This study develops a machine learning-based approach to retrieve significant wave height (SWH) from Soil Moisture Active Passive (SMAP) radiometer data under tropical cyclone (TC) conditions, accounting for limited fetch effects. Using over 1000 SMAP measurements within 400 km of the eye from 400 TCs in 2018-2024, we employed the Symmetric Hurricane Estimates for Wind (SHEW) model to determine TC eye locations and maximum wind radii (RMW). The three machine learning methods (i.e., eXtreme Gradient Boosting (XGBoost), Random Forest (RF) and Convolutional Neural Networks (CNNs)) were trained to establish the relationship between SMAP wind speeds and WAVEWATCH-III (WW3)-hindcasted SWH, incorporating key parameters including wind speed, distance to TC eye, RMW, and TC translation speed/direction. Validation against both WW3 hindcasts and Haiyang-2 (HY-2) altimeter measurements in 2023-2024 demonstrated the superior performance of the XGBoost model compared to empirical wave-growth, RF and CNNs models, i.e., lower root mean square error (RMSE) of 0.50 m (WW3) and 0.71 m (HY-2), along with higher correlation (Cor) of 0.96 (WW3) and 0.92 (HY-2).

This letter aims to assess the applicability of synchronous wind and wave observations using Chinese Gaofen-3 (GF-3) synthetic aperture radar (SAR) during tropical cyclones (TCs) in the North West Pacific (NWP). From 2023 to 2024, 28 GF-3 images featuring visible TC eyes were available for this study. The well-known machine learning method, denoted as eXtreme Gradient Boosting (XGBoost), is implemented to retrieve sea surface wind and significant wave height (SWH) from dual-polarized GF-3 images acquired in Wide ScanSAR (WSC) mode. The inversion is supported by 20 GF-3 images collocated with re-constructed winds based on European Centre for Medium-Range Weather Forecasts (ECMWF) reanalysis (ERA-5) data and Japan Meteorological Agency (JMA) best-track data along with SWH simulations from the WAVEWATCH-III (WW3). Wind speeds inverted from additional 8 GF-3 images are validated against Soil Moisture Active Passive (SMAP) measurements, yielding a root mean squared error (RMSE) of 3.31 m/s, a correlation (COR) of 0.80, and a scatter index (SI) of 0.29. The SWHs inverted from 8 images are compared with Haiyang-2 (HY-2) altimeter data and WW3 simulations, showing an RMSE of 0.74 m, a COR of 0.83, and an SI of 0.25. Our study demonstrates the robustness of GF-3 SAR in monitoring extreme sea states.

The inversion of the one-dimensional wave spectrum from dual-polarized synthetic aperture radar (SAR) data is performed using machine learning methods, namely Random Forest (RF), eXtreme Gradient Boosting (XGBoost), Support Vector Regression (SVR) and Convolutional Neural Networks (CNNs). This process incorporates an improved hydrodynamic modulation transfer function (MTF), calibrated with more than 8000 Sentinel-1 (S-1) Ground Range Detected (GRD) images from 2021–2024, to account for shear currents induced by oceanic eddies. This study primarily aims to the differences between algorithm for one-dimension wave spectrum retrieval from S-1 SAR image as passing oceanic eddy and examine the characteristics of sea surface waves change observed by SAR passing through oceanic eddies. SAR retrievals are compared with significant wave heights (SWHs) from Haiyang-2 (HY-2) altimeter and wave spectrum from the Surface Wave Investigation and Monitoring (SWIM) and National Data Buoy Center (NDBC) buoys. The XGBoost model is established as superior for retrieving one-dimensional wave spectra and SWH, recording the lowest RMSE (0.33 m vs. SWIM, 0.30 m vs. HY-2, 0.68 m vs. Buoy) and highest correlation in validation. Applied to a North Atlantic case study, it effectively derives sea states from S-1 image. This high-resolution analysis reveals that oceanic eddies significantly amplify wave energy, resulting in elevated SWH inside their cores. While model accuracy declines in high kinetic energy eddies, the SAR-derived retrievals maintain reliability across eddy-affected regions.

In our study, machine learning is applied for wind speed retrieval in tropical cyclones (TCs) utilizing the wave spectrum measured by Surface Wave Investigation and Monitoring (SWIM) onboard the Chinese–French Oceanography SATellite (CFOSAT). These measured waves with a spatial resolution of 18 km are collocated with wind products of 0.25° spatial resolution derived from a Soil Moisture Active Passive (SMAP) microwave radiometer in the western Pacific Ocean from 2019–2021. Through our abundant dataset, we find that wind speeds up to 45 m/s are linearly correlated with significant wave height (SWH) with a 0.8 correlation (COR) and cross-zero mean wave period (MWP) with a 0.56 COR. Based on this finding, a machine learning method, denoted as Adaptive Boosting (AdaBoost), is applied to relate wind speed with two parameters (i.e., SWH and MWP). The wind speeds retrieved from SWIM-measured wave spectra are compared with the wind products obtained from SMAP radiometers in the China Seas during the TC season of 2021: we obtain a 2.78 m/s root mean square error (RMSE), a 0.85 COR, and a 0.21 scatter index (SI). These results are better than those obtained using parametric formulas among the wind-wave triplets, i.e., an RMSE > 4 m/s of wind speed, a COR0.25. We conclude that cyclonic winds and waves can be synchronously measured by SWIM without any prior information.

