HY-2B Scatterometer Wind Field Inversion Based On Neural Network

HY-2B scatterometer can observe the wind field on the global ocean surface in all weather and all day.The wind field product data of HY-2B scatterometer is widely used in weather forecast and climate research.Due to the complex influence of the physical characteristics of sea surface and the environment,the geophysical model function used for wind field inversion of scatterometer is not universal,especially in the case of low wind speed or high wind speed,the effect of wind field inversion is not ideal.In order to make better use of HY-2B scatterometer observation data and improve the application ability of HY-2B scatterometer wind field products in weather forecast and natural disaster monitoring,this dissertation analyzes the quality of HY-2B scatterometer wind field products,and finds appropriate and effective methods to improve the accuracy of HY-2B scatterometer wind field inversion.Using TAO buoy and ERA5 atmospheric reanalysis wind field data,the quality of L2B product wind field of HYB scatterometer was assessment based on statistical indexes such as mean deviation,mean absolute deviation and root mean square error.The results show that the L2B wind field meets the operational accuracy requirements of HY-2B scatterometer within the effective wind speed observation range.The MAE and RMSE of wind speed are 0.75 m·s^-1 and 1.02 m·s^-1,respectively.The MAE and RMSE of wind direction are 10.81°and 14.86°,respectively.The accuracy of L2B wind field was evaluated more carefully and comprehensively using ERA5 data.The results showed that the MAE and RMSE of wind speed were 1.07 m·s^-1 and 1.42 m·s^-1,and the MAE and RMSE of wind direction were 13.90°and 19.46°,respectively,in the range of medium and low wind speed.In the middle and high wind speed range,the MAE and RMSE of wind speed are 1.44 m·s^-1 and 1.81 m·s^-1,the MAE and RMSE of wind direction are 10.75°and 14.61°respectively.On the whole,the accuracy of wind speed of L2B is better than that of medium-high wind speed at medium-low wind speed,and with the increase of wind speed,the overestimation of wind speed of L2B is gradually changed to underestimation,while the accuracy of wind direction is better than that of medium-low wind speed at medium-high wind speed.According to the accuracy of L2B wind field products at different wind speeds,based on BP neural network,the inversion models of medium-low wind speed(4～15 m·s^-1),medium-high wind speed(15～24 m·s^-1)and total wind speed(4～24 m·s^-1)were established by using HY-2B scatterometer L2A data and ERA5 wind speed data are used.In the test set,MAE and MSE of medium-low wind speed model are 0.55 m·s^-1and 0.67 m·s^-1,respectively.MAE and MSE of medium-high wind speed model are0.44 m·s^-1 and 0.41 m·s^-1,respectively.MAE and MSE of total wind speed models are 0.57 m·s^-1 and 0.70 m·s^-1,respectively.The wind speed inversion accuracy of the three models is better than that of L2B wind speed,and they have good generalization ability.The trained neural network model is applied to the typhoon"Maysak"and"Haishen”for inversion,the results showed that the three kinds of neural network model are improved to typhoon wind velocity inversion effect,especially in the medium-low wind speed model and medium-high wind speed model used separately,the medium-high wind speed model can be more targeted to inversion of high wind speed near the typhoon center.Thus,the inversion accuracy of wind speed is improved,and the overestimation of L2B wind speed near the typhoon center area is better corrected.Based on the spatial continuity of wind field,adopt field-to-field approach,the inversion model of wind direction based on neural network is established.L2A data,output wind speed of BP neural network and ERA5 wind direction as real reference are used for neural network training.The training and test results show that the model can fit the wind direction under the wind speed of 4～24 m·s^-1,and has good generalization ability,and the wind direction inversion accuracy is better than the L2B wind direction.In the test set,MAE and MSE of model wind direction and ERA5 wind direction were0.090 rad（5.16°）and 0.065 rad（3.73°）,respectively.Finally,the trained wind direction inversion model is used to invert an anticyclone in the southern hemisphere to test the inversion capability of the model when the wind direction changes greatly.The results show that compared with the L2B wind direction,the deviation between the model wind direction and ERA5 wind direction decreases obviously,which verifies the inversion ability of the model for cyclone or anticyclone.In wind speed inversion,wind speed inversion is carried out by dividing wind speed interval,and the mutual influence of different environmental factors in each wind speed interval is reduced.In wind direction inversion,the wind direction inversion strategy of field-to-field is adopted,and the wind field in a specific range is output at the same time.Through the above inversion methods,the inversion model of wind speed and direction based on neural network is established,which effectively improves the inversion accuracy and efficiency of HY-2B scatterometer wind field.