Item

Abstract

Ocean surface wind is vital to the Earth’s meteoro-
logical system, and their properties can be detected by spaceborne
global navigation satellite system reflectometry (GNSS-R) mea-
surements. With the growing number of GNSS-R signal sources,
machine learning technology exhibits prominent advantages in
wind speed estimation. Currently, the deep learning techniques
that establish relationships between GNSS-R measurements and
ocean surface wind speeds generally apply grids and sequence
structures and lack flexibility and robustness. Additionally, con-
structing models with individual GNSS-R observations results in
the loss of valuable temporal correlation within delay–Doppler
maps (DDMs). Therefore, this study proposes a novel spa-
tiotemporal graph-based deep neural network (STG-DNN) for
retrieving wind speed, which incorporates a graph module with a
transformer module to fully exploit the spatial–temporal depen-
dencies of DDMs. Results demonstrate that the graph module
significantly improves both the accuracy and reliability in wind
speed retrieval. Meanwhile, the transformer module effectively
captures temporal features from various DDMs. Validations with
cyclone GNSS (CYGNSS) test data support the superior accuracy
of STG-DNN, revealing a correlation coefficient of 0.92 for the
wind speeds. The results indicate that the root-mean-square error
(RMSE) of STG-DNN for wind speed is 1.27 m/s, representing
improvements of approximately 33.2%, 20.6%, and 13.6% over
the minimum variance estimator (MVE), convolutional neural
network (CNN), and vision graph (VIG) neural networks, respec-
tively. Additionally, a promising agreement is observed between
STG-DNN and ERA5 in the spatial distributions of wind speed
retrieval, indicating a robust spatial performance in STG-DNN.
As for the temporal scale, the daily variations in retrieval
accuracy of STG-DNN exhibit smaller fluctuations compared to
both CNN and VIG wind data in the test dataset.