝑅𝑅𝑅𝑅𝑅𝑅 TS GSMaP and 𝑅𝑅𝑅𝑅𝑅𝑅 TD GSMaP likelihoods cannot be assumed to follow a normal distribution and is therefore constructed instead from the radial distribution of azimuthally averaged precipitation normalized by its maximum value (Fig. 7a). Since + 𝑅𝑅𝑅𝑅𝑅𝑅 TS GSMaP is originally derived from the radius of maximum precipitation, the normalized precipitation is relocated (Fig. 7b) using an equation consistent with Eq. (1): 𝑟𝑟𝑟 𝑟 𝑟𝑟𝑟𝑟𝑟𝑟 𝑟 𝑟𝑟𝑟𝑟, where 𝑟𝑟𝑟 is the relocated radius. The peak of the relocated precipitation is then enhanced to obtain the final distribution (Fig. 7c). More specific procedures for the + TD error distributions of 𝑅𝑅𝑅𝑅𝑅𝑅 TS GSMaP and 𝑅𝑅𝑅𝑅𝑅𝑅 GSMaP are given TS+ in Appendix B. Finally, the 𝑅𝑅𝑅𝑅𝑅𝑅 GSMaP and 𝑅𝑅𝑅𝑅𝑅𝑅 TD GSMaP likelihoods also include a uniform outlier distribution, as in Eq. (4). 4.3 Estimation procedure As consensus members, all RMW estimates available at each analysis time are used. In addition, when 𝑅𝑅𝑅𝑅𝑅𝑅 IR or 𝑅𝑅𝑅𝑅𝑅𝑅 PMW , which has high accuracy but limited availability, is not available at a given analysis time, (i) the estimate from 3 h prior to the analysis time, or otherwise (ii) the estimate from 3 h after the analysis time, is included as a consensus member, if available. Consequently, the same 𝑅𝑅𝑅𝑅𝑅𝑅 IR or 𝑅𝑅𝑅𝑅𝑅𝑅 PMW estimates can be used in the consensus for up to three consecutive analysis times. After the posterior distribution is computed, a 1–1-1 filter is applied to it in both the radial (1-km resolution) and temporal (3-h resolution) directions to suppress irregular fluctuations (Fig. 8). The filter is applied iteratively five times, as determined empirically, enabling the RMW to evolve smoothly over time (Fig. 8b). The final RMW is determined by the MAP estimate [i.e., arg max𝑝𝑝 (𝑋𝑋𝑋𝑋𝑋1 , 𝑥𝑥2 𝑥𝑥3 , … , 𝑥𝑥𝑇𝑇 )]. However, spike-like fluctua𝑋𝑋 tions occasionally remain in the temporal evolution of the final RMW. Such residual spikes are identified and corrected at each analysis time t as follows. Define ΔRMW1 = RMW(t) – RMW(t−3h) and ΔRMW2 = RMW(t+3h) – RMW(t). If |ΔRMW1|> 10 km, |ΔRMW2|> 10 km, ΔRMW1*ΔRMW2 < 0, and min ( , |Δ𝑅𝑅𝑅𝑅𝑅𝑅𝑅| |Δ𝑅𝑅𝑅𝑅𝑅𝑅𝑅| |Δ𝑅𝑅𝑅𝑅𝑅𝑅𝑅| |Δ𝑅𝑅𝑅𝑅𝑅𝑅𝑅| ) > 0.2, then the RMW(t) is replaced by the average of RMW(t−3h) and RMW(t+3h). This procedure is applied iteratively three times. Fig. 6 (top) Posterior probability distributions obtained with and without an outlier distribution. (bottom) Individual error distributions corresponding to the RMW estimates 13 Journal of the Meteorological Society of Japan (2026) 104:38 Page 11 of 19 38 + Fig. 7 Hovmöller diagrams of a linearly normalized precipitation, b relocated precipitation, and c the error distribution term of the 𝑅𝑅𝑅𝑅𝑅𝑅 TS GSMaP likelihood for Typhoon Mawar (2023) Fig. 8 Hovmöller diagrams of the posterior probability distribution of the RMW a before and b after smoothing for Typhoon Mawar (2023) 13 38 Page 12 of 19 5 Verification Verification of the consensus RMW estimates is conducted using matchup samples with SAR observations. Since SARderived RMWs during the period 2015–2024 are used to develop the individual RMW estimation methods in Sect. 3, the matchup samples are divided into dependent samples from 2015–2024 and independent samples from 2025. Here, SAR scenes with concentric eyewalls are included. Two case studies are then presented. Figures 9a and b show the bias of the consensus RMW estimates as a function of JMA Vmax and peak posterior probability, respectively. There is little difference in the error distributions between dependent and independent samples. Errors decrease with increasing Vmax and with increasing posterior probability. Weak TCs tend to have poorly defined structures, making RMW estimation difficult and causing larger estimation errors for Vmax < 35 m s–1. Additionally, seven of the nine samples with Vmax > 40 m s–1 and large errors (> 30 km) are associated with concentric eyewalls. For such cases, large errors can occur if the estimated RMW is associated with the inner eyewall while the true RMW corresponds to the outer eyewall, or vice versa. A more distinct relationship is found between the error and the posterior probability, with errors being smaller for posterior probabilities > 0.04. Therefore, the posterior probability itself can serve as an indicator of the reliability of the RMW estimates. The MAE calculated using both dependent and independent samples is 13.6 km (Fig. 9c). Although the errors are widely distributed, 62% of the samples have absolute errors of ≤ 10 km (not shown). A negative bias is evident, particularly for 25 m s–1 ≤ Vmax < 40 m s–1 (Fig. 9d) and for posterior probabilities < 0.045 (Fig. 9e). This negative bias is mainly + attributable to the Vmax-dependent bias of 𝑅𝑅𝑅𝑅𝑅𝑅 TS GSMaP : a negative bias with many negative outliers for Vmax of 25–35 m s–1, and