Figure 1 shows the observed tracks and Saffir-Simpson scale intensities of the four 2022 Atlantic hurricanes selected for this study. These four TCs were selected because they were the most significant of the 2022 Atlantic season in terms of strength, longevity, and overall impact. Due to computational constraints, additional cases could not be included. Data are taken from the National Hurricane Center (NHC) Best Track archive (Jarvinen et al., 1984). Category 2 Hurricane Earl (2–10 September) and Category 4 Hurricane Fiona (14–23 September) were long-tracked storms that spent most of their time at hurricane intensity, defined as having maximum surface wind speed (VMAX) exceeding 65 kn, outside the tropics (Blake, 2023; Pasch et al., 2023). Both storms originated as African Easterly Wave (AEW) disturbances in the eastern tropical Atlantic, intensified as they encountered a moister environment around the Lesser Antilles, recurved northward around the Western Atlantic Ridge, and eventually underwent extratropical transition. Earl passed close enough to Bermuda to generate tropical-storm-force winds, and its outer rainbands caused heavy rainfall over Puerto Rico, leading to isolated flooding reports. Compared to Earl, Fiona was a more impactful storm. Fiona dropped 10–20 in. (254–508 mm) of rainfall over Guadeloupe and Puerto Rico, causing significant to catastrophic flash flooding on those islands. Fiona later made landfall as an extratropical storm on Nova Scotia with hurricane-force winds and a minimum central pressure of 931 hPa, becoming the deepest cyclone on record to strike Canada. Fiona's flooding and winds caused seven direct fatalities in the Caribbean islands and Canada.

Major Hurricane Ian (23–30 September) also originated as an AEW and was slow to intensify initially while moving westward through the eastern and central Caribbean (Bucci et al., 2023). Late on 25 September, upper-level winds around the storm relaxed, enabling the vortex's upper- and lower-level centers to align vertically. Ian then underwent a 39 h rapid intensification episode (Kaplan and DeMaria, 2003) while moving over an area of high ocean heat content, which ended with landfall in western Cuba. During this period Ian's VMAX increased from 40 to 110 kn. After slightly weakening from the land interaction, Ian resumed intensifying in the southeastern Gulf of Mexico, and the storm made landfall at 19:05 UTC on 28 September near Cape Coral, Florida, as a 130 kn Category 4 hurricane with a 4.6 m peak storm surge. Ian caused 156 fatalities and an estimated USD 113 billion of damage in the US, making it the third most costly US hurricane on record. Category 1 Hurricane Julia (6–10 October 2022) developed and intensified in the southwestern Caribbean and struck Nicaragua (Cangialosi 2023). Moving across Central America, Julia caused severe flooding, leading to 35 fatalities. Please refer to the NHC tropical cyclone reports [Blake, 2023 (Earl); Pasch et al., 2023 (Fiona); Bucci et al., 2023 (Ian); and Cangialosi, 2023 (Julia)] for further details on the meteorology, damage impacts, and operational forecast assessments of these storms.

This study uses version 2.0 of NOAA's HAFS-A, a regional forecasting model operationally activated by the National Centers for Environmental Prediction (NCEP) to forecast a single TC following its cyclogenesis in the Atlantic or eastern Pacific basins. HAFS-A v2.0 is dual nested with two-way feedback between a fixed ∼ 75° × 75° sized outer domain with 5.4 km horizontal resolution and a TC-following ∼ 12° × 12° sized nest with 1.8 km horizontal resolution. The 81-level sigma pressure hybrid vertical coordinate system has a 2 hPa model top, and its lower boundary is coupled to the Modular Ocean Model version 6 (MOM6). Initial conditions and lateral boundary conditions for the outer domain are provided by NOAA's operational Global Forecasting System (GFS). The Gridpoint Statistical Interpolation (GSI; Wang et al., 2013) software assimilates observations into HAFS-A's moving nest every 6 h using a Four-dimensional Ensemble Variational (4DEnVar) algorithm that extracts flow-dependent background error covariances from the 80-member operational GFS ensemble. Assimilated observations include temperature, humidity, and winds from conventional platforms (radiosondes, ships, and buoys); GNSS RO bending angles; microwave and infrared satellite radiances; and in-situ data from aircraft reconnaissance missions (dropsondes, High Density Observations from aircraft sensors, and Tail Doppler Radar-measured winds).

