College of Information Science and Engineering, Ocean University of China, Qingdao, Shandong, China.
National Satellite Ocean Application Service, State Oceanic Administration, Beijing, China.
PLoS One. 2018 Dec 17;13(12):e0208989. doi: 10.1371/journal.pone.0208989. eCollection 2018.
This paper constructed a stacked-autoencoder neural network model (SAE model) to estimate sea state bias (SSB) based on radar altimeter data. Six cycles of the geophysical data record (GDR) from Jason-1/2 radar altimeters were used as a training dataset, and the other 2 cycles of the GDR from Jason-1/2 were used for testing. The inputs to this SAE model include the significant wave height (SWH), wind speed (U), sea surface height (SSH), backscatter coefficient (σ0) and automatic gain control (AGC), and the model outputs the SSB. The model includes one input layer, three hidden layers and one output layer. The SSBs in the GDR of Jason-1/2 were obtained from a nonparametric model based on the SWH and U as input variables; thus, the model has high accuracy but low efficiency. The SSBs in the GDR of HY-2A were computed using a four-parameter parametric model that uses the SWH and U as input variables; therefore, this model's computational speed is high but its accuracy is low. Thus, we used the HY-2A radar altimeter as an unseen validation dataset to evaluate the performance of the SAE model. Then, we analyzed the contrasting results of these methods, including the differences in the SSB, explained variance, residual error and operational efficiency. The results demonstrate not only that the accuracy of the SAE model is superior to that of the conventional parametric model but also that its operational efficiency is better than that of the nonparametric model.
本文构建了一个堆叠自编码器神经网络模型（SAE 模型），基于雷达高度计数据估计海洋状态偏差（SSB）。使用 Jason-1/2 雷达高度计的六个周期地球物理数据记录（GDR）作为训练数据集，另外两个周期的 GDR 用于测试。该 SAE 模型的输入包括有效波高（SWH）、风速（U）、海面高度（SSH）、反向散射系数（σ0）和自动增益控制（AGC），模型输出 SSB。模型包括一个输入层、三个隐藏层和一个输出层。Jason-1/2 的 GDR 中的 SSB 是从基于 SWH 和 U 作为输入变量的非参数模型中获得的；因此，该模型具有高精度但低效率。HY-2A 的 GDR 中的 SSB 是使用四参数参数模型计算的，该模型将 SWH 和 U 作为输入变量；因此，该模型的计算速度快但精度低。因此，我们使用 HY-2A 雷达高度计作为未见过的验证数据集来评估 SAE 模型的性能。然后，我们分析了这些方法的对比结果，包括 SSB、解释方差、残差和操作效率的差异。结果不仅表明 SAE 模型的准确性优于传统参数模型，而且其操作效率也优于非参数模型。