a positive bias for Vmax greater than 50 m s–1 (not shown), although its overall bias is very small (–0.2 km). In the consensus estimation, estimates including 𝑅𝑅𝑅𝑅𝑅𝑅 IR and 𝑅𝑅𝑅𝑅𝑅𝑅 PMW help reduce the positive bias, whereas no other estimates are generally available to reduce the negative bias and outliers. Consequently, the negative bias and outliers remain in the consensus estimates, resulting in an overall negative bias. This issue will be addressed in future work as more SAR samples for weak TCs become available. Figure 10 shows the temporal evolution of the consensus RMWs together with the individual estimates for Typhoon Mawar (2023), as a representative case of an intense TC. + The consensus primarily follows 𝑅𝑅𝑅𝑅𝑅𝑅 TS GSMaP , with appropriate corrections from 𝑅𝑅𝑅𝑅𝑅𝑅 IR and 𝑅𝑅𝑅𝑅𝑅𝑅 PMW when available. The comparison with the SAR-derived RMWs shows that the consensus successfully determines appropriate 13 Journal of the Meteorological Society of Japan (2026) 104:38 RMW estimates. Figure 8b shows a Hovmöller diagram of the posterior probability distribution for Mawar. Mawar is characterized by sharply peaked distributions during periods with a well-defined eye, whereas broader distributions appear during its initial and decaying stages. The rapid outward shift of the radius of maximum posterior probability suggests that eyewall replacement cycles (ERCs) occurred at least twice, at ~1800 UTC 23 May and ~1200 UTC 26 May. Furthermore, the posterior distribution agrees closely with SAR-derived RMWs when the distribution is sharply peaked, whereas discrepancies increase when it is broad. Therefore, the posterior distribution obtained by the consensus method is useful not only for RMW estimation but also for assessing its reliability and diagnosing ERCs. Figure 11 shows Tropical Storm Barijat (2024) as a representative case of a weak TC. A spike-like fluctuation in the MAP estimate at 1800 UTC 8 October is corrected by the post-processing procedure (Fig. 11a). In general, RMW estimation is difficult for weak TCs (e.g., Shimada 2025). This case exhibits a broad posterior distribution (Fig. 11b), suggesting large estimation uncertainty. Because the ERA5 wind field of Barijat exhibits a small-sized vortex throughout its life cycle (not shown), 𝑅𝑅𝑅𝑅𝑅𝑅 TS Rankine remains small as well, mostly below 30 km. There are two noticeable features of the RMW estimates for Barijat. First, the RMW is estimated to be less than 30 km during the initial TD period, which is smaller than that during the TS period. The ASCAT wind field shows a small vortex with a possible RMW of less than 50 km (Fig. 11c), indicating that the consensus RMW estimates are reasonable. Second, the consensus RMW reaches 300 km (the maximum possible value in this method) on 8 October. The ASCAT observations and visible imagery show a small low-level vortex without deep convection in its vicinity, along with a relatively strong wind region on the eastern side of the TC (Fig. 11d), representing a typical “shear pattern” in the Dvorak technique. 𝑅𝑅𝑅𝑅𝑅𝑅 TS Rankine corresponds to the RMW of the small-sized low-level vortex, + whereas 𝑅𝑅𝑅𝑅𝑅𝑅 TS GSMaP corresponds to the strong wind region on the eastern side. These two potential RMWs are appropriately reflected in the posterior distribution. 6 Long-term dataset A long-term RMW dataset is constructed for TCs that occurred in the western North Pacific since 2000, because GSMaP data with homogeneous quality have been available from that year onward. A total of 631 TCs during the period 2000–2025 are included in the dataset. The RMW is estimated at 3-h intervals for periods during which the JMA best track data classify the TCs as TDs or stronger. The dataset includes the consensus RMW, the posterior probability Journal of the Meteorological Society of Japan (2026) 104:38 Page 13 of 19 38 Fig. 9 Relationships between a bias (consensus minus SAR-derived RMW; km) and JMA Vmax (m s–1) and b bias (km) and posterior probability. Black circles denote dependent samples, and red circles denote independent samples. c Scatter plot of consensus RMW (km) versus SAR-derived RMW (km). The dashed line denotes the 1:1 line. Sample size (N), mean absolute error (MAE), root-mean-square error (RMSE), bias, correlation coefficient (Corr), and coefficient of determination R2 are shown. (d–e) Boxplots of absolute errors stratified by JMA Vmax (m s–1) and posterior probability, respectively. A cross denotes the MAE for each group, and circles denote outliers. Sample size (N) and bias are shown at the RMW, the posterior probability distribution, as well as the RMW estimates from the available methods (Shimada and Tsukada 2026). Figure 12 shows the two-dimensional histogram of the JTWC best track Vmax and the consensus RMW for the period 2001–2025, in comparison with the corresponding JTWC best track RMW3 (Howell et al. 2025). The consensus RMW dataset shows a frequency distribution that differs from the JTWC RMW dataset. In particular, pronounced differences are found in two regions: area I, defined by Vmax = 10–40 m s–1 and RMW = 12.5–35.0 km, and area II, defined by Vmax = 40–60 m s–1 and RMW = 60–120 km. Area I is 3 JTWC best track RMWs have been included in the dataset since 2001. 