Track-Based Cyclone Maximum Wind Retrievals Using the Cyclone Global Navigation Satellite System (CYGNSS) Mission Full DDMs

Cyclone maximum wind retrievals are demonstrated using tracks of full delay-Doppler map (DDM) measurements of the Cyclone Global Navigation Satellite System (CYGNSS) constellation. The method is based on the production of a library of simulated DDMs for a synthetic storm model, as the parameters of the synthetic storm are varied. A matched filter approach between normalized simulated and measured DDMs is then applied to retrieve storm parameters. Because the method uses normalized DDMs in the retrieval, it is insensitive to the absolute power calibration of the DDM and instead relies on the DDM shape. Results from the method are shown for CYGNSS full DDM overpasses of Hurricane Irma as a case study. The maximum wind speed estimates obtained were associated with a retrieval root-mean-square error (RMSE) of 6.89 m/s and an unbiased RMSE 5.20 m/s with a mean difference of 4.52 m/s compared to reported National Hurricane Center Best Track forecasts.

SECTION I. Introduction

Monitoring the maximum wind speeds of hurricanes from space is of high interest due to the importance of this parameter in forecasting storm impact and evolution. According to reports by the National Oceanic and Atmospheric Administration (NOAA) [1]–​[3], the 2017 Atlantic hurricane season was associated with economic damages in excess of $200 billion. The ability to characterize and monitor storm properties and their evolution is, therefore, crucial to improve readiness and to minimize the damage incurred to human life and property. The forecasting of hurricane intensity in terms of the maximum sustained wind speed remains an ongoing challenge. This is compounded by the fact that storms can undergo substantial changes in their structure over a relatively short time span. Current numerical models, such as the Hurricane Weather Research and Forecasting model, can predict such rapid changes with only limited accuracy.

Several existing spaceborne platforms are able to provide cyclone information. Scatterometers operating in the $C$ or $K_u$ bands, such as ASCAT, RapidSCAT. and QuickSCAT, have provided surface measurements of storm vector winds, and precipitation radar measurements as in the Global Precipitation Measurement mission can provide information on rain rates and storm vertical structure. All these measurements, however, can experience limitations in determining the maximum winds in a cyclone inner core due to the high rain rates and high attenuation within these, $C$ and $K_u$, frequency bands.

The Cyclone Global Navigation Satellite System (CYGNSS) mission was developed to address these challenges. CYGNSS's L-band (1.575 GHz) frequency operation makes it insensitive to adverse weather conditions including heavy rain, making observations of winds in the central core of storms possible; this together with its frequent temporal revisits provides an opportunity to investigate means, through which its returns can be used for the purposes of cyclone maximum sustained surface wind (SSW) retrievals.

Due to these desirable properties and the mission's focus on cyclone wind sensing, several investigations into the retrieval of cyclone wind speeds have previously been reported [4]–​[7]. These investigations are based on the use of CYGNSS's Level 2 wind speed retrievals, which are derived from the calibrated power measurements of the Level 1 delay-Doppler map (DDM) observation. While CYGNSS's wind retrievals are continuing to advance with time, the difficulties of achieving precise absolute calibration for eight observatories acquiring the reflections of 32 distinct GPS transmitters have made high wind retrieval performance an ongoing focus area for improvement. The tendency of the received signal power to decrease monotonically at a slow rate with the increasing surface wind speed also makes this process difficult.

This study, therefore, attempts to make use of the measurements made under one of CYGNSS’ special modes of operation, the full DDM mode, in order to investigate the utility of retrieving SSW through the matching of CYGNSS observations to a reference library of waveforms produced under varying storm conditions. It is noted that a previous work explored a simulation study using a method similar to that reported here [8]; this study extends the method described in [8] and demonstrates its performance using measured CYGNSS data.

