A Study Of Sea Surface Wind Field Retrieval For The HY-2B Scatterometer Based On Deep Learning

The sea surface wind field plays a crucial role in marine weather forecasting,natural disaster monitoring,and navigation,and is an essential physical parameter for studying the interaction between ocean and atmosphere and ocean storms.HY-2B is China’s second operational ocean dynamic environment satellite,equipped with a Kuband scatterometer that can observe global ocean surface wind fields around the clock.Its wind field product has been widely used in weather forecasting and climate change research.Currently,wind field retrieval methods widely accepted by people are pointto-point methods based on geophysical model functions,which are complicated to calculate and require fuzzy solution removal,and do not consider the continuity of the sea surface wind field.Based on the spatial continuity characteristics of the wind field,this study uses L2 A level data from the HY-2B scatterometer and reanalysis wind field data from the European Centre for Medium-Range Weather Forecasts(ECMWF)from April to June2021 to establish a sea surface wind speed retrieval model based on convolutional neural networks(CNN-SPD)and a sea surface wind speed retrieval model based on residual networks(Resnet-SPD)for the entire wind speed range(0～20m/s).Both models can simultaneously retrieve wind speed for specific sizes of continuous wind field regions.Using the test set to compare the two deep learning models with the traditional maximum likelihood method(MLE),in the test set,the bias mean(ME)of CNN-SPD was-0.0071 m/s,the mean absolute deviation(MAE)was 0.4703 m/s,and the root mean square error(RMSE)was 0.5872 m/s;the ME of Resnet-SPD was-0.0449 m/s,the MAE was 0.4766 m/s,and the RMSE was 0.5951 m/s;the ME of the MLE method was 0.0798 m/s,the MAE was 0.6352 m/s,and the RMSE was 0.7937m/s.Both models have higher wind speed retrieval accuracy than the MLE method and have good generalization capabilities.Different wind speed ranges were used to verify the retrieval ability of the three methods in different wind speeds.The results show that compared with the standard MLE method,the retrieval performance of the two deep learning models is improved for different wind speeds,especially the CNN-SPD model,which improves the RMSE by 0.2065 m/s compared with the traditional method while reducing the retrieval error in low and high wind speed regions.Using the wind speed data obtained by CNN-SPD and Resnet-SPD,this study also establishes two wind direction retrieval models based on convolutional neural networks(CNN-DIR)and residual networks(Resnet-DIR),respectively.Both models fit wind directions well.In the test set,the ME of the CNN-DIR model is 2.2778°,the MAE is7.5725°,and the RMSE is 9.7144°;the ME of the Resnet-DIR model is 2.1089°,the MAE is 6.5661°,and the RMSE is 8.4365°;the RMSE of the MLE method is 10.074°,the ME is 2.3929°,and the MAE is 7.8755°.The wind direction retrieval accuracy of the two deep learning models is better than that of the traditional method,especially when the wind speed is between 3-15m/s.The study compares the retrieval accuracy of the three methods in different wind direction areas to test the model’s retrieval ability for different wind directions.The results show that the deviation between the retrieval wind direction of the two deep learning models and ECMWF wind direction is significantly smaller than that of the traditional method,especially the Resnet-DIR model,which increases the RMSE by 1.6375° and has better model stability.This proves that the two models have the ability to retrieve wind direction for different wind directions.In wind field retrieval,taking into account the spatial correlation and continuity between adjacent observation points,and adopting a method of simultaneously inverting multiple points solves the ambiguous solution problem of traditional wind field retrieval methods.Through the above retrieval methods,this study establishes wind speed and wind direction retrieval models based on CNN and Resnet.It has been demonstrated that ResNet can be applied to microwave scatterometer sea surface wind speed retrieval,providing a new research idea for improving the retrieval effect of sea surface wind field.