13 38 Page 14 of 19 Fig. 10 Temporal evolution of the consensus RMW estimates (thick black line) together with the individual estimates for Typhoon Mawar (2023). White circles denote SAR-derived RMWs as reference Fig. 11 a As in Fig. 10, but for Tropical Storm Barijat (2024). Vertical lines indicate the times shown in (c) and (d). The Vmax of 18 m s–1 at 0600 UTC 5 October results from extrapolation during the TD stage (see Sect. 2). b Hovmöller diagram of the posterior probability distribution of the RMW. The thick black line denotes the radius of maximum posterior probability. (c–d) ASCAT wind fields at ~2330 UTC 4 13 Journal of the Meteorological Society of Japan (2026) 104:38 characterized by the absence of a distinct frequency peak in the JTWC RMW dataset, whereas a clear frequency peak appears in each Vmax bin in the consensus RMW dataset. The presence of these frequency peaks is consistent with SAR observations for cases with Vmax > 20 m s–1, although further verification is needed because the consensus RMWs for weak storms tend to have negative biases (Fig. 9d). Area II is characterized by very few samples in the JTWC RMW dataset, but by moderate frequencies in the consensus RMW dataset, in agreement with SAR observations. In reality, cases with Vmax ≈ 50 m s–1 and RMW ≈ 100 km have been observed (e.g., Typhoon Trami 2018; Hirano et al. 2022), suggesting that the consensus RMW dataset is capable of realistically representing a wide range of RMW values. Figure 13 compares the bias and absolute error of the JTWC and consensus RMWs against SAR-derived RMWs. The consensus RMW dataset generally has a narrower range of boxplot whiskers for both bias and absolute error. In particular, both the mean and median of the absolute error are smaller for the consensus RMWs than for the JTWC RMWs, indicating that the consensus RMW dataset has October and ~2310 UTC 7 October, respectively, with the corresponding visible imagery overlaid. The yellow circles indicate a radius of 50 km in c, and radii of 100 km and 300 km in d. In (d), the TC center in the ASCAT wind field appears displaced from that in the visible imagery, possibly due to the coarse spatial resolution of ASCAT (25 km) Journal of the Meteorological Society of Japan (2026) 104:38 Fig. 12 Two-dimensional histograms of 6-hourly (a) JTWC Vmax (m s–1) versus consensus RMW (km) and (b) JTWC Vmax (m s–1) versus JTWC best track RMW (km) during 2001–2025. White circles Page 15 of 19 38 denote SAR-derived RMWs and their corresponding JTWC Vmax values during 2015–2025. Black and red boxes indicate areas I and II, respectively better accuracy. Note that the JTWC RMW values included in the best track dataset are not reanalyzed after the season and have been estimated without SAR observations, except in very recent years (Howell et al. 2025). Furthermore, very small and large RMW estimates tend to be avoided for use in TC vitals and guidance suites. These factors likely contribute to the large difference in RMWs between the consensus and JTWC datasets. The new RMW dataset may alter the climatology of TC RMWs in the western North Pacific. In particular, since this dataset extends back to the 2000s, a period during which SAR observations were unavailable, the representation of small RMWs, which may not have been captured in previous datasets and studies, is expected to be substantially improved. A detailed evaluation and further analyses will be conducted in a future study. 7 Summary Fig. 13 Boxplots of a bias and b absolute error of JTWC and consensus RMWs against SAR-derived RMWs, stratified by Saffir-Simpson hurricane wind scale categories based on JTWC Vmax. A cross denotes the MAE for each group, and circles denote outliers. Sample size (N) is shown below the category name. The numbers indicated in teal represent out-of-range outliers beyond ±100 km Given the importance of RMW information for a wide range of applications, there is a need to develop new techniques and to construct a long-term, homogeneous RMW dataset (Rogers 2023; Ito et al. 2025). Although multiple RMW estimates using different methods have been available for the same analysis time based on previous studies, no consensus method suitable for RMW estimation has been developed to obtain a single RMW estimate from multiple estimates. To this end, this study develops a consensus method based on a Bayesian