The comprehensiveness of the raw input data and the effectiveness of feature engineering are two key factors affecting the performance of machine learning. To improve the data comprehensiveness for Global Navigation Satellite System Reflectometry (GNSS-R) ocean wind speed retrieval, this article introduces a new input data structure, which is composed of Delay–Doppler maps (DDM) and all satellite receiver status (SRS) parameters. Then, to overcome the difficulty of handcrafted feature engineering and effectively fusion the information of DDM and SRS, we presented a heterogeneous multimodal deep learning (HMDL) method to retrieve the wind speed according to the heterogeneity of the input data. The proposed model is verified by the performance evaluation of realistic data sets obtained from TDS-1. The new input data structure improves the prediction accuracy at 13.5% to 30.7% on mean absolute error (MAE) at 10.6% to 29.5% on the root mean square error (RMSE). The HMDL improves the prediction accuracy at 7.7% on MAE and 7.1% on RMSE. The whole proposed solution improves the prediction accuracy at 36.3% on MAE and 36.8% on RMSE, comparing with the traditional neural network-based solution. The results clearly show that both the introduction of the new input data structure and HMDL effectively improve the accuracy and robustness of GNSS-R wind speed retrieval.

中文翻译：

---

用于异构 GNSS-R 数据融合和海洋风速检索的多模态深度学习

原始输入数据的全面性和特征工程的有效性是影响机器学习性能的两个关键因素。为了提高全球导航卫星系统反射计（GNSS-R）海洋风速反演的数据综合性，本文引入了一种新的输入数据结构，它由延迟-多普勒图（DDM）和所有卫星接收器状态（SRS）参数组成。 . 然后，为了克服手工特征工程的困难并有效融合DDM和SRS的信息，我们提出了一种异构多模态深度学习（HMDL）方法，根据输入数据的异质性来检索风速。通过对从 TDS-1 获得的真实数据集的性能评估来验证所提出的模型。新的输入数据结构将平均绝对误差 (MAE) 的预测精度提高了 13.5% 至 30.7%，均方根误差 (RMSE) 的预测精度提高了 10.6% 至 29.5%。HMDL 在 MAE 上提高了 7.7% 的预测精度，在 RMSE 上提高了 7.1%。与传统的基于神经网络的解决方案相比，整个提出的解决方案在 MAE 上的预测精度提高了 36.3%，在 RMSE 上提高了 36.8%。结果清楚地表明，新的输入数据结构和HMDL的引入都有效地提高了GNSS-R风速反演的准确性和鲁棒性。与传统的基于神经网络的解决方案相比。结果清楚地表明，新的输入数据结构和HMDL的引入都有效地提高了GNSS-R风速反演的准确性和鲁棒性。与传统的基于神经网络的解决方案相比。结果清楚地表明，新的输入数据结构和HMDL的引入都有效地提高了GNSS-R风速反演的准确性和鲁棒性。