GSWR-DARN:GNSS-R Sea Surface Wind Speed Retrieval Based on Data Augmentation and Residual Network

Global sea surface wind speed is a key parameter for weather forecasting and climate studies. However, retrieving it from Global Navigation Satellite System Reflectometry (GNSS-R) signals reflected from the ocean surface requires complex data processing and modelling. In addition, conventional GNSS-R methods have limited accuracy in the relatively high wind speed range of 12-20m/s. A novel model - GSWR-DARN - is presented here that combines data augmentation techniques with a residual network to improve the performance of GNSS-R wind speed retrieval. The model transforms one-dimensional (1D) data from the Cyclone Global Navigation Satellite System (CYGNSS) into two-dimensional (2D) data that can be fed into a Convolutional Neural Network (CNN), while enhancing the interconnectivity between different physical variables. By adding a residual network module, the model achieves higher accuracy and better distribution of wind speed estimates in the 0–20 m/s range than traditional methods. The model is shown to reduce the average Root Mean Square Error (RMSE) by 22.48% and increase the average Pearson correlation coefficient by 22%. It is also shown that the model significantly reduces the error distribution in the range of 12-20m/s wind speed range. The model is compared with other models of varying complexity - including Artificial Neural Networks (ANN), the CNN model, the ResNet18 model, the ResNet34 model and the Vision Transformer (Vit) model - and it is found that accuracy decreases with increasing number of residual blocks due to inherent characteristics of CYGNSS data.