基于深度学习的南海北部MASNUM海浪数据的高分辨率重构研究

何英强; 金权; 江龙宇; 江兴杰; 王硕; 周剑; 邹仲水; 朱学明; 张守文

doi:10.12284/hyxb20250123

基于深度学习的南海北部MASNUM海浪数据的高分辨率重构研究

doi: 10.12284/hyxb20250123

1.
中山大学海洋科学学院/南方海洋科学与工程广东省实验室（珠海），广东珠海 519000
2.
海南热带海洋学院海洋科学技术学院，海南三亚 572022
3.
浙江大学海洋学院，浙江舟山 316021
4.
自然资源部第一海洋研究所，山东青岛 266061

基金项目: 南方海洋科学与工程广东省实验室（珠海）资助项目（SML2023SP202）；珠海市基础与应用基础课题研究项目（2320004002806）；智慧地球重点实验室开放基金课题（KF2023YB03-08）。

详细信息

作者简介:
何英强（2001—），男，广东省清远市人，从事海洋环境要素预报应用研究。E-mail：2569821171@qq.com

通讯作者:
朱学明，男，研究员，主要从事海洋环境预警预报研究。E-mail：zhuxueming@sml-zhuhai.cn

中图分类号: P731.33
计量
- 文章访问数: 13
- HTML全文浏览量: 13
- PDF下载量: 0
- 被引次数: 0
出版历程
- 收稿日期: 2025-09-03
- 修回日期: 2025-12-04
- 刊出日期: 2025-12-31

Deep learning-based high-resolution reconstruction of MASNUM wave data in the northern South China Sea

1.
School of Marine Sciences, Sun Yat-sen University & Southern Marine Science and Engineering Guangdong Laboratory (Zhuhai), Zhuhai 519000, China
2.
College of Marine Science and Technology, Hainan Tropical Ocean University, Sanya 572022, China
3.
Ocean College, Zhejiang University, Zhoushan 316021, China
4.
First Institute of Oceanography, Ministry of Natural Resources, Qingdao 266061, China

摘要

摘要: 海浪通常指海洋中的波动现象，极端情况下海浪高度可达20余米，与大气运动、海洋动力、热力过程以及海洋环境密切相关。为了探究海浪数值模式在高分辨率地形下计算量大、速度慢的解决方案，本文采用MASNUM海浪数值模式数据，基于深度学习算法开展对南海北部海浪的高分辨率重构研究。通过对传统线性插值方法与多种深度学习算法，如卷积神经网络、生成式对抗神经网络、图像重构的扩散模型，在南海北部海浪数据高分辨率重构上进行多方面的性能评估，结果显示：相较于传统线性插值方法，深度学习算法在挖掘海浪数据的物理变化规律中表现更佳，且图像重构的扩散模型重构效果明显优于卷积神经网络和生成式对抗神经网络，综合平均均方根误差仅为0.0103 m，表明重构的高分辨海浪数据是可靠的，为建立高分辨率海浪数据重构模型提供了新的方案建议。
- 深度学习 /
- 超分辨率 /
- 海浪模式
Abstract: Ocean waves generally refer to wave phenomena in the ocean. Under extreme conditions, wave heights can exceed 20 meters. Waves are closely related to atmospheric motion, ocean dynamics, thermodynamic processes, and the marine environment. To address the issues of high computational load and slow speed in wave numerical models under high-resolution topography, this study utilizes MASNUM wave model data and conducts high-resolution reconstruction research for waves in the northern South China Sea based on deep learning algorithms. Through comprehensive performance evaluation of traditional linear interpolation methods and various deep learning algorithms—Convolutional Neural Networks, Generative Adversarial Networks, and diffusion models for image reconstruction—in high-resolution wave data reconstruction, results show that compared to traditional linear interpolation methods, deep learning algorithms perform better in uncovering the physical variation patterns of wave data. Furthermore, the diffusion model for image reconstruction significantly outperforms both convolutional neural networks and generative adversarial networks, achieving a comprehensive average root mean square error of merely 0.0103 meters. This finding substantiates the reliability of the reconstructed high-resolution wave data and provides a novel methodological framework for establishing advanced high-resolution ocean wave data reconstruction models.
- deep learning /
- super resolution /
- numerical wave models