framework. In the consensus approach, the posterior probability distribution of the RMW given multiple RMW estimates is estimated using likelihoods 13 38 Page 16 of 19 Journal of the Meteorological Society of Japan constructed from method-specific error distributions combined with an outlier distribution, and the final RMW is determined as the MAP estimate. By explicitly accounting for an outlier distribution, the consensus method can appropriately determine a reasonable RMW close to one of the clusters even when the estimates split into two distinct clusters. Using this method, we construct a long-term RMW dataset at 3-h intervals for 631 TCs in the western North Pacific during 2000–2025. As consensus members, the input estimates include those from methods using a satellite-based + TD global precipitation dataset (𝑅𝑅𝑅𝑅𝑅𝑅 TS GSMaP and 𝑅𝑅𝑅𝑅𝑅𝑅 GSMaP ), methods based on IR and PMW satellite data when an eye is observed (𝑅𝑅𝑅𝑅𝑅𝑅 IR and 𝑅𝑅𝑅𝑅𝑅𝑅 PMW ), a modified Rankine + vortex method (𝑅𝑅𝑅𝑅𝑅𝑅 TS Rankine ), and an ERA5-based method (𝑅𝑅𝑅𝑅𝑅𝑅 TD ERA5 ). Verification of the consensus RMW estimates shows that errors decrease with increasing TC intensity and posterior probability, and that the posterior probability itself serves as an indicator of estimate reliability. This new RMW dataset enables a more realistic representation of the inner-core structure, particularly for small RMWs that were difficult to estimate without aircraft observations before the advent of SAR-derived winds. However, developing a more robust RMW estimation method for weak TCs remains an important future task. Because the proposed Bayesian consensus method allows the incorporation of any available estimates with their associated error characteristics, observation-derived RMWs, such as those obtained from SAR, ASCAT, and L-band radiometers, can also be included in the consensus framework, in addition to the RMW estimates used in this study, if longterm homogeneity of the final RMW estimates is not strictly required. Incorporating such additional RMWs would likely further reduce the overall estimation error compared with that of the dataset constructed in this study and would be useful for operational applications. Although this study focuses on TCs in the western North Pacific, the proposed framework can be readily applied to TCs worldwide. This extension is left for future work. The proposed framework provides a practical basis for improving RMW estimation and supporting both research and operational applications. Appendix A List of the notation used in this paper. Notation 𝑅𝑅12 𝑅𝑅TB40 13 Description Radius of the 12 m s–1 azimuthally averaged𝑉𝑉𝑉ERA5 Radius of the azimuthally averaged brightness temperature at −40 °C Notation (2026) 104:38 Description Candidate RMW 𝑅𝑅cand 𝑅𝑅eye Radius of the minimum enclosing circle of the –45 °C IR brightness temperature contour RMW derived from the ERA5 wind field 𝑅𝑅mERA5 𝑅𝑅mp Radius of maximum azimuthally averaged precipitation within 300 km of the TC center The first, second, and third peaks of the azimuthally averaged precipitation beyond a radius of 10 km Radius of maximum local precipitation 𝑅𝑅mp1 ,𝑅𝑅mp2 ,𝑅𝑅mp3 𝑅𝑅mp−local 𝑅𝑅𝑅𝑅𝑅𝑅 TD ERA5 RMW estimate using ERA5 wind data for TD strength RMW estimate using GSMaP data for TD strength RMW estimate using GSMaP data for TS strength and above RMW estimate using IR satellite data 𝑅𝑅𝑅𝑅𝑅𝑅 TD GSMaP + 𝑅𝑅𝑅𝑅𝑅𝑅 TS GSMaP 𝑅𝑅𝑅𝑅𝑅𝑅 IR 𝑅𝑅𝑅𝑅𝑅𝑅 PMW RMW estimate using PMW satellite data + 𝑅𝑅𝑅𝑅𝑅𝑅 TS Rankine RMW estimate using a modified Rankine vortex profile and ERA5 wind data for TS strength and above ERA5 10-m wind speeds 𝑉𝑉ERA5 Bias-corrected𝑉𝑉ERA5 𝑉𝑉𝑉ERA5 𝑉𝑉mERA5 Vmax derived from the azimuthally averaged𝑉𝑉𝑉ERA5 Appendix B + The error distribution term of the 𝑅𝑅𝑅𝑅𝑅𝑅 TS GSMaP and TD 𝑅𝑅𝑅𝑅𝑅𝑅 GSMaP likelihoods is specifically derived as follows: i) Linearly normalize the 12-h running mean precipitation in step (i) of Sect. 3.2 by its maximum value within a radius of 400 km from the TC center at each time (Fig. 7a). ii) Relocate the normalized precipitation using an equation consistent with Eq. (1): 𝑟𝑟𝑟 𝑟 𝑟𝑟𝑟𝑟𝑟𝑟 𝑟 𝑟𝑟𝑟𝑟 where 𝑟𝑟𝑟 is the relocated radius (Fig. 7b). For example, precipitation originally located at 𝑟𝑟 = 100 km is relocated to 𝑟𝑟𝑟 = 63.68 km. iii) Further normalize the precipitation relocated in step (ii) to follow a normal-distribution-shaped function defined as 𝑃𝑃𝑃𝑃nml = exp [− (1 − 𝑃𝑃𝑃𝑃𝑃) ] 2 2𝜎𝜎nml 2 Journal of the Meteorological Society of Japan (2026) 104:38 where 𝑃𝑃𝑃𝑃𝑃 is the normalized precipitation derived in step (i), and 𝜎𝜎nml = 0.3. The value of 𝜎𝜎nml is chosen so that the maximum likelihood derived in step (v) becomes comparable to those of other methods (0.03–0.1). This normalization is necessary because the 𝑃𝑃𝑃𝑃𝑃 distribution yields a maximum likelihood of only ~0.01, which