基于卷积神经网络的GNSSGR海面风速反演方法研究  

GNSS-R ocean wind speed retrieval method based on convolutional neural network

作　　者：陈趁新 杨志[1] 王晓宇[1] 白照广[1] CHEN Chenxin;YANG Zhi;WANG Xiaoyu;BAI Zhaoguang(DFH Satellite Co.,Ltd,,Beijing 100094,China)

机构地区：[1]航天东方红卫星有限公司,北京100094

出　　处：《先进小卫星技术（中英文）》2024年第4期8-13,共6页Advanced Small Satellite Technology

摘　　要：传统基于地球物理模型函数(geophysical model function,GMF)的全球导航卫星系统反射测量(global navigation satellite system reflectometry,GNSS-R)海面风速反演存在特征提取准确度低、模型复杂度高等问题。针对上述问题,提出了一种基于卷积神经网络的GNSS-R海面风速反演方法。通过构建卷积模块自动提取时延-多普勒映射图像(delay-Doppler map,DDM)中的观测特征,特征融合模块将提取的特征与辅助特征关联,全连接模块将上述特征向量逐级映射到海面风速。以“捕风一号”卫星观测数据为例验证了上述方法的有效性,较传统GMF方法,风速反演精度在均方根误差(root mean square error,RMSE)和平均偏差(mean bias error,MBE)上分别降低了0.51 m/s和0.19 m/s,反演效果分别提升了21%和16%。试验结果表明:该方法能够有针对性地自动提取DDM特征,有效提高特征提取的精度,同时显著降低模型的复杂度。本研究为同类卫星各种地表参数反演提供了新思路。The traditional global navigation satellite system reflectometry(GNSS-R)ocean wind speed retrieval based on geophysical model function(GMF)has the problems of low feature extraction accuracy and high model complexity.To solve these problems,a GNSS-R ocean wind speed retrieval method based on convolutional neural network was proposed.The observation features in delay-Doppler map(DDM)images were automatically extracted by constructing a convolutional module.The feature fusion module correlated the extracted features with auxiliary features,and the full connection module mapped the above feature vectors to the ocean wind speed step by step.The validity of the above method was verified using the BuFeng-1 Satellite observation data as an example.Compared with the traditional GMF method,the root mean square error(RMSE)and mean bias error(MBE)are reduced by 0.51 m/s and 0.19 m/s,respectively,which improve the retrieval effect by 21%and 16%.Experimental results show that the proposed method can automatically extract the DDM features,effectively improve the accuracy of feature extraction,and significantly reduce the complexity of the model.This study provides a new idea for retrieving ocean surface parameters from similar satellites.

关 键 词：深度学习 GNSS-R “捕风一号”卫星 海面风速反演 卷积神经网络 

分 类 号：TP75[自动化与计算机技术—检测技术与自动化装置]