Assimilating observations into a background vortex that poorly estimates a TC's intensity, structure, or position can cause analysis degradations because a TC's inner core can feature sharp horizontal gradients of wind, temperature, and water vapor (Christophersen et al., 2022). To mitigate this potential problem, HAFS-A adjusts 6 h forecast fields within the moving nest using a vortex initialization (VI) algorithm during each DA cycle, prior to assimilating observations (Liu et al., 2020). The VI procedure relocates the background vortex's kinematic and thermodynamic fields to the position estimated by the real-time Tropical Cyclone Vitals Database (TCVitals; Trahan and Sparling, 2012). It then adjusts these fields to better match the TCVitals VMAX, PMIN, and storm size while keeping relative humidity (RH) unchanged. The background vortex is normally extracted from the 6 h HAFS-A forecast initialized from the previous cycle, unless the storm is very weak (TCVitals VMAX < 20 m s⁻¹), in which case the 6 h operational GFS-forecast vortex is used. In either case, HAFS-A pastes the background vortex onto the operational GFS's 6 h forecast environmental fields interpolated to the moving nest. Therefore, the HAFS-A DA system does not cycle background fields between the TC vortex edge and the moving nest boundaries, although these fields are modified by observations assimilated in the current cycle.

HAFS-A assimilates RO bending angles occurring within a ±3 h window around the analysis time using the NCEP Bending Angle Model (NBAM; Cucurull et al., 2013; Cucurull and Purser, 2023), a one-dimensional forward operator. NBAM estimates RO bending angle observation errors using a piecewise exponential function of the observation's impact height with default tunable coefficients that have been determined from GFS background and analysis departure statistics using the Desroziers et al. (2005) method. The HAFS-GSI quality control (QC) algorithm rejects RO observations that could adversely affect the analysis. Most rejections are triggered either by a “statistical check” when the fractional observation-minus-background (O-B) innovation |O-B|/O exceeds a latitude, height, and temperature-dependent cutoff value, or by the RO ray's tangent point being located within or below a background super-refractivity (SR) layer where the vertical refractivity gradient exceeds 75 % of the -157 N-units km⁻¹ critical threshold. The SR rejections are necessary because NBAM can yield an indeterminate solution for observations located inside or below a SR layer (Cucurull, 2015; Cucurull and Purser, 2023).

The ROMEX dataset used in this study contains approximately 26 000 daily global RO bending-angle profiles. EUMETSAT processed all ROMEX profiles, except for the COSMIC-2 data, which UCAR provided. The largest data volume comes from Spire, which provides about 16 750 profiles per day globally, followed by the COSMIC-2, PlanetiQ, and Metop-B/C missions, which contribute about 4900, 2750, and 1150 profiles per day, respectively (Anthes et al., 2024). Prior to a December 2024 code upgrade, GSI implementations with NCEP's forecasting models rejected all Metop bending angles below the 8 km impact height (∼ 650 hPa) based on prior work finding larger Metop-A/B departures against the GFS background (Dutta et al., 2019). Since our HAFS-GSI build predates the removal of this QC check, the assimilated Metop-B/C profiles are expected to have little impact on the forecast lower- to middle-tropospheric temperature and water vapor fields. The government-owned TerraSAR-X, TanDEM-X, and KOMPSAT-5 missions together contribute up to 500 daily global profiles. The ROMEX commercial bending angles are assigned the same observation errors and QC settings as those used for COSMIC-2, except that their QC cutoff threshold rejects all data above the 30 km impact height, whereas COSMIC-2 and other government RO mission data are rejected above 50 km. This follows the default settings for commercial RO data in the operational HAFS-GSI.

To demonstrate the impacts of assimilating the ROMEX commercial observations, COSMIC-2, and all lower-tropospheric RO data, four independent HAFS-A observing system experiments (OSEs) are run as follows: i.

Control, assimilating only data from government-owned missions (primarily COSMIC-2 and Metop);

Figure 2a shows the number of Spire, PlanetiQ, and COSMIC-2 profiles available for assimilation in HAFS-A during each Hurricane Fiona analysis cycle. As expected, Spire provides the largest volume of RO data, although the number of profiles overlapping the ∼ 12° × 12° domain varies substantially across different analysis cycles, ranging from one to thirty-one. Anthes et al. (2024) observed a bimodal distribution in the local time sampling density of ROMEX Spire profiles, with peaks at 09:00–12:00 and 21:00–00:00 LT (see their Fig. 12), which they attributed to the Spire satellites' sun-synchronous polar orbits. Spire's bimodal local time sampling pattern is not easily discernible during periods when the HAFS-A nest has limited zonal movement, such as 20–22 September, when Fiona remains over the US Eastern Daylight Time zone's (GMT-4 h) longitudinal range while moving mostly northward east of the Bahamas (Fig. 1). This is not surprising, given the HAFS-A nest's relatively small size and wide (6 h) DA window. Therefore, we attribute the differences in Spire profile availability over DA cycles mostly to variability in Spire occultation swath locations over the TC case study periods.