The rest of this article is organized as follows. Section II presents an overview of the CYGNSS mission and the various data products it provides under its standard and special modes of operation. Section III provides an overview of the proposed retrieval approach and its major elements. This includes the tools used to forward model CYGNSS returns under different storm conditions, the parametric modeling of storms, and the retrieval process. Section IV summarizes the results obtained by extending the retrieval approach to CYGNSS full DDM observations over Hurricane Irma. A discussion of these results along with other considerations follows in Section V.

SECTION II. Background

A. CYGNSS Mission

The CYGNSS mission launch on December 15, 2016 placed a constellation of eight small satellites in Low Earth Orbit at altitudes of approximately 520 km. Each satellite hosts the CYGNSS Delay-Doppler Mapping Instrument (DDMI) payload that receives GPS signals reflected from Earth's surface [9], [10]. Each CYGNSS satellite receiver, therefore, operates as a bistatic radar observing the specular forward scattering from Earth of a particular GPS transmitter. By cross correlating received GPS transmissions with a locally generated copy of the transmitted GPS C/A-code, fundamental GNSS-R measurement, the DDM, is formed. “Pixels” in the DDM represent the power scattered from Earth's surface at specific offsets in delay and Doppler from the specular reflection point. Each CYGNSS satellite is equipped with two nadir antennas and is capable of tracking four simultaneous reflections at any given second. Together, the constellation produces up to 32 DDM measurements every second for latitudes within $\approx \pm 38^\circ$. CYGNSS measurements occur in tracks that repeat irregularly in time and space due to the motion of both the CYGNSS receivers and GPS transmitters; mean and median revisit times are $\approx$7 and 3 h, respectively [4], [11]. These properties make CYGNSS measurements favorable for achieving frequent storm coverage.

The primary scientific mission of CYGNSS is to measure incoherent scattering from the ocean's surface and to utilize these measurements for ocean wind speed retrieval [12]–​[14]. Wind speed retrievals are based on the relationship between ocean surface roughness (which impacts incoherent forward scatter) and ocean surface wind speed using empirically developed descriptions of the specular scattering process.

B. Data Products

CYGNSS receivers are configured to produce three different data products. The first is the Level-1 (L1) DDM produced continuously in normal science mode observations. The second and third, called the “full DDM” and “Raw Intermediate Frequency” modes, respectively, are used only in special situations, for example, when a CYGNSS receiver observes a scene of particular importance, such as a tropical cyclone.

On board the receiver, the observed surface scatter is mapped into a DDM comprised of 128 delay bins (approx. $\pm$16 chips at the $\approx$0.25 $\mu$s CYGNSS sampling rate) and 20 Doppler bins (approx. $\pm 5$ kHz at the 500 Hz CYGNSS sampling rate). This is the CYGNSS full DDM, which is downlinked to ground stations when CYGNSS operates in the full DDM mode. Over the ocean, one half of the delay rows contain receiver noise only with the mean of the first 45 delay rows used to estimate the measurement's noise floor. A DDMI compression algorithm onboard the receiver then estimates the location of the specular bin as the location of the maximum value of the DDM about which a bounding box is formed. The box has a 17-bin extent in delay (approx. 3.5 chips with the first few delay rows containing thermal noise) and an 11-bin extent in Doppler (approx. $\pm$2.5 kHz). This is the L1 DDM, a subset of the full DDM formed on board the DDMI. A smaller portion of the standard DDM values corresponding to the specular point are averaged to produce a “DDM average” (DDMA) absolute power product that is used for wind speed retrieval. Illustrations of the full DDM and standard DDM data products are provided in Fig. 1. Finally, the Raw I/F mode records the CYGNSS intermediate frequency data stream in units of counts before any correlation with the GNSS code or integration. This enables the creation of DDMs of arbitrary extent and sampling in ground processing, at the cost of a greatly increased downlink datarate. Raw I/F mode is not used in the results reported in this article.

Fig. 1.

Illustration of different types of CYGNSS DDMs. (a) Full DDM observed by CYG05 on DOY-Year 280-2017 over the ocean at 22.62$^\circ$N 95.57$^\circ$W. (b) Standard CYGNSS Level-1 DDM cropped from Full DDM.

