关键词: 深度学习, 神经网络, GNSS高程转换, 精度分析

Abstract: It is of great practical value to transform the GNSS geodetic height into the normal height with higher precision in the mountains. This paper uses the high-precision GNSS leveling data of the GSVS2017 project to analyze the accuracy of GNSS height transformation based on deep BP/ELMAN neural network, general regression neural network (GRNN), radial basis function neural network (RBFNN), support vector machine regression (SVR), quadratic curve fitting and surface fitting, etc. The results show that: ①When the distance between training points is 50, 30, 15, 10 and 5 km, the deep BP/ELMAN neural network with the hidden layer activation function ReLU can obtain higher precision results, and they are more accurate than GRNN, RBFNN, SVR, quadratic curve fitting and surface fitting. ②The deep BP/ELMAN neural network with the hidden layer activation function ReLU are used for GNSS height transformation. Among the five kinds of training point spacing, more than 90% of the height difference can meet the fourth-order leveling accuracy, and more than 75% of height difference can meet the third-order leveling accuracy; when the distance between training points spacing is 5 km, more than 55% of the height difference can meet the second-order leveling accuracy.

Key words: deep learning, neural network, GNSS height transformation, accuracy analysis