CNN-SENet: a GNSS-R ocean wind speed retrieval model integrating CNN and SENet attention mechanism
Yimin Xia, Dongliang Guan, Zhiling Zhou
Abstract
Abstract The retrieval of sea surface wind speed is a key application of Global Navigation Satellite System-Reflectometry (GNSS-R). The continuous advancement of deep learning technologies has enabled the application of Convolutional Neural Network (CNN) models to retrieve sea surface wind speed from GNSS-R observables. However, the standard CNN models assign equal weight to all features, overlooking the more relevant ones, which reduces training efficiency and accuracy. To address this issue, this paper proposes a CNN model that incorporates the Squeeze-and-Excitation Network (SENet) attention mechanism, named CNN-SENet. The CNN-SENet model increases the weight for important features while suppressing the weight for less relevant ones, thereby improving accuracy and training efficiency. Results indicate that the CNN-SENet demonstrates a significant advantage in training efficiency over the standard CNN, reducing training time by nearly half. Additionally, the CNN-SENet model predicts the wind speeds in the range of 0–40 m/s with a Root Mean Square Error (RMSE) of 1.29 m/s and a coefficient of determination ( $$R^{2}$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"> <mml:msup> <mml:mi>R</mml:mi> <mml:mn>2</mml:mn> </mml:msup> </mml:math> ) of 62.4%. It also outperforms both the standard CNN and the Geophysical Model Function (GMF), improving wind speed accuracy by 0.14 m/s and 0.62 m/s, respectively. Furthermore, the CNN-SENet model exhibits superior temporal generalization compared to the standard CNN.