C. On the Delay Extent of CYGNSS Products

The distinction that is of most interest between the different data products is their varying extents in delay. This refers to the total time delay a transmitted signal undergoes in its propagation from the transmitter, scattering off the Earth's surface and its forward propagation to a CYGNSS receiver, and is typically expressed relative to the delay for scattering arising from the specular point (i.e., the specular scatter delay = 0 $\mu$s). On the surface, this translates to concentric ellipses of constant delay with major and minor axes $a$ and $b$ given by

$$\begin{align*} a &= \sqrt{\frac{2\,c\,\tau \,R_r\,R_t\, \text{sec}^2\theta _i}{\left(R_r+R_t\right)}}\tag{1}\\ b &= a \cos \theta _i \tag{2} \end{align*}$$ where $\tau$ is the excess delay relative to the specular path, $\theta _i$ is the incidence angle, $R_r$ is the range from the receiver to the specular point, and $R_t$ is the range from the transmitter to the specular point. An effective diameter of the delay ellipse $d$ can further be defined as $d = 2\sqrt{a\cdot b}$. An illustration of the maximum spatial extent of each of CYGNSS’ DDM products on a track that is estimated to have come within 20 km of Hurricane Irma's storm center is shown in Fig. 2; the significantly larger spatial extent of the full DDM is evident. Note that the maximum surface area spanned by a given measurement is incidence angle dependent; a plot of the effective diameter of the standard and full DDM products is illustrated in Fig. 3 for a typical Medium Earth Orbit GPS transmitter and a CYGNSS receiver.

Fig. 2.

Illustration of maximum delay (spatial) extent at $\theta _i\cong$ 36$^\circ$ of different data products based on transmitter–receiver geometries for CYGNSS L1/full DDM track intersecting a simulated representation of Hurricane Irma on DOY-Year 247-2017 based on the Willoughby model. Each blue bullseye represents a specular point, plotted once every five specular points. Each cyan ellipse indicates a four chip separation. Track estimated to have come within 20 km of Irma's storm center.

Fig. 3.

Illustration of angle dependence profile different CYGNSS DDM types.

In observation scenarios, where CYGNSS tracks do not align exactly with the storm center (i.e., near miss geometries) and instead pass through the storm at varying radial separations from its center, the effectiveness of the L1 DDM may be limited as it only furnishes information about a small portion of the storm. The full DDM spatial extent instead “samples” a larger portion of the storm profile so that information from the storm center, the transition region, and the portion of the storm that is beyond the transition region is included. As a result, the full DDM measurement has the potential of furnishing more information about the state of the surface and, therefore, more information about the state of the storm from a single track of measurements.

SECTION III. Retrieval Methodology

Several recent works have investigated the utility of CYGNSS measurements for the purposes of retrieving one or more storm parameters. Many such efforts have relied on the use of the standard retrieved Level-2 wind speeds [6] from the CYGNSS mission in order to produce a storm characterization having parameters, such as the storm maximum wind speed $V_{\max }$, measures of storm radii ($R_{64}$, $R_{50}$, and $R_{34}$), and the radius of maximum surface wind speed $R_{\max }$ through the fitting of retrieved CYGNSS wind tracks to modeled storms [7]. Due to the reliance of these methods on CYGNSS L2 winds, they make use of only the small portion of the DDM used to form the DDMA [15]–​[17]. This portion typically extends over five Doppler bins (approx. 2 kHz at the CYGNSS 500-Hz Doppler sampling rate) by three delay bins (approx. $\approx \!0.75$ $\mu$s at the CYGNSS 0.25 $\mu$s sampling rate) surrounding the specular return. The surface area represented by the DDMA varies with the angle, but is generally $\leq$50 × 50 km. While these efforts continue to show promise, highlighting CYGNSS's potential in improving storm feature characterization, more recent investigations using a similar approach have highlighted challenges caused by the dependence of DDMA retrieved wind speeds on CYGNSS absolute power calibration[18]–​[21] and its associated uncertainties [7].

