Machine Learning-Based Algorithm for SAR Wave Parameters Retrieval During a Tropical Cyclone

The major objective of our research is to retrieve wave parameters from synthetic aperture radar (SAR) images during a tropical cyclone (TC) based on a machine learning method. In this study, more than 2000 Sentinel-1 images obtained in interferometric-wide and extra wide mode are collected during 200 TCs, which are collocated with hindcasted waves by a third-generation numeric model, namely WAVEWATCH-III (WW3). It is found that wave parameters, i.e., significant wave height (SWH), mean wave period (MWP), and mean wave length (MWL), are correlated with several SAR-measured image variables. Based on these findings, a machine learning method, namely eXtreme Gradient Boosting (XGBoost), is developed through the training dataset using 1600 images. The trained algorithm is tested over 400 images and the retrievals are compared with WW3 simulations. The statistical analysis shows that the root mean squared error (RMSE) and scatter index (SI) of SWH are 0.19 m and 0.06, respectively. The RMSE and SI of MWP are 0.19 s and 0.03, respectively. The RMSE of the MWL is 3.77 m and the SI is 0.04. Comparisons between inverted SWH by XGBoost methods and the altimeter measurements presents a 0.59 m RMSE of SWH with and 0.19 SI. This result is improved comparing to the results (i.e., a 1.44 m RMSE of SWH with a 0.45 SI) achieved by a previous algorithm. Collectively, it is considered that machine learning is a valuable method to extract wave parameters from dual-polarization SAR images.