Information fusion for GNSS-R wind speed retrieval using statistically modified convolutional neural network

Abstract

Listen

Translate article

Spaceborne global navigation satellite system reflectometry (GNSS-R) has recently been applied for wind speed retrieval over oceans, where the wind speed is often retrieved using features extracted from delay-Doppler map (DDM) and empirical geophysical model functions (GMFs). However, it is challenging to utilize the other factors related to the GNSS-R process, such as the geometry and sea state, as the input in GMF given their complicated effects. The use of a fully connected network (FCN) has been recently proposed, but the overfitting occurs at high wind speed due to the non-uniform wind distribution, and some information is undesirably forfeited due to artificially extracting features from DDMs. To this end, we propose a deep learning-based end-to-end modified convolutional neural network (CNN) model, which applies the cumulative distribution function (CDF) matching. A multimodal approach is utilized where the convolutional layers extract effective DDM features at first. Then, they are fused with the auxiliary information, including the geolocation of the specular point, incidence angle, range corrected gain (RCG), uncertainty of bistatic radar cross section (BRCS), and significant wave height (SWH). Further, multiple fully connected layers compose the remainder of the network's layers. At the last step, CDF matching is applied to correct the system deviation of CNN winds. We found that the root mean square error (RMSE) and bias of wind speed retrievals are 1.53 m/s and − 0.097 m/s, respectively, within 0–25 m/s when the proposed method is used. Notably, the bias is reduced by 51%, compared with the FCN architecture, while the RMSE of the retrievals at 12–25 m/s is improved by 12%. Moreover, the RMSE and bias against the incidence angle differ no more than 0.35 m/s and 0.078 m/s, while those against RCG differ less than 0.24 m/s and 0.061 m/s, respectively. Time-series analyses indicated that the wind speed retrievals of the proposed model are in line with the referenced wind speeds. We identified a temporal increase of the retrieval bias, driven by the downward trend of DDM observation in the cyclone global navigation satellite system (CYGNSS) Version 2.1 inadequately calibrated products.