The method used in this article extends that of [8] by matching CYGNSS full DDM measurements to a library of simulated DDMs created for a storm parametric model as the parameters of the storm are varied. The retrieved storm parameter values then correspond to those of the library storms, whose DDMs best match a particular full DDM measurement or track of full DDM measurements. Both the measured and reference DDMs are normalized by their maximum values before the matching is performed, so that the retrieval approach focuses on the use of DDM “shape” rather than “amplitude.” The impact of any uncertainties associated with CYGNSS absolute power calibration on retrieval performance is, therefore, eliminated.

A. Forward Modeling CYGNSS Returns

To produce the reference template library of waveforms, the CYGNSS end-to-end simulator (E2ES) [22]–​[25] is used. The E2ES provides a simulation of CYGNSS measurements for a given transmitter and receiver geometry using a gridded representation of the wind field on Earth's surface in the vicinity of the specular point. The forward model applied is that of [26], and the surface-normalized bistatic radar cross section is modeled using the geometrical optics approximation with the empirically determined mean square slopes (MSS) of [28], [29]. The MSS model assumes that the sea is roughened solely by surface wind and may not fully account for dependencies on storms’ radial and/or azimuthal profiles. See [22]–​[25] for additional information on the E2ES.

Due to the impact the surface wind field distribution can have on the DDM, it is important that the surface winds used in the E2ES grid provide a reasonable representation of the nonuniform winds encountered in cyclones. Parametric wind fields can be used for this purpose. Several such models exist in the literature [30]–​[33]; the choice of which model to use in this article was made as a compromise between the model complexity (in terms of the number of parameters used) and the physical representativeness of its winds. As in [8], the Willoughby storm model [34], [35] with incorporated translational effects [36] was selected as a reasonable compromise between these goals. The model is a function of storm latitude $\varphi$ and SSW $V_{\max }$ and has been shown to outperform preceding models such as those in [30] and [33], as shown in [34] and [35].

The model describes an exponentially decaying wind speed function within the eye wall radius $R_1$ given by $V_{{\rm in}}$, a transition wind speed function between $R_1$ and the radial extent of the transition region $R_2$ given by $V_{{\rm tr}}$, and a second exponentially decaying wind speed function beyond $R_2$ given by $V_{{\rm out}}$ where $V_{{\rm tr}}$ is a combination of $V_{{\rm in}}$ and $V_{{\rm out}}$ weighted by the parameter $A_w$. Azimuthal asymmetries are incorporated by accounting for the storm's translational speed $V_g$ and direction $\theta _g$. Additional information on the Willoughby model is provided in [34]–​[36]. A representation of Hurricane Irma at different points in its life cycle using the Willoughby model is shown in Fig. 4.

Fig. 4.

Illustration of surface wind fields based on Willoughby storm model functions for Hurricane Irma with/without translational effects incorporated. (a) DOY 242, $V_{\max }$ = 23.14 m/s without translation effects. (b) DOY 243, $V_{\max }$ = 33.44 m/s without translation effects. (c) DOY 242, $V_{\max }$ = 23.14 m/s, $V_{g}$ = 17.81 km/h, and $\theta _{g}$ = 270.14$^\circ$. (d) DOY 243, $V_{\max }$ = 33.44 m/s, $V_{g}$ = 16.94 km/h, and $\theta _{g}$ = 289.19$^\circ$.

