Spatial-Temporal Convolutional Gated Recurrent Unit Network for Significant Wave Height Estimation From Shipborne Marine Radar Data

Spatial-temporal features are extracted from X-band marine radar backscatter image sequences via deep neural networks to estimate sea surface significant wave heights (SWHs). A convolutional neural network (CNN) is first constructed based on the pretrained GoogLeNet to estimate SWH using multiscale deep spatial features extracted from each radar image. Since the CNN-based model cannot analyze the temporal behavior of wave signatures in radar image sequences, a gated recurrent unit (GRU) network is concatenated after the deep convolutional layers from the CNN to build a convolutional GRU (CGRU)-based model, which generates spatial-temporal features for SWH estimation. Both the CNN and CGRU-based models are trained and tested using shipborne marine radar data collected during a sea trial off the East Coast of Canada, while simultaneous SWHs measured by nearby buoys are used as ground truths for model training and reference. Experimental results show that compared to the classic signal-to-noise ratio (SNR)-based method, both models improve estimation accuracy and computational efficiency significantly, with a reduction of RMSD by 0.32 m (CNN) and 0.35 m (CGRU), respectively. It is also found that under rainy conditions, CGRU outperforms SNR and CNN-based models by reducing the RMSD from around 0.90 to 0.54 m.

Automated summary

In this study, deep neural networks are utilized to extract spatial-temporal features from X-band marine radar backscatter image sequences for sea surface significant wave height estimation. The models constructed based on CNN and CGRU show improved estimation accuracy and computational efficiency, with CGRU outperforming SNR and CNN-based models under rainy conditions.