Table 1 lists the number of RO profiles assimilated per day in the Control_noC2, Control, and ROMEX experiments for each TC case. Over the full evaluation period, the addition of Spire and PlanetiQ data in ROMEX increases the assimilated profile count 4.4-fold for Earl, 3.8-fold for Fiona, 3.3-fold for Ian, and 1.8-fold for Julia, relative to Control. Earl and Fionas' larger gain in RO data volume from Control to ROMEX is consistent with those storms spending more time outside the tropics (Fig. 1), where COSMIC-2 data have sparser coverage (Ho et al., 2020b). When combining all analyses from the four TC cases, Control_noC2 assimilates 129 profiles, Control assimilates 717 profiles, and ROMEX assimilates 2513 profiles.

Figure 2b highlights a key challenge: for some COSMIC-2 and commercial RO profiles, all bending angles below 800 hPa are rejected by HAFS-GSI's QC checks and not assimilated. This is critical because TC intensification is highly sensitive to water vapor distribution in the lower-to-middle troposphere around the storm (e.g., Nolan, 2007; Doyle et al., 2012; Teng et al., 2021). The GSI QC algorithm is known to reject the largest percentage of RO data in the lower-to-middle troposphere (Miller et al., 2023, 2025; Lien et al., 2021), primarily due to greater bending angle retrieval uncertainty caused by water vapor effects (Kursinski et al., 1997; Kuo et al., 2004; Anthes et al., 2022). Further work is needed to determine whether HAFS-GSI's QC criteria are optimally set for all RO data platforms.

Figure 3 presents histograms of COSMIC-2, Spire, and PlanetiQ bending angle rejection percentages binned by pressure level for all checks (Fig. 3a) and the statistical O-B outlier check (Fig. 3b). The bins extend up to 15 hPa, which is below the 30 km high altitude cutoff used for commercial RO data. All three platforms show similar retention statistics above 950 hPa that are mostly caused by the statistical O-B check (cf. Fig. 3a, b), with the total rejection rate increasing from ∼ 20 % in the 700–450 hPa layer to ∼ 55 % in the 950–850 hPa layer (Fig. 3a). Notably, below 950 hPa, Spire's O-B rejection percentage is almost 10 % smaller than that of COSMIC-2. Since GSI uses the same empirically-determined |O-B|/O rejection thresholds for COSMIC-2 and commercial RO data, this indicates that Spire has a lower frequency of extreme O-B outliers within and near the boundary layer than COSMIC-2. Almost all other tropospheric RO observation rejections are triggered by the super-refractivity check, which mainly depends on the background refractivity field (Cucurull, 2015).

Figure 4 compares profiles of COSMIC-2, Spire, and PlanetiQ observation error standard deviations estimated from the ROMEX experiment's O-B and observation-minus-analysis (O-A) data collected from all four TC cases. These diagnosed observation errors include contributions from instrument, retrieval, and representative sources, and we estimated them using the Desroziers et al. (2005) algorithm, which has been widely employed by operational NWP centers (Cucurull et al., 2013; Bonavita, 2014; Bowler, 2020; Lien et al., 2021). Some caution is warranted in drawing broader conclusions about the uncertainty characteristics of these three RO observation platforms, given our relatively small DA diagnostic data sample drawn from a TC-following regional domain. From our experiments, we find that all three platforms have similar observation uncertainty in the upper troposphere. However, Spire observations show the smallest uncertainty below the 5 km impact height in the tropics (Fig. 4a) and below the 7 km impact height in the extratropics (Fig. 4b).

We now evaluate the impacts of assimilating ROMEX and COSMIC-2 RO observations on HAFS-A TC track and intensity forecasts verified against the NHC's Best Track dataset. Forecasts initialized during the first 48 cycling hours (containing eight analysis cycles) are not used for computing verification statistics to allow HAFS-A time to “spin up” and respond to the accumulative impacts of assimilated observations. Figure 5a compares the mean absolute TC position errors across the four experiments as a function of forecast lead time. Also shown is the number of Control forecasts used for averaging, which decreases with increasing lead time because the HAFS-A TC vortex tracker does not provide output for storms that have weakened after landfall or moved near the edge of the fixed outer domain. Differences in TC position error among the four experiments are generally small, which is not surprising given that TC motion is often most sensitive to larger-scale flows driven by synoptic weather features far from the storm (Chan and Gray, 1982; Galarneau and Davis, 2013). The HAFS-A model's 12° × 12° storm-following nest receives the same GFS-provided initial and lateral boundary conditions from its surrounding outer domain in all four experiments (Sect. 2.2).