To assess the impact of including a parametric storm model in the DDM computation, Fig. 5 illustrates the difference between the DDM predicted using a uniform wind field equal to the wind speed at the specular point and the nonuniform synthetic storm case. The particular geometry considered is shown in the upper portion of the figure. The differences obtained in this case (e.g., strictly positive and strictly negative differences below and above zero Doppler, respectively) are affected by the position of the specular point relative to the storm center. The differences at the individual specular point level are modest, typically on the $\pm 10\%$ levels. However, the effect of these differences can be amplified by performing the retrieval over all points on a CYGNSS track within a specified distance of the storm center instead of for a single 1 s CYGNSS measurement. This approach is adopted in what follows. The small differences obtained between the uniform and nonuniform winds cases make it clear that a highly accurate forward model, such as the E2ES, is necessary to conduct successful storm parameter retrievals using this method. While the results to be shown will demonstrate that use of the Willoughby model can produce reasonable retrievals, it is also possible to apply the retrieval methodology with other parametric wind field descriptions. Such efforts are under current consideration due to some of the limitations of the Willoughby model that will be discussed further in Section V.

Fig. 5.

(a) CYGNSS track with projection of iso-Doppler (green) lines in kHz and iso-Delay (cyan) ellipsis corresponding to maximum full DDM spatial extent. (b) Illustration of the DDM percentage shape difference under the assumptions of Willoughby-model-based surface wind fields and uniform wind fields.

B. Maximum Wind Estimators

The retrieval is based on use of the E2ES forward model for CYGNSS returns, which can produce predicted returns for synthetic storm models having varying storm features, in particular the maximum sustained wind speed ($V_{\max }$). A “matched filter” approach is then adopted by comparing predicted returns with those observed throughout the time when a CYGNSS track was within, approximately, $\pm$150 km of the storm center. The matching is performed between the predicted $S(\tau,f_D; V_{\max })$ and measured $M(\tau,f_D; V_{\max })$ DDMs normalized by their rms amplitudes. Multiple metrics for assessing agreement were examined. The first measure finds storm $V_{\max }$ parameter that maximizes the correlation between waveforms over a track of $P$ specular points

$$\begin{align*} V_{\max }^{*} & = \underset{\mathrm{V}_{\max }}{\mathrm{argmax}} \sum _{p=1}^{P}\,R\left(V_{\max }\right) \tag{3} \end{align*}$$ where $$\begin{align*} & R\left(V_{\max }\right) =\\ & \frac{\left|\langle M_p^{\tau,f_D}(V_{\max }),S_p^{\tau,f_D}(V_{\max })\rangle \right|^2}{\langle M_p^{\tau,f_D}(V_{\max }),M_p^{\tau,f_D}(V_{\max})\rangle \langle S_p^{\tau,f_D}(V_{\max }),S_p^{\tau,f_D}(V_{\max})\rangle }\tag{4} \end{align*}$$ in which the bracket notation refers to a pointwise multiplication of DDM “pixels” followed by a summation of all pixel products. This provides the estimate $V_{\max }^R$ of SSW based on maximum correlation. A secondary measure for assessing agreement based on the RMSE between $S(\tau,f_D; V_{\max })$ and $M(\tau,f_D; V_{\max })$ was also examined $$\begin{align*} & V_{\max }^* = \underset{\mathrm{V}_{\max }}{\mathrm{argmin}}\sum _{p=1}^{P}\\ &\sqrt{\frac{1}{N\tau }\frac{1}{N_f}\sum _{i=1}^{N_\tau }\sum _{j=1}^{N_f}\left|M_p(\tau ^i,f_D^j;V_{\max })-S_p(\tau ^i,f_D^j;V_{\max })\right|^2}\tag{5} \end{align*}$$ where $N_\tau$ and $N_f$ are the total number of delay and Doppler bins in a single DDM, respectively. This provides the estimate $V_{\max }^{\text{RMSE}}$ of SSW based on the minimum RMSE. Due to the potential of slight misalignment of the exact locations of the specular bins within the predicted and measured DDMs, the predicted DDM is shifted, relative to the measured DDM, over a limited range in delay and Doppler until the point of the best alignment between the two for a given specular point is reached.

Performance has been found to improve when a full track of measurements is used rather than a single “snapshot” point measurement to infer the parameters of a cyclone model, since multiple measurements in a known geometry with respect to the storm location can then be combined with a physical model for the cyclone spatial structure. Such an approach can compensate at least in part for the relatively coarse CYGNSS spatial resolution when compared to cyclone maximum wind features [5], [15].