is much smaller than those of other methods (Figs. 3f and 5). iv) Apply a correction to the 𝑃𝑃𝑃𝑃nml distribution to address + the constraint that 𝑅𝑅𝑅𝑅𝑅𝑅 TS GSMaP is set to 19 km when Rmp < 16 km (Sect. 3.2). Here, the derivation of Rmp follows the same procedures as in Sect. 3.2, except for step (iii), which is redefined as follows: When Rmp < 10 km, compute the radius of maximum azimuthally averaged precipitation prior to applying a 12-h running pre pre mean (𝑅𝑅mp ), and set Rmp to the smaller of Rmp1 and 𝑅𝑅mp that lies between 10 and 40 km; if no such value exists, Rmp remains unchanged. Then, the 𝑃𝑃𝑃𝑃nml distribution is corrected as follows: When Rmp < 16 km: (𝑟𝑟𝑟𝑟𝑟𝑟 ⎧ exp [− 2 ] , 𝑟𝑟 ≤ 19𝑘𝑘𝑘𝑘 { 2𝜎𝜎 𝑃𝑃𝑃𝑃nml = (𝐵𝐵𝐵𝐵 2 (𝑟𝑟𝑟𝑟𝑟) ⎨ { max (exp [− 2 ] , 𝑃𝑃𝑃𝑃nml ) , 𝑟𝑟 𝑟 𝑟𝑟𝑟𝑟𝑟𝑟 2𝜎𝜎 ⎩ 2 where 𝜇𝜇 = 19 km and 𝜎𝜎 = 10.617, corresponding to the + standard deviation of the 𝑅𝑅𝑅𝑅𝑅𝑅 TS GSMaP error distribution (Table 1). When Rmp≥ 16 km: 𝑃𝑃𝑃𝑃nml = { 𝑦𝑦0 exp [− (𝑟𝑟𝑟𝑟𝑟𝑟 2𝜎𝜎 2 2 ] , 𝑟𝑟 ≤ 15𝑘𝑘𝑘𝑘 (𝐵𝐵𝐵𝐵 𝑃𝑃𝑃𝑃nml , 𝑟𝑟 𝑟 𝑟𝑟𝑟𝑟𝑟𝑟 where 𝑦𝑦0 = 𝑃𝑃𝑃𝑃nml evaluated at r = 14 km in step (iii), 𝜇𝜇 = 14 km, and 𝜎𝜎 is the same as that defined above. v) Divide the 𝑃𝑃𝑃𝑃nml distribution by its sum over the range 0–300 km to obtain the error distribution term of the + 𝑅𝑅𝑅𝑅𝑅𝑅 TS GSMaP likelihood (Fig. 7c). For the error distribution term of the 𝑅𝑅𝑅𝑅𝑅𝑅 TD GSMaP likelihood, step (i) above is applied to obtain 𝑃𝑃𝑃𝑃nml . Steps (ii) and (iii) are skipped. In step (iv), when Rmp < 20 km, the same correction as Eq. (B1) is applied. When Rmp≥ 20 km, a correction based on Eq. (B2) with 𝜇𝜇 = 19 km is applied for 𝑟𝑟 ≤ 19km. Step (v) is then applied. The maximum likelihood is less than ~0.01, which is comparable to that of 𝑅𝑅𝑅𝑅𝑅𝑅 TD ERA5 (not shown). Acknowledgements U. Shimada is deeply grateful to his colleagues at MRI and JMA Headquarters for their encouragement and fruitful discussions. The insightful comments and suggestions from Drs. Chris Velden and Chris Slocum are greatly appreciated. The views expressed in this paper are those of the authors and should not be regarded as official JMA views. U. Shimada was supported by the Japan Society Page 17 of 19 38 for the Promotion of Science (JSPS) KAKENHI Grant 23K13172 and the Japan Science and Technology Agency (JST) Moonshot JPMJMS2282. This paper was also supported by the collaborative research program (2025GC-06) of the Disaster Prevention Research Institute of Kyoto University. T. Tsukada was supported by the Cooperative Institute for Research in the Atmosphere (CIRA) and National Oceanic and Atmospheric Administration (NOAA) under Cooperative Agreement NA24OARX432C0007 and by the Office of Naval Research under Grant N000142512067. Author contributions Both authors contributed to the conception and design of the study. U.S. performed data collection and analysis, wrote the main text of the manuscript, and prepared the figures and the table. T.T. wrote some parts of the manuscript and prepared a few figures. All authors reviewed the manuscript. Funding Japan Society for the Promotion of Science,KAKENHI Grant 23K13172,Japan Science and Technology Agency,Moonshot JPMJMS2282,Cooperative Institute for Research in the Atmosphere (CIRA) and National Oceanic and Atmospheric Administration (NOAA),NA24OARX432C0007 Data availability GSMaP data are available at ​h​t​t​​p​s​:​/​​/​s​h​​a​r​a​​k​u​.​e​o​r​c​.​j​a​x​ a​.​j​p​/​G​S​M​a​P​/​i​n​d​e​x​.​h​t​m​​​​.​ ERA5 data are available from the Copernicus Climate Change Service (C3S) Climate Data Store (CDS; Hersbach et al. 2023). The Tropical Cyclone Precipitation, Infrared, Microwave, and Environmental Dataset (TC PRIMED) is available from the NOAA National Centers for Environmental Information (NCEI; Razin et al. 2023a). Information on the availability of successive geostationary IR satellite data is provided at ​h​t​t​​p​s​:​/​​/​w​w​​w​.​d​​a​t​a​​.​j​m​​a​.​g​o​​.​j​​p​/​m​s​c​w​e​ b​/​e​n​/​p​r​o​d​u​c​t​/​l​i​b​r​a​r​y​_​d​a​t​a​.​h​t​m​l​​​​.​ SAR wind data are available from the website of STAR (2026). ASCAT wind data are available from NASA (2026). JMA and JTWC best track data are obtained from the International Best Track Archive for Climate Stewardship (IBTrACS) Version 4r01 dataset available at ​h​t​t​​p​s​:​/​​/​w​w​​w​.​n​​c​e​i​​.​n​o​​a​a​.​g​​o​v​​/​p​r​o​d​u​c​t​s​/​i​n​t​e​r​n​a​t​i​ o​n​a​l​-​b​e​s​t​-​t​r​a​c​k​-​a​r​c​h​i​v​e​​​​.​ The RMW dataset constructed in this study is available at https:/​/doi.or​g/10.52​81/ze​nodo.23131986. Declarations Conflict of interest The authors declare that they have no competing interests. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit ​h​t​t​​p​:​/​/​​c​r​e​​a​t​i​​v​e​c​o​m​m​o​n​s​.​o​ r​g​/​l​i​c​e​n​s​e​s​/​b​y​/​4​.​0​/​​​​.