Significant wave height estimation from shipborne marine radar data using convolutional and self-attention network

Abstract

In this paper, a fusion model based on convolution and self-attention with multi-subimage input model (CNN-SA-MS) is proposed to estimate significant wave height (SWH) from shipborne X-band radar images. The model takes multiple radar subimages as input simultaneously, which not only improves the accuracy of SWH inversion by including more information, but also avoids the restriction of selecting a single subimage in the upwind direction and dependence on external devices for wind data provision. Based on the characteristics of radar images and computational efficiency considerations, this paper selects three radar subimages as the input for the model. The comparison data from buoys and ECMWF are used for training and testing. After averaging the results of 64 radar images, the root mean square error (RMSE) and correlation coefficient (CC) of the CNN-SA-MS model are 0.197 m and 0.903, respectively. The results show that the CNN-SA-MS model improves the accuracy and stability of SWH estimation compared to single-subimage CNN regression model. For the two time periods with significant discrepancies between radar data and ECMWF predictions, we introduce satellite altimeter information as a source of reference for evaluation. The resulting analysis indicates that the significant wave height estimates generated by CNN-SA-MS model are more reliable.