The retrieval method can also be extended to retrieve additional parameters beyond maximum wind speed due to the extensive amount of data used in the retrieval process (the multiple “pixels” in each DDM in multiple DDM measurements). The retrieval of storm location was also examined, in terms of offsets in the storm center location from the ancillary information in latitude and longitude, as described in the following section. It is noted that there is some potential for the retrieval to become ill-posed as the number of parameters is increased; the results shown in this article, however, all showed a unique minimum in the retrieval process.

SECTION IV. Results

Retrievals were conducted for CYGNSS full DDM observations of Hurricane Irma during the 2017 Atlantic Hurricane season. An illustration of these tracks over a single acquisition is illustrated in Fig. 6. The 15 individual tracks each contain more than 500 specular points spanning distances of more than 3000 km; only a portion of the measurements are within sufficient distance of the storm to be useful for retrieval. To identify the relevant measurements, a priori storm location information from National Hurricane Center (NHC) Best Track forecasts [2] is used. This information is updated on 6-h intervals based on numerical modeling and reconnaissance, and the maximum sustained winds from the forecasts are later used in retrieval performance assessment. Using known Best Track storm locations, the translational speed and direction of the storm is computed at the same temporal resolution, based on a WGS84 curved earth, and used as a priori information for the Willoughby model. The storm is also propagated in space assuming a constant translational velocity over 6-h intervals to obtain approximate knowledge of the storm's location in space and time at 1-s intervals. For each selected full DDM track, 51 specular points centered on the closest point are used in the retrieval, representing a $\approx$300 km transect of the storm. The reference library of waveforms is generated for $30\leq V_{\max }\leq \text{80}$ m/s at 1 m/s increments; note that this is the expected range, over which the storm will have developed from a tropical depression and/or storm to a Category 1–5 hurricane based on the Saffir–Simpson Hurricane Scale [39]. To enable storm location retrievals, the reference library also includes sweeps over a $\approx$$200\times 200$ km sampled every 20 km range of potential storm centers centered about the estimated storm center location at the time of the CYGNSS track. A simultaneous retrieval of storm location and maximum winds is possible [8] by searching for the $V_{\max }$ and offset parameters that produce the best fit. However, better performance was found by first retrieving $V_{\max }$ and then subsequently performing a second retrieval process to find location offsets using the $V_{\max }$ value already determined.

Fig. 6.

Full DDM coverage of Hurricane Irma on Sep. 3, 2017. Storm locations denoted by cyan circles at 00:00:00, 06:00:00, 12:00:00, 18:00:00, and 00:00:00 right to left.

Fig. 7 plots the obtained RMSE and correlation values obtained as a function $V_{\max }$ for a CYGNSS track through Hurricane Irma on DOY 246, 2017. The two metrics obtain similar peaks in correlation or minimum in RMSE for $V_{\max } \approx 58$ m/s. Other tests also showed the two statistics to provide similar wind speed retrieval results, so an additional approach was adopted of averaging the $V_{\max }$ values from the two methods (4, 5) to obtain a final averaged value. Table I summarizes results for the ten tracks considered with each of the observing CYGNSS space vehicles’ (SV) mean along track signal-to-noise ratio (SNR), receiver gain G$_R$, and incidence angle $\theta _i$ listed. The results show a tendency of the retrieval method to overestimate the best track values with differences generally contained to less than 10 m/s differences, RMSE was 6.89 m/s, the unbiased RMSE was 5.20 m/s, and mean difference was 4.52 m/s, for an average truth wind speed of approximately 60 m/s. This performance is within the CYGNSS performance goal of 10% error for wind speeds greater than 20 m/s, although it is noted that the “track-based” retrieval process used is significantly different from the “snapshot” 1 s standard CYGNSS products, for which this requirement applies. A direct comparison with the standard CYGNSS Level 2 (L2) products for eight of the ten tracks listed in Table I was not possible due to the fact that the space vehicles were observing reflections from block IIF PRNs for which the current v2.1 retrieval algorithm does not provide wind speed estimates due to limited knowledge of their variable EIRP levels [21]. For tracks 8 and 10 observing reflections from block IIRM PRNs, the maximum along track CYGNSS L2 Young Seas Limited Fetch [13] Normalized Radar Cross Section based estimates were 19.2 and 67.0 m/s, respectively, resulting in a difference of $-$42.0 and 5.57 m/s difference relative to NHC estimates. The limited nature of the datasets available for comparison prevents further assessment of the relative merits of the proposed approach with the respect to that provided by the existing L2 product.