Nevertheless, we find that assimilating the commercial RO observations has a small positive impact on TC position errors during the t=18–102 h period, as evidenced by the positive ROMEX relative skill versus Control (Fig. 5b). When analyzing mean absolute error (MAE) statistics from a set of forecasts, we define relative skill for each sensitivity experiment (EXP) as (MAEControl-MAEEXP)/MAEControl. Control_noC2's negative relative skill during the same period (Fig. 5b) indicates that COSMIC-2 observations are also beneficial for TC track forecasts. This is particularly true for the subset of analyses that initialize TCs equatorward of 24° N, for which removing COSMIC-2 from Control yields negative 5 %–10 % relative skill during t=54–96 h (not shown). We hypothesize that these small- to medium-range track forecast improvements, gained from assimilating COSMIC-2 and commercial RO data in the HAFS-A moving nest, could come from improved model predictions of TC vortex size, depth, or circulations in the near-storm environment, which in turn influence the storm's response to larger-scale flow fields (Galarneau and Davis, 2013).

Figure 6 compares the HAFS-A experiments' PMIN mean absolute error time series. ROMEX commercial observations are modestly beneficial through t=36 h, when they provide a ∼ 5 %–15 % relative skill improvement, which is statistically significant for 0 h and 12 h forecasts. A substantial portion of ROMEX's t=0–36 h relative skill improvement is lost when RO data below the 5 km impact height are removed (compare the green and orange curves in Fig. 6b). COSMIC-2 impacts on forecast PMIN intensity mean absolute errors are mixed. Assimilating the ROMEX commercial observations nearly eliminates Control's negative 2–3 hPa mean PMIN bias over the t=24–90 h period (Fig. 7). COSMIC-2 assimilation partially contributes to Control's PMIN over-intensification bias, as shown by comparing the red and black curves in Fig. 7.

We next compare the HAFS-A experiments in terms of their forecast temperature and water vapor errors, using the ECMWF Reanalysis version 5 (ERA5) as the verification truth. Figure 8a shows that assimilating ROMEX commercial RO data reduces the temperature Root-Mean-Square Deviation (RMSD) in the 200–400 hPa layer by several percentage points relative to Control over most of the forecast period. This improvement in ROMEX temperature RMSD extends down to around 800 hPa between t=18 h and t=48 h. In the lower troposphere, the impacts of ROMEX commercial RO observation impacts on temperature RMSD are mixed, showing degradation over the first 18 h but mostly improvement after t=54 h. Regarding bias, Fig. 8b shows that Control has a cool bias in the lower troposphere and a warm bias in the mid-to-upper troposphere. Assimilating the ROMEX commercial RO observations has little impact on the HAFS-A temperature bias patterns, except for a slight mitigation of the warm bias between 200 and 400 hPa in the later forecast period. Compared to the ROMEX commercial RO data, the assimilated COSMIC-2 observations have more mixed and muted impacts on the HAFS-A temperature fields, as evidenced by the Control and Control_noC2 temperature RMSD and bias field comparisons shown in Fig. 8c and d, respectively.

Differences between the Control-normalized temperature RMSDs shown in Fig. 8a and e reflect the impacts of RO data assimilated below 5 km. Removing all lower-tropospheric RO observations from the ROMEX configuration (Fig. 8e) results in a substantial portion of the ROMEX temperature RMSD improvement (relative to Control) below 600 hPa (Fig. 8a) being lost. Predictably, the temperature RMSD improvement in the upper troposphere is largely retained in ROMEX_noLowlevRO. This outcome is expected, as localization in the HAFS-A background-error covariance matrix should cause assimilated observation impacts to dampen with increasing vertical distance from the observation.

Overall, assimilating ROMEX commercial RO data reduces random errors in HAFS-A forecast tropospheric water vapor. Like temperature, the ROMEX specific humidity RMSD is reduced by several percentage points relative to Control over most layers and times in the middle and upper troposphere, with mixed impacts in the lower troposphere (Fig. 9a). Some of the most substantial improvement occurs during analyses and short-range forecasts around 500 hPa, where the ROMEX specific humidity RMSD is 3 %–6 % smaller than Control. The ROMEX commercial RO data have a modest drying impact; they mitigate a Control moist bias in the 400–700 hPa layer but exacerbate a Control dry bias in short-to-medium range forecasts below 850 hPa (Fig. 9b). However, these ROMEX-minus-Control specific humidity mean error differences are quite small, generally less than 0.1 g kg⁻¹.

