摘要
针对南海区域,使用3种重力信号(垂线偏差、重力异常、垂直重力梯度异常)训练卷积神经网络模型,并将预测结果与船测数据和国外模型结果进行对比分析。将3种重力信号分成4组数据:重力异常,重力异常与垂直重力梯度异常,重力异常与垂线偏差,以及重力异常、垂线偏差和垂直重力梯度异常。4种组合方式的反演结果与船测水深之间的标准差分别为104.780 m、102.778 m、93.788 m、88.289 m,表明随着不同类型重力数据的加入,水深预测精度明显提高,并且在深度大于2000 m时,反演结果精度提升效果更为显著。将训练集占总数据集的比例分别设置为80%、70%、60%和50%,反演结果与船测水深之间的标准差分别为88.289 m、91.256 m、92.833 m、96.022 m,表明数据量的增多可以有效提高模型学习结果的精度。
For the South China Sea region,we use three types of gravity field signals(vertical deflection,gravity anomaly,and vertical gravity gradient anomaly)to train a convolutional neural network model,which is compared and analyzed with shipborne depth and foreign models.The three gravity signals are divided into four groups:gravity anomaly,gravity anomaly combined with vertical gravity gradient anomaly,gravity anomaly combined with vertical deflection,and gravity anomaly combined with vertical deflection and vertical gravity gradient anomaly.The standard deviations between the inversion results of four combinations and shipborne depths are 104.780 m,102.778 m,93.788 m,and 88.289 m,respectively,indicating that the accuracy of bathymetry prediction improves significantly with the increase of different types of gravity data.At depth greater than 2000 m,the accuracy improvement of inversion results is more significant.By setting the proportion of training set to total dataset to 80%,70%,60%and 50%,respectively,the standard deviations between the inversion results and shipborne depths are 88.289 m,91.256 m,92.833 m and 96.022 m,respectively,indicating that the increase of data can effectively improve the accuracy of model learning results.
作者
王怀兵
万晓云
Richard Fiifi Annan
WANG Huaibing;WAN Xiaoyun;Richard Fiifi Annan(School of Land Science and Technology,China University of Geosciences,29 Xueyuan Road,Beijing 100083,China)
出处
《大地测量与地球动力学》
CSCD
北大核心
2024年第3期287-292,共6页
Journal of Geodesy and Geodynamics
基金
国家自然科学基金(42074017)。
关键词
卷积神经网络
重力场信息
深度学习
海底地形
频域法
convolutional neural network
gravity field information
deep learning
seafloor topography
frequency domain method
