Widespread bathymetric outliers detection and elimination based on conditional variational autoencoder generative adversarial network

Abstract

In view of the complexity and variability of bathymetric data missing and exception, an algorithm named CVAE-GAN to detect and eliminate the widespread bathymetric outliersis proposed. Firstly, the proposed model is an alternative to traditional generative adversarial network (GAN) training methods, combined with the advantages of conditional variational autoencoder (CVAE) and deep convolutional generative adversarial network (DCGAN).Secondly, the network structure is introduced in detail.The generalized CVAE algorithm is added to change and reshape the sample distribution, having a better ability of dimensionality reduction.The GAN method improves the robustness of the whole algorithm.Thirdly,using electronic chart data containing widespread outliers, long-time experiments were carried out to train the CVAE-GAN till optimality. Finally, compared with median filtering method and trend filtering algorithm(TFA), the results show that the proposed method has an improvement in accuracy, stability and robustness.It is also verified that the feasibility of the proposedmethod in bathymetric data processing.