Heavy rain is associated with strong winds and extreme waves in a tropical cyclone (TC). In this paper, a practical algorithm for rain rate retrieval in TCs is proposed through 24 dual-polarized (vertical-vertical (VV) and vertical-horizontal (VH)) Sentinel-1 (S-1) synthetic aperture radar (SAR) images acquired in interferometric-wide (IW) swath mode, in which 13 images are collocated with the observations from stepped-frequency microwave radiometers (SFMRs). TC winds are directly obtained from VH-polarized images utilizing the geophysical model function (GMF) S-1 IW mode wind speed retrieval model after noise removal (S1IW.NR). The normalized radar cross section (NRCS) at VV-polarization channel is simulated using GMF CMOD5N and VH-polarized SAR wind. It is found that the difference between the simulated NRCSs and measurements from SAR is linearly related to the rain rate and oscillates with the incidence angle. Following this finding, an empirical algorithm for SAR rain rate retrieval is developed, denoted as CRAIN2_S1, which considers the influence of the radius of the maximum wind speed. The proposed algorithm is applied to 11 images in the dataset, and the validation of the rain rate (up to 35 mm/hr) against the products from global precipitation measurements (GPMs) has a root mean square error of 1.74 mm/hr, a correlation coefficient of 0.92 and a scatter index of 0.29. Collectively, it is concluded that the algorithm CRAIN2_S1 can be practically applied for dual-polarized SAR rain rate retrieval without any external information.

In our work, three empirical algorithms for significant wave height (SWH) retrieval have been tuned for horizontal-horizontal (HH) polarization Sentinel-1 SAR. We extracted more than ten thousand sub-scenes from available 200 images, which were treated as a dataset with collocated SWH from European Centre for Medium-Range Weather Forecasts (ECMWF) reanalysis wave data at a 0.125 ° grid. Two algorithms are based on the relation between SWH and azimuthal cutoff wavelength named CSAR_WAVEs and the other is called XWAVE which was originally explored for SWH retrieval from X-band SAR data. The three empirical algorithms have been tuned for HH-polarization Sentinel-1 SAR through the dataset. Additional 83 HH-polarization Sentinel-1 SAR images were carried out to retrieve SWH by using the four algorithms. Although the three algorithms allow estimating SWH from HH-polarization Sentinel-1 SAR, the algorithm herein called CSAR_WAVE_S is recommended for use due to less comparative error of a 0.57m STD of SWH.

Significant wave height (Hs), is an important parameter, represented as the integration of directional wave spectra. Although many researchers have directly extractedHsfrom SAR images and got a great accuracy of retrieval, those approaches are not suitable for GF-3 SAR data. In this paper, we propose an empirical approach for SARHsretrieval, using λcestimated from the real part of image cross spectra obtain from VV-polarized Gaofen-3 (GF-3) wave mode data acquired in different radar beams (called wave-code). Results using GF-3 wave mode data from January to February 2017 indicate that the bias and RMSE errors are: 189 wave-code, -0.13 m and 0.57 m; 190 wave-code, -0.07 m and 0.34 m; 193 wave-code, -0.3 m and 0.59 m; 199 wave-code, 0.16 m and 0.68 m; 215 wave-code, 0.2 m and 0.87 m. they show a relative behavior between the retrievedHsand theHsextracted from WAVEWATCH-III (WW3). However, there is a significant error when WW3-extractedHsexceed 4 m. It seems that the model is not suitable forHsretrieval on high sea conditions.

The motivation of this work is to investigate the characteristics of Normalized Radar Cross Section (NRCS) of X-band SAR under high winds using a TerraSAR-X (TS-X) image taken in Typhoon Megi in the South China Sea. NRCS derived from TS-X is compared with the simulated NRCS using the C-band GMF CMOD5 and the Ku-band GMF NSCAT4. The wind velocity is ranged from 25 to 36 m s-1in the analysis. Results show the Root-Mean-Square Error (RMSE) between the CMOD5 simulations and the observations of TS-X is 1.7 dB, while it decreases to 0.86 dB in the comparisons between NSCAT4 simulations and TS-X observations. We also investigate the effect of NRCS by rain on X-band SAR, related to high winds. Analysis shows that NRCS from NSCAT4 is closer to TS-X than CMOD5, even at high rain rates. Besides, rainfall enlarges the difference of NRCS on Ku-band, X-band and C-band at a low rain rate (-1), while rainfall lessens the difference when the rain rate is greater than 5 mm h-1. However, NRCS on Ku-band, X-band and C-band SAR tend to be approached when wind velocity increases at various rain rates.

In this study, we present the analysis of measurements of normalized radar cross section (NRCS) and the polarisation ratio (PR) in dual-polarisation TerraSAR-X (TS-X) images. Based on two PR functions proposed for C-band SAR (denoted as the Thompson model and Elfouhaily model respectively) and the relationship between NRCS and incidence angle for X-band (it is called X-PR), three PR models are tuned using 45 dual-polarisation TS-X images. The PR is found to be dependent of the incidence angle and VV-polarisation NRCS is larger than HH-polarisation NRCS when the incidence angle is larger than 23°. A total of 20 HH-polarisation TS-X images are analyzed for retrieving the wind field using three PR models and XMOD algorithm together. The results are compared to QuikSCAT data and buoy measurements for validation. The root-mean-square error (RMSE) of wind speed is 2.34m/s, 2.16m/s and 1.99m/s with a correlation of 0.82, 0.87 and 0.88 by using Thompson model, Elfouhaily model and X-PR model respectively for the cases.