HY-2B卫星散射计神经网络多区间风速反演

以欧洲中期天气预报中心ECMWF(European Centre for Medium Range Weather Forecasts)的ERA5风场数据为真实风速参考值,利用HY-2B卫星散射计L2A数据,使用反向传播神经网络方法对风速进行了反演,分别建立了中高风速、中低风速和全风速反演模型。与基于NSCAT-4地球物理模式函数得到的L2B风速相比,在训练集中,3种网络模型反演风速的均方误差(Mean Square Error,MSE)分别达到了0.18,0.14和0.32 m/s,平均绝对值误差(Mean Absolute Error,MAE)分别达到了0.27,0.24和0.34 m/s;在测试集中,3种网络模型反演风速的均方误差(MSE)分别达到了0.54,0.27和0.46 m/s,平均绝对值误差(MAE)分别达到了0.48,0.35和0.42 m/s。研究结果表明,中高和中低风速模型优于全风速模型,其中中低风速模型反演风速的MSE和MAE最低,中高风速模型反演风速的MSE和MAE下降幅度最大;3种模型都具有良好的泛化能力。

Multiple Interval Wind Speed Inversion for HY-2B Satellite Scatterometer Based on Neural Network

Taking the ERA5 wind field data of European Centre for Medium Range Weather Forecasts(ECWMF)as the reference value of the real wind speed,and using the L2A data of HY-2B satellite scatterer,the wind speed was inverted by using the back propagation neural network method.Thereafter,the inversion models of Medium-high wind speed,Medium-low wind speed and full wind speed were established respectively.Compared with the L2B wind speed product obtained by the NSCAT-4 geophysical model functions,the mean square errors(MSE)of the three network models for wind speed inversion were 0.18,0.14 and 0.32 m/s respectively,and the mean absolute error(MAE)were 0.27,0.24 and 0.34 m/s respectively in the training set.In the test set,the MSE of the three network models for wind speed inversion were 0.54,0.27 and 0.46 m/s respectively,and the MAE were 0.48,0.35 and 0.42 m/s respectively.The results show that the medium-high and medium-low wind speed models are superior to the full wind speed model.The MSE and MAE of the medium-low wind speed model are the lowest,and the MSE and MAE of the medium-high wind speed model decreased the most.All three models have good generalization ability.