Fig. 7.

Retrieval metric profile across range of storm maximum winds across which reference DDMs are generated.

TABLE I Retrieval Results Based on CYGNSS Full DDM Observations for Hurricane Irma

SECTION V. Discussion

While the results highlight the potential of the proposed method, several areas for possible improvement remain. It is clearly important that the parametric storm model provides an accurate representation of the storm structure. While the Willoughby model is known to outperform some descriptions, it is also known to have a tendency to overestimate surface wind speeds. As an example, Fig. 8 compares the Willoughby model radial profile of wind speed with wind fields from Ocean Weather Inc. (OWI) [40]. A consistent overestimation is evident. One method to address this issue involves “tuning” the Willoughby model functions to minimize differences with OWI wind fields or other reanalyses of storms. Furthermore, while azimuthal asymmetries are included in the model used through the incorporation of translational speed/direction, the changes do not introduce large quadrant specific variations across the various storm radii ($R_{64}$, $R_{50}$, and $R_{34}$). More sophisticated models [41] account for these effects. Future efforts will aim to assess the impact of using synthetic storm models, such as [41], which capture these effects, as part of the retrieval.

Fig. 8.

Comparison of radial profile of Willoughby storm against reanalysis product.

Other practical considerations relate to the nature of the observed CYGNSS measurements. While CYGNSS full DDM Hurricane Irma measurements were made for DOYs 246, 247, 250, 251, and 253, retrievals were possible only for DOYs 246 and 247. This is due to the fact that measurements for all other days coincided with Hurricane Irma's storm location being near islands. Islands can cause DDMs that are a mixture of incoherent, low-SNR, and coherent returns that are not currently captured by the E2ES model. Future extension of the E2ES forward model to include such scenarios may allow expansion in retrieval coverage. Further improvements in retrieval performance may also be facilitated by accounting for PRN specific variability in the width of their GPS L1 C/A code delay spreading functions, as part of the forward model [42].

Future work will also include an examination of the use of the even larger DDMs available by processing CYGNSS raw I/F measurements. The larger delay ranges available in such cases may provide additional retrieval improvements, or extend the number of tracks that can be considered. However, it is important to recognize that the high data rates associated with the raw I/F mode are not amenable to frequent measurements. Future efforts may also explore the retrieval of additional storm parameters such as storm wind radii and their variation in azimuth, as in [6] and [7].

SECTION VI. Conclusion

This article described retrievals of storm maximum winds using CYGNSS full DDM measurements that had extended the simulation studies of [8] into the use of the full DDM and measured CYGNSS data. Comparisons against NHC Best Track showed the promise of the method. A major advantage of the proposed retrieval methodology was that it used the “shape” of the CYGNSS waveforms as opposed to its amplitude thereby bypassing uncertainties associated with CYGNSS absolute power calibration. The authors are continuing retrievals of storm maximum wind speeds using this approach as the archive of CYGNSS full DDM measurements of cyclones continues to increase. Several opportunities for improvement and future work have also been identified and are currently in process.

ACKNOWLEDGMENT

The authors would like to thank NASA EOSDIS Physical Oceanography Distributed Active Archive Center (PO.DAAC) at the Jet Propulsion Laboratory, Pasadena, CA, USA, for making GNSS-R data derived from the CYGNSS constellation available.