​​​ References Aijaz S, Kepert JD, Ye H, Huang Z, Hawksford A (2019) Bias correction of tropical cyclone parameters in the ECMWF ensemble 13 38 Page 18 of 19 prediction system in Australia. Mon Weather Rev 147:4261– 4285. https:/​/doi.or​g/10.11​75/MW​R-D-18-0377.1 Aizawa M, Ito K, Shimada U (2024) Revisiting Koba’s relationship to improve minimum sea-level pressure estimates of western North Pacific tropical cyclones. J Meteorol Soc Japan 102:377–390. https:/​/doi.or​g/10.21​51/jm​sj.2024-018 Avenas A, Mouche A, Tandeo P, Piolle J, Chavas D, Fablet R, Knaff J, Chapron B (2023) Reexamining the estimation of tropical cyclone radius of maximum wind from outer size with an extensive synthetic aperture radar dataset. Mon Weather Rev 151:3169–3189. https:/​/doi.or​g/10.11​75/MW​R-D-22-0277.1 Avenas A, Tsukada T, Cévaër T, Knaff J, Mouche A, Chapron B (2026) The global kilometric tropical cyclone wind database (CyclObs) enabling high-resolution data-driven studies. Bull Am Meteorol Soc. https:/​/doi.or​g/10.11​75/MW​R-D-22-0277.1 Bessho K, Date K, Hayashi M, Ikeda A, Imai T, Inoue H, Kumagai Y, Miyakawa T, Murata H, Ohno T, Okuyama A, Oyama R, Sasaki Y, Shimizu Y, Shimoji K, Sumida Y, Suzuki M, Taniguchi H, Tsuchiyama H, Uesawa D, Yokota H, Yoshida R (2016) An introduction to Himawari-8/9—Japan’s new-generation geostationary meteorological satellites. J Meteorol Soc Japan 94:151–183. https:/​/doi.or​g/10.21​51/jm​sj.2016-009 Black ML, Willoughby HE (1992) The concentric eyewall cycle of Hurricane Gilbert. Mon Weather Rev 120:947–957. ​h​t​t​​p​s​:​/​​/​d​o​​i​.​ o​​r​g​/​1​0​.​1​1​7​5​/​1​5​2​0​-​0​4​9​3​(​1​9​9​2​)​1​2​0​%​3​c​0​9​4​7​:​T​C​E​C​O​H​%​3​e​2​.​0​.​C​ O​;​2​​​​ Carrasco C, Landsea C, Lin Y (2014) The influence of tropical cyclone size on its intensification. Weather Forecast 29:582–590. ​h​t​t​​p​s​:​/​​/​d​ o​​i​.​o​​r​g​/​1​0​.​1​1​7​5​/​W​A​F​-​D​-​1​3​-​0​0​0​9​2​.​1​​​​ Chavas DR, Knaff JA (2022) A simple model for predicting the tropical cyclone radius of maximum wind from outer size. Weather Forecast 37:563–579. https:/​/doi.or​g/10.11​75/WA​F-D-21-0137.1 Chavas DR, Lin N (2016) A model for the complete radial structure of the tropical cyclone wind field. Part II: Wind field variability. J Atmos Sci 73:3093–3113. ​h​t​t​​p​s​:​/​​/​d​o​​i​.​o​​r​g​/​1​0​.​1​1​7​5​/​J​A​S​-​D​-​1​5​-​0​ 0​1​4​.​1​​​​ Chen B-F, Chen C, Hsiao H, Teng C, Lee H (2025) Constructing deep learning datasets to reveal climate trends of tropical cyclone intensity and structure extremes. Artif Intell Earth Syst 4:e240064. https:/​/doi.or​g/10.11​75/AI​ES-D-24-0064.1 Combot C, Mouche A, Knaff J, Zhao Y, Zhao Y, Vinour L, Quilfen Y, Chapron B (2020) Extensive high-resolution synthetic aperture radar (SAR) data analysis of tropical cyclones: comparisons with SFMR flights and best track. Mon Weather Rev 148:4545–4563. https:/​/doi.or​g/10.11​75/MW​R-D-20-0005.1 DeMaria M, Franklin JL, Chirokova G, Radford J, DeMaria R, Musgrave KD, Ebert-Uphoff I (2025) An operations-based evaluation of tropical cyclone track and intensity forecasts from artificial intelligence weather prediction models. Artif Intell Earth Syst 4:e240085. https:/​/doi.or​g/10.11​75/AI​ES-D-24-0085.1 Dvorak VF (1984) Tropical cyclone intensity analysis using satellite data. NOAA Tech Rep 11 Emanuel K, Rotunno R (2011) Self-stratification of tropical cyclone outflow. Part I: implications for storm structure. J Atmos Sci 68:2236–2249. https:/​/doi.or​g/10.11​75/JA​S-D-10-05024.1 Hersbach H et al (2020) The ERA5 global reanalysis. Q J R Meteorol Soc 144:1999–2049. https:/​/doi.or​g/10.10​02/qj​.3803 Hersbach H, Bell B, Berrisford P, Biavati G, Horányi A, Muñoz Sabater J, Nicolas J, Peubey C, Radu R, Rozum I, Schepers D, Simmons A, Soci C, Dee D, Thépaut J-N (2023) ERA5 monthly averaged data on pressure levels from 1940 to present. ​h​t​t​​p​s​:​/​​/​d​o​​i​.​ o​​r​g​/​1​0​.​2​4​3​8​1​/​c​d​s​.​6​8​6​0​a​5​7​3​​​​. Copernicus Climate Change Service (C3S) Climate Data Store (CDS) Hirano S, Ito K, Yamada H, Tsujino S, Tsuboki K, Wu C-C (2022) Deep eye clouds in Tropical Cyclone Trami (2018) during 13 Journal of the Meteorological Society of Japan (2026) 104:38 T-PARCII dropsonde observations. J Atmos Sci 79:683–703. ​h​t​ t​​p​s​:​/​​/​d​o​​i​.​o​​r​g​/​1​0​.​1​1​7​5​/​J​A​S​-​D​-​2​1​-​0​1​2​9​.​1​​​​ Howell B, Kucas M, Cowan L, Strahl B, Howard A, Sampson CR (2025) The Joint Typhoon Warning Center tropical cyclone best tracks, 1945–2025. NRL Formal Rep NRL/7550/FR--2025/3. ​h​t​ t​​p​s​:​/​​/​w​w​​w​.​m​​e​t​o​​c​.​n​​a​v​y​.​​m​i​​l​/​j​t​w​c​/​p​r​o​d​u​c​t​s​/​b​e​s​t​-​t​r​a​c​k​s​/​J​T​W​C​_​T​C​_​ B​e​s​t​_​T​r​a​c​k​s​_​N​R​L​_​R​e​p​o​r​t​.​p​d​f​​​​.