Global navigation satellite system reflectometry (GNSS-R) has shown a capability in recent years to be applied as a novel remote sensing technique to retrieve ocean wind speeds. The combination of GNSS-R observable delay-Doppler maps (DDMs) and deep learning algorithms provides the possibility to build an end-to-end pipeline for improving wind speed estimations. Recent studies have proven that data-driven approaches can be applied to generate enhanced estimation products. However, these are usually trained with convolutional neural networks (CNNs), which include inductive bias throughout the models. The inbuilt translation equivariance in CNNs represents an inexactitude for the feature extraction on DDMs. To address this issue, we propose Transformer-based models, named DDM-Former and DDM-Sequence-Former (DDM-Seq-Former), to exploit delay-Doppler correlation within and between DDMs, respectively. The advantages of our methods over conventional retrieval algorithms and other deep learning-based approaches are presented based on the Cyclone GNSS (CYGNSS) version 3.0 dataset. In addition, a comprehensive study on the attention mechanism for our models is demonstrated. The proposed DDM-Former yields the best overall performance with a root mean square error (RMSE) of $1.43$ $\mathrm{m}/\mathrm{s}$ and a bias of $-0.02$ $\mathrm{m}/\mathrm{s}$ over the nine months test period. Moreover, with an RMSE of $2.89\mathrm{m}/\mathrm{s}$ and a bias of $-1.88$ $\mathrm{m}/\mathrm{s}$, the proposed DDM-Seq-Former can promisingly improve the estimations in the cases with wind speeds higher than $12$ $\mathrm{m}/\mathrm{s}$. There are still opportunities for further enhancements in creating more robust models that could perform well in all wind regimes given a non-uniform wind distribution. We will make our code publicly available.

近年来，全球导航卫星系统反射计 (GNSS-R) 已显示出可作为一种新型遥感技术用于反演海洋风速的能力。GNSS-R 可观测延迟多普勒地图 (DDM) 和深度学习算法的结合提供了构建端到端管道以改进风速估计的可能性。最近的研究证明，数据驱动的方法可用于生成增强的估算产品。然而，这些通常是用卷积神经网络 (CNN) 训练的，其中包括整个模型的归纳偏差。CNN 中内置的平移等变性代表了 DDM 上特征提取的不精确性。为了解决这个问题，我们提出了基于 Transformer 的模型，命名为 DDM-Former 和 DDM-Sequence-Former (DDM-Seq-Former)，分别利用 DDM 内部和之间的延迟多普勒相关性。基于 Cyclone GNSS (CYGNSS) 3.0 版数据集，展示了我们的方法相对于传统检索算法和其他基于深度学习的方法的优势。此外，还展示了对我们模型的注意力机制的全面研究。所提出的 DDM-Former 产生了最佳的整体性能，均方根误差 (RMSE) 为$1.43$ $\mathrm{米}/\mathrm{秒}$和偏见$-0.02$ $\mathrm{米}/\mathrm{秒}$在九个月的测试期间。此外，RMSE 为$2.89\mathrm{米}/\mathrm{秒}$和偏见$-1.88$ $\mathrm{米}/\mathrm{秒}$，所提出的 DDM-Seq-Former 可以在风速高于$12$ $\mathrm{米}/\mathrm{秒}$. 在风分布不均匀的情况下，仍然有机会进一步增强创建更强大的模型，这些模型可以在所有风况下表现良好。我们将公开我们的代码。