Unlike the mixed results for temperature, COSMIC-2 observations show a clearer positive impact on water vapor fields, as evidenced by the widespread specific humidity RMSD degradation in Control_noC2 relative to Control (Fig. 9c). COSMIC-2 assimilation also yields mixed and generally very small changes in HAFS-A water vapor mean biases (Fig. 9d). Finally, removing all (commercial- and government-mission) RO data below 5 km in ROMEX_noLowlevRO causes nearly all water vapor RMSD improvement below 400 hPa in ROMEX to be lost (cf. Fig. 9a, e). This implies that lower-tropospheric RO observations can also improve HAFS-A water vapor estimates in the middle troposphere via the background-error covariance matrix.

NOAA P-3 reconnaissance aircraft sampled Hurricanes Earl, Fiona, and Ian at 12–24 h intervals during their intensification and mature phases. Flying around 3 km in altitude, the P-3 aircraft crossed the TC inner core in several transect legs while releasing GPS dropsondes and collecting other in-situ observations. Two NOAA Gulfstream-IV missions also flew into Fiona's near-storm environment and released upper-tropospheric dropsondes. For verification, we compute 12 h HAFS-A forecast errors for temperature, specific humidity, and wind speed from 21 initialization cycles (7 from Earl, 7 from Fiona, 6 from Ian, and 1 from Julia; see Table 2) against dropsondes available within ±3 h of the verification time. Dropsondes were equipped with GPS sensors, enabling tracking of their three-dimensional trajectories during free fall. Since most dropsondes were collected by the lower-flying P-3 missions, our verification is restricted to the sub-700 hPa layer.

Figure 10a compares the 12 h forecast temperature biases and RMSDs of the four HAFS-A experiments against collocated dropsonde temperature observations. Control shows a negative 0.5–1 K cool bias in the sub-700 hPa layer. Compared to Control, the ROMEX temperature bias profile is similar, with one exception: it shows a ∼ 0.25 K stronger cool bias in the 700–750 hPa layer. Removing the sub-5 km RO data from our ROMEX experiment reduces the cool-temperature bias in this layer, bringing it closer to that of Control, suggesting that sub-5 km RO observations drive the ROMEX experiment's degradation there. The Control_noC2 cool bias is also ∼ 0.25 K stronger than Control's between 700 and 750 hPa, indicating a positive impact from assimilating COSMIC-2 in that layer. For temperature RMSD, both the ROMEX and ROMEX_noLowlevRO are comparable to or larger than Control's over the sub-700 hPa layer. However, assimilating COSMIC-2 in Control yields a 0.1–0.2 K RMSD improvement over Control_noC2 within a large portion of the sub-700 hPa layer.

For the 12 h forecast specific humidity verification (Fig. 10b), we find that Control has a ∼ 0.5 g kg⁻¹ dry bias below 700 hPa, and that ROMEX's dry bias is similar or slightly worse. RO data assimilated below 5 km are slightly beneficial, as removing them increases the dry bias by ∼ 0.1 g kg⁻¹ over portions of the sub-700 hPa layer. COSMIC-2 assimilation has little impact on Control's dry bias. While the ROMEX specific humidity RMSD profile resembles that of Control, the RO data assimilated below 5 km are generally beneficial for reducing random water vapor errors, particularly around 900 hPa. At this level, the ROMEX_noLowlevRO specific humidity RMSD is a statistically significant ∼ 0.2 g kg⁻¹ larger than those of both ROMEX and Control. Above 800 hPa, COSMIC-2 assimilation in Control also yields a statistically significant ∼ 0.2 g kg⁻¹ specific humidity RMSD improvement over Control_noC2.

The assimilated ROMEX RO data improves the horizontal wind speed RMSD by ∼ 1–2 m s⁻¹ in the 750–950 hPa layer, although this result is not statistically significant. Removing all sub-5 km RO data more than reverses all of ROMEX's wind RMSD improvement, and additionally, removing the COSMIC-2 observations degrades the Control wind RMSD. Therefore, the lower-tropospheric RO dataset from Spire, PlanetiQ, and COSMIC-2 is beneficial for reducing random horizontal wind-speed errors below 750 hPa in HAFS-A. Even though the RO forward operator does not depend on kinematic variables, this result is consistent with Chen et al. (2020), who showed that cycled RO refractivity assimilation in a regional model could improve short-range forecasts of vertical velocity and vorticity during TC cyclogenesis via dynamical coupling between thermodynamic and kinematic fields.