​ Accessed 4 Mar 2026 Ito K, Miyamoto Y, Wu C-C, Didlake A, Hlywiak J, Huang Y-H, Lai T-K, Pattie L, Qin N, Shimada U, Tao D, Yamada Y, Zhang J, Kanada S, Herndon D (2025) Recent research and operational tools for improved understanding and diagnosis of tropical cyclone inner core structure. J Meteorol Soc Japan 103:147–180. https:/​/doi.or​g/10.21​51/jm​sj.2025-008 Japan Aerospace Exploration Agency (2026) JAXA Global Rainfall Watch. https:/​/sharak​u.eorc.​jaxa.​jp/GSMaP/guide.html. Accessed 23 Aug 2026 Jorgensen DP (1984) Mesoscale and convective-scale characteristics of mature hurricanes. Part II: inner-core structure of Hurricane Allen (1980). J Atmos Sci 41:1287–1311. ​h​t​t​​p​s​:​/​​/​d​o​​i​.​o​​r​g​/​1​0​.​1​1​7​5​ /​1​5​2​0​-​0​4​6​9​(​1​9​8​4​)​0​4​1​%​3​c​1​2​8​7​:​M​A​CS ​ ​C​O​%​3​e​2​.​0​.​C​O​;​2​​​​ Knaff JA, Longmore SP, DeMaria RT, Molenar DA (2015) Improved tropical-cyclone flight-level wind estimates using routine infrared satellite reconnaissance. J Appl Meteorol Climatol 54:463–478. https:/​/doi.or​g/10.11​75/JA​MC-D-14-0112.1 Kossin JP, Knaff JA, Berger HI, Herndon DC, Cram TA, Velden CS, Murnane RJ, Hawkins JD (2007) Estimating hurricane wind structure in the absence of aircraft reconnaissance. Weather Forecast 22:89–101. https:/​/doi.or​g/10.11​75/WA​F985.1 Kubota T et al (2007) Global precipitation map using satelliteborne microwave radiometers by the GSMaP project: production and validation. IEEE Trans Geosci Remote Sens 45:2259–2275. https:/​/doi.or​g/10.11​09/TG​RS.2007.895337 Kubota T, Shige S, Aonashi K, Okamoto K (2009) Development of nonuniform beamfilling correction method in rainfall retrievals for passive microwave radiometers over ocean using TRMM observations. J Meteorol Soc Japan 87A:153–164. ​h​t​t​​p​s​:​/​​/​d​o​​i​.​o​​r​ g​/​1​0​.​2​1​5​1​/​j​m​s​j​.​8​7​A​.​1​5​3​​​​ Lajoie F, Walsh K (2008) A technique to determine the radius of maximum wind of a tropical cyclone. Weather Forecast 23:1007–1015. https:/​/doi.or​g/10.11​75/20​08WAF2007077.1 Li Y, Wang Y, Tan Z-M (2022) How frequently does rapid intensification occur after rapid contraction of the radius of maximum wind in tropical cyclones over the North Atlantic and eastern North Pacific? Mon Weather Rev 150:1747–1760. ​h​t​t​​p​s​:​/​​/​d​o​​i​.​o​​r​g​/​1​0​.​1​ 1​7​5​/​M​WR ​ ​-​D​-​2​1​-​0​3​2​2​.​1​​​​ Liu G, Jiang S, Zheng M (2025) A global ERA5-based tropical cyclone wind field dataset enhanced by integrated parametric correction methods. Sci Data 12:1429. ​h​t​t​​p​s​:​/​​/​d​o​​i​.​o​​r​g​/​1​0​.​1​0​3​8​/​s​4​1​5​9​7​-​0​2​5​ -​0​5​7​8​9​-​w​​​​ Lu X, Wong W, Yu H, Yang X (2022) Tropical cyclone size identification over the western North Pacific using support vector machine and general regression neural network. J Meteorol Soc Japan 100:927–941. https:/​/doi.or​g/10.21​51/jm​sj.2022-048 Mallen KJ, Montgomery MT, Wang B (2005) Reexamining the nearcore radial structure of the tropical cyclone primary circulation: implications for vortex resiliency. J Atmos Sci 62:408–425. https:/​/doi.or​g/10.11​75/JA​S-3377.1 Mouche A, Chapron B, Zhang B, Husson R (2017) Combined co- and cross-polarized SAR measurements under extreme wind conditions. IEEE Trans Geosci Remote Sens 55:6746–6755. ​h​t​t​​p​s​:​/​​/​d​o​​ i​.​o​​r​g​/​1​0​.​1​1​0​9​/​T​G​R​S​.​2​0​1​7​.​2​7​3​2​5​0​8​​​​ Mouche A, Chapron B, Knaff JA, Zhao Y, Zhang B, Combot C (2019) Copolarized and cross-polarized SAR measurements for high-resolution description of major hurricane wind structures: application to Irma Category 5 hurricane. J Geophys Res Oceans 124:3905–3922. https:/​/doi.or​g/10.10​29/20​19JC015056 Journal of the Meteorological Society of Japan (2026) 104:38 Mouche A, Avenas A et al (2026) Fostering tropical cyclone research and applications with synthetic aperture radar. Remote Sens Environ 333:115139. https:/​/doi.or​g/10.10​16/j.​rse.2025.115139 Mueller KJ, DeMaria M, Knaff JA, Kossin JP, Vonder Haar TH (2006) Objective estimation of tropical cyclone wind structure from infrared satellite data. Weather Forecast 21:990–1005. ​h​t​t​​p​s​:​/​​/​d​o​​i​ .​o​​r​g​/​1​0​.​1​1​7​5​/​W​A​F​9​5​5​.​1​​​​ NASA (2026) Earthdata search. https:/​/search​.earthd​ata.n​asa.gov. Accessed 4 Mar 2026 Nishimura S, Sasaki M, Nonaka N, Ueno M, Mochizuki Y, Miura M (2023) The Japan Meteorological Agency's Dvorak re-analysis data. RSMC Tokyo – Typhoon Center Tech Rev. ​h​t​t​​p​s​:​/​​/​w​w​​w​.​j​​m​a​ .​​g​o​.​​j​p​/​j​​m​a​​/​j​m​a​-​e​n​g​/​j​m​a​-​c​e​n​t​e​r​/​r​s​m​c​-​h​p​-​p​u​b​-​e​g​/​t​e​c​h​r​e​v​/​t​e​x​t​2​6​-​1​.​ p​d​f​​​​​. Accessed 4 Mar 2026 Razin MN, Slocum CJ, Knaff JA, Brown PJ, Bell MM (2023b) Tropical cyclone precipitation, infrared, microwave, and environmental dataset (TC PRIMED). Bull Am Meteorol Soc 104:E1980–E1998. https:/​/doi.or​g/10.11​75/BA​MS-D-21-0052.1 Razin MN, Slocum CJ, Knaff JA, Brown PJ (2023a) Tropical cyclone precipitation, infrared, microwave, and environmental dataset (TC PRIMED). https:/​/doi.or​g/10.25​921/d​my1-0595 Rogers R (2023) The World Meteorological Organization Tenth International Workshop on Tropical Cyclones (IWTC-10): a summary. Trop Cyclone Res Rev 12:1–9. ​h​t​t​​p​s​:​/​​/​d​o​​i​.