于彩彩, 楚晓亮, 王曙曜, 2024. 基于卷积神经网络的高频地波雷达有效波高反演[J]. 海洋科学进展, 42(1): 126-136. doi: 10.12362/j.issn.1671-6647.20221027001.
引用本文: 于彩彩, 楚晓亮, 王曙曜, 2024. 基于卷积神经网络的高频地波雷达有效波高反演[J]. 海洋科学进展, 42(1): 126-136. doi: 10.12362/j.issn.1671-6647.20221027001.
YU C C, CHU X L, WANG S Y, 2024. Significant wave height inversion of high frequency surface wave radar based on convolutional neural network[J]. Advances in Marine Science, 42(1): 126-136. doi: 10.12362/j.issn.1671-6647.20221027001
Citation: YU C C, CHU X L, WANG S Y, 2024. Significant wave height inversion of high frequency surface wave radar based on convolutional neural network[J]. Advances in Marine Science, 42(1): 126-136. doi: 10.12362/j.issn.1671-6647.20221027001

基于卷积神经网络的高频地波雷达有效波高反演

Significant Wave Height Inversion of High Frequency Surface Wave Radar Based on Convolutional Neural Network

  • 摘要: 高频地波雷达海面回波多普勒谱中蕴含着非常丰富的海浪信息,针对经验模型通常存在对回波谱信息利用不充分的问题,本文提出了一种利用卷积神经网络反演波高的方法。首先基于雷达后向散射截面方程和有效波高反演的参数化经验模型,并结合实测数据分析,本文选取了二阶与一阶谱能量比值、二阶谱能量和、二阶与一阶谱峰值比、Bragg频率处无向波高谱值、二阶谱能量、二阶谱峰值、左右一阶谱峰值及其比值、1/\sqrt 2 布拉格频率处峰值、一阶谱能量共11种与有效波高相关的回波谱特征参数;为进一步说明一阶和二阶信息对有效波高的作用,将11种特征参数分成4个组合,分别搭建了多层深度卷积神经网络并进行高频地波雷达有效波高反演;然后将高、低两种海况下多层深度卷积神经网络模型与经验模型反演结果进行对比分析。研究结果表明,利用11种特征参数构建的模型波高反演精度更高,在高、低海况测试集中雷达反演有效波高与浮标波高序列相关系数R分别为0.92和0.78,均方根误差(Root Mean Square Error, RMSE)分别为0.32 m和0.21 m,平均绝对误差(Mean Absolute Error, MAE)分别为0.25 m和0.16 m,平均相对误差(Mean Relative Error, MRE)分别为0.12和0.27。综合利用这11种特征参数,能够提高雷达在复杂海况下对有效波高的反演精度。

     

    Abstract: The Doppler spectrum of sea surface echo from High Frequency Surface Wave Radar (HFSWR) contains very rich wave information. In order to solve the problem of inadequate use of echo spectrum information in empirical models, a method of wave height inversion using convolutional neural network is proposed in this paper. In this paper, based on the radar backscattering cross section equation and the parametric empirical model of Significant Wave Height (SWH) inversion, combined with the analysis of measured data, 11 kinds of echo spectral characteristic parameters related to SHW are selected, including the ratio of second-order to first-order spectral energy, the sum of second-order spectral energy, the ratio of second-order to first-order peak power, the undirected peak power at the Bragg frequency, the second-order spectral energy, the second-order peak power, the left and right first-order peak power and their ratios, the peak power at the 1/ \sqrt2 Bragg frequency, and the first-order spectral energy. Secondly, in order to further explain the role of first and second order information on SWH, the 11 characteristic parameters are divided into 4 combinations, and multi-layer deep convolutional neural networks are built to retrieve the SWH. Then, the inversion results of the multi-layer deep convolutional neural network model and the empirical model are compared and analyzed in high and low sea states. The results show that the inversion accuracy of SWH based on 11 characteristic parameters is higher. In the high and low sea state test sets, the correlation coefficients (R) of SWH retrieved by radar and by buoy are 0.92 and 0.78, respectively, and the Root Mean Square Error (RMSE) is 0.32 m and 0.21 m, respectively. The Mean Absolute Error (MAE) is 0.25 m and 0.16 m, and the Mean Relative Error (MRE) is 0.12 and 0.27, respectively. The comprehensive use of these 11 characteristic parameters can improve the inversion accuracy of significant wave height under complex sea conditions.

     

/

返回文章
返回