大面积水深异常检测的条件变化自编码算法

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

针对大面积海底地形数据缺失或异常的复杂及多变性特点,结合条件变分自编码器(CVAE)与深度卷积生成对抗网络(DCGAN),构建了条件变分自编码生成对抗网络(CVAE-GAN)大面积海底伪地形的检测与剔除方法。本文方法利用条件变分自编码算法改变原有的样本分布,通过对训练样本的学习重新构建样本之间的分布规律,有效提高了高维到低维映射的稳定性;结合生成对抗网络,提高了整体算法的稳健性,最终得到较优的检测与剔除结果。采用水深格网数据进行试验,并与中值滤波法、趋势面滤波法进行比较。结果表明,本文方法在精度、稳定性及噪声稳健性方面有所提高,验证了本文方法在海底地形数据处理上具有可行性。

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.