​o​​r​g​/​1​0​.​1​0​1​6​/​j​.​t​c​r​r​.​2​0​2​ 3​.​0​4​.​0​0​1​​​​ Shimada U (2025) Estimation of the radius of maximum wind using passive microwave satellite data. SOLA 21:265–272. ​h​t​t​​p​s​:​/​​/​d​o​​i​.​ o​​r​g​/​1​0​.​2​1​5​1​/​s​o​l​a​.​2​0​2​5​-​0​3​3​​​​ Shimada U, Aonashi K, Miyamoto Y (2017) Tropical cyclone intensity change and axisymmetricity deduced from GSMaP. Mon Weather Rev 145:1003–1017. https:/​/doi.or​g/10.11​75/MW​R-D-16-0244.1 Shimada U, Owada H, Yamaguchi M, Iriguchi T, Sawada M, Aonashi K, DeMaria M, Musgrave KD (2018) Further improvements to the statistical hurricane intensity prediction scheme using tropical cyclone rainfall and structural features. Weather Forecast 33:1587–1603. https:/​/doi.or​g/10.11​75/WA​F-D-18-0021.1 Shimada U, Hayashi M, Mouche A (2024) A comparison between SAR wind speeds and western North Pacific tropical cyclone best track estimates. J Meteorol Soc Japan 102:575–593. ​h​t​t​​p​s​:​/​​/​d​o​​i​.​o​​ r​g​/​1​0​.​2​1​5​1​/​j​m​s​j​.​2​0​2​4​-​0​3​1​​​​ Shimada U, Tsukada T (2026) Data for “A Long-Term Dataset of Tropical Cyclone Radius of Maximum Wind in the Western North Pacific Based on a Bayesian Consensus of Multiple Estimates”. 10.5281/zenodo.23131986 Accessed 4 October 2026 Sitkowski M, Kossin JP, Rozoff CM (2011) Intensity and structure changes during hurricane eyewall replacement cycles. Mon Weather Rev 139:3829–3847. ​h​t​t​​p​s​:​/​​/​d​o​​i​.​o​​r​g​/​1​0​.​1​1​7​5​/​M​W​R​-​D​-​1​ 1​-​0​0​0​3​4​.​1​​​​ STAR (2026) Tropical cyclone winds. ​h​t​t​​p​s​:​/​​/​w​w​​w​.​s​​t​a​r​​.​n​e​​s​d​i​s​​.​n​​o​a​a​.​g​ o​v​/​s​o​c​d​/​m​e​c​b​/​s​a​r​/​s​a​r​w​i​n​d​s​_​t​r​o​p​i​c​a​l​.​p​h​p​​​​.​ Accessed 4 Mar 2026 Page 19 of 19 38 Stern DP, Vigh JL, Nolan DS, Zhang F (2015) Revisiting the relationship between eyewall contraction and intensification. J Atmos Sci 72:1283–1306. https:/​/doi.or​g/10.11​75/JA​S-D-14-0261.1 Tao D, Nystrom RG, Chavas DR, Avenas A (2026) A fast analytical model for the complete radial structure of tropical cyclone lowlevel wind field. Geophys Res Lett 53:e2025GL118320. ​h​t​t​​p​s​:​/​​/​d​ o​​i​.​o​​r​g​/​1​0​.​1​0​2​9​/​2​0​2​5​G​L​1​1​8​3​2​0​​​​ Tsukada T, Horinouchi T (2023) Strong relationship between eye radius and radius of maximum wind of tropical cyclones. Mon Weather Rev 151:569–588. https:/​/doi.or​g/10.11​75/MW​R-D-22-0106.1 Tsukada T, Knaff J, Shimada U, Mouche A, Miller S (2025) Guidance for estimating mean winds in tropical cyclone using SAR derived winds. Weather Forecast 40:2699–2716. ​h​t​t​​p​s​:​/​​/​d​o​​i​.​o​​r​g​/​1​0​.​1​1​7​5​/​ W​A​F​-​D​-​2​5​-​0​0​5​9​.​1​​​​ Vickery PJ, Wadhera D (2008) Statistical models of Holland pressure profile parameter and radius to maximum winds of hurricanes from flight-level pressure and H*Wind data. J Appl Meteorol Climatol 47:2497–2517. https:/​/doi.or​g/10.11​75/20​08JAMC1837.1 Vinour L, Jullien S, Mouche A, Combot C, Mangeas M (2021) Observations of tropical cyclone inner-core fine-scale structure, and its link to intensity variations. J Atmos Sci 78:3651–3671. ​h​t​t​​p​s​:​/​​/​d​o​​ i​.​o​​r​g​/​1​0​.​1​1​7​5​/​J​A​S​-​D​-​2​0​-​0​2​4​5​.​1​​​​ Willoughby HE, Rahn ME (2004) Parametric representation of the primary hurricane vortex. Part I: observations and evaluation of the Holland (1980) model. Mon Weather Rev 132:3033–3048. https:/​ /doi.or​g/10.11​75/MW​R2831.1 Willoughby HE, Clos JA, Shoreibah MB (1982) Concentric eye walls, secondary wind maxima, and the evolution of the hurricane vortex. J Atmos Sci 39:395–411. ​h​t​t​​p​s​:​/​​/​d​o​​i​.​o​​r​g​/​1​0​.​1​1​7​5​/​1​5​2​0​-​0​4​6​9​(​ 1​9​8​2​)​0​3​9​%​3​c​0​3​9​5​:​CE ​ ​W​S​W​M​%​3​e​2​.​0​.​CO ​ ​;2​ ​​​​ Wu Q, Ruan Z (2021) Rapid contraction of the radius of maximum tangential wind and rapid intensification of a tropical cyclone. J Geophys Res Atmos 126:e2020JD033681. ​h​t​t​​p​s​:​/​​/​d​o​​i​.​o​​r​g​/​1​0​.​1​0​2​ 9​/​2​0​2​0​J​D​0​3​3​6​8​1​​​​ Yamaguchi M, Ikuta Y, Ito K, Satoh M (2025) Tropical cyclone track and intensity predictions in the western North Pacific basin using Pangu-Weather and JMA initial conditions. J Meteorol Soc Japan 103:357–370. https:/​/doi.or​g/10.21​51/jm​sj.2025-018 Zhuo JY, Tan ZM (2021) Physics-augmented deep learning to improve tropical cyclone intensity and size estimation from satellite imagery. Mon Weather Rev 149:2097–2113. ​h​t​t​​p​s​:​/​​/​d​o​​i​.​o​​r​g​/​1​0​.​1​1​7​5​/​ M​W​R-​ ​D​-​2​0​-​0​3​3​3​.​1​​​​ Zhuo JY, Tan ZM (2023) A deep-learning reconstruction of tropical cyclone size metrics 1981–2017: examining trends. J Clim 36:5103–5123. https:​​​//d​oi.​or​g/10​.117​​5​/J​CLI-​D-22-0714.1 Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. 13