Significant Wave Height Retrieval Based on Multivariable Regression Models Developed With CYGNSS Data

Abstract

This study utilizes L1B level data from reflected global navigation satellite system (GNSS) signals from the Cyclone GNSS (CYGNSS) mission to estimate sea surface significant wave height (SWH). The normalized bistatic radar cross Section (NBRCS), the leading edge slope (LES), the signal-to-noise ratio (SNR), and the delay-Doppler map average (DDMA) are used as the key variables for the SWH retrieval. Eight other parameters, including instrument gain and scatter area, are also utilized as auxiliary variables to enhance the SWH retrieval performance. A variety of multivariable regression models are investigated to clarify the relationship between the SWH and the variables by using the following five methods: stepwise linear regression, Gaussian support vector machine, artificial neural network, sparrow search algorithm–extreme learning machine, and bagging tree (BT). Results show that, among the five regression models developed, the BT model performs the best with the root mean square error (RMSE) of 0.48 m and the correlation coefficient (CC) of 0.82 when testing one million sets of data randomly selected, while the RMSE and CC of BT model are 0.44 m and 0.73 in the 4500 National Data Buoy Center (NDBC) buoy testing dataset. Meanwhile, the BT model also has the best generalization ability, which means that it performs well in practical applications. In addition, the impacts of different input variables, the size of the training dataset, and the sea surface wind speed are also investigated. These findings are anticipated to serve as helpful guides for creating future SWH retrieval algorithms that are more advanced.