HTML全文

图 1 本文研究的南海北部区域范围

Fig. 1 The scope of the northern South China Sea region studied in the article

下载: 全尺寸图片幻灯片

图 2 SRCNN的模型架构和高分辨率重构流程

Fig. 2 SRCNN model architecture and high-resolution reconstruction process

下载: 全尺寸图片幻灯片

图 3 SRGAN的模型架构和高分辨率重构流程

Fig. 3 SRGAN model architecture and high-resolution reconstruction process

下载: 全尺寸图片幻灯片

图 4 DiffIR的模型架构和高分辨率重构流程

Fig. 4 DiffIR model architecture and high-resolution reconstruction process

下载: 全尺寸图片幻灯片

图 5 DiffIR架构细节

Fig. 5 Architecture details of DiffIR

下载: 全尺寸图片幻灯片

图 6 海面风速与有效波高的相关系数（a）；海面风速与有效波高的时间序列对比（b）

Fig. 6 Correlation coefficients between sea surface wind speed and significant wave height (a); time series comparison of sea surface wind speed and significant wave height (b)

下载: 全尺寸图片幻灯片

图 7 不同高分辨率重构方法的RMSE空间分布特征及时变规律

Fig. 7 Spatial distribution characteristics of RMSE and temporal variations across different high-resolution reconstruction methods

下载: 全尺寸图片幻灯片

图 8 空间平均Hs的概率分布（a，b）以及各模型重构空间平均RMSE随Hs的变化关系（c，d，e，f）

散点图采用0.1 m等间距分箱统计，横坐标标示各区间中点位置，如0.05 m代表0～0.1 m区间平均值

Fig. 8 Probability distribution of spatially averaged Hs for training set (a) and test set (b), and variation of reconstruction RMSE with spatially averaged Hs across different models (c, d, e, f)

Scatter plot data were binned at 0.1 m intervals; the x-axis indicates bin midpoints (e.g., 0.05 m corresponds to the 0 – 0.1 m bin average)

下载: 全尺寸图片幻灯片

图 9 各模型逐点重构ln (MAE)随Hs的变化关系

散点图采用0.1 m等间距分箱统计，横坐标标示各区间中点位置，如0.05 m代表0～0.1 m区间平均值

Fig. 9 The variation of point-wise ln (MAE) with Hs for different reconstruction models

Scatter plot data were binned at 0.1 m intervals; the x-axis indicates bin midpoints (e.g., 0.05 m corresponds to the 0 – 0.1 m bin average)

下载: 全尺寸图片幻灯片

图 10 11月27日海浪场重构结果时序对比

a. 各时刻真实场（I-a：00:00, II-a：06:00, III-a：12:00, IV-a：18:00）；b − e. 对应时刻重构结果（b：Bilinear，c：SRCNN，d：SRGAN，e：DiffIR）

Fig. 10 Time-series reconstruction on November 27

a. Ground truth at each time point (I-a: 00:00, II-a: 06:00, III-a: 12:00, IV-a: 18:00); b − e. reconstructed results (b: Bilinear, c: SRCNN, d: SRGAN, e: DiffIR)

下载: 全尺寸图片幻灯片

图 11 不同高分辨率重构方法的RMSE空间分布特征及时变规律

Fig. 11 Spatial distribution characteristics of RMSE and temporal variations across different high-resolution reconstruction methods

下载: 全尺寸图片幻灯片

图 12 2021年海浪场重构结果时序对比

a. 各时刻真实场（I-a：01～01, II-a：01～02, III-a：01～03, IV-a：01～04）；b−e. 对应时刻重构结果（b：Bilinear, c：SRCNN, d：SRGAN, e：DiffIR）

Fig. 12 Time-series reconstruction in 2021

a. Ground truth at each time point (I-a: 01～01, II-a: 01～02, III-a: 01～03, IV-a: 01～04); b − e. reconstructed results (b: Bilinear, c: SRCNN, d: SRGAN, e: DiffIR)

下载: 全尺寸图片幻灯片

表 1 深度学习模型超参数设置

Tab. 1 Deep learning model hyper-parameter settings

参数	值
训练轮数	100
批大小/个	8
学习率	0.001
优化器	Adam
损失函数	MSE（Mean Square Error）

下载: 导出CSV

表 2 各种模型的高分辨率重构性能评估

Tab. 2 Evaluation of high-resolution reconstruction performance across various models

模型	MAE/m ↓	R² ↑	COR ↑	RMSE/m ↓
Bilinear	0.144 2	0.974 4	0.987 5	0.221 8
SRCNN	0.028 0	0.997 3	0.998 7	0.072 4
SRGAN	0.009 3	0.999 8	0.999 9	0.014 1
DiffIR	0.005 1	0.999 9	0.999 9	0.010 3
注：↑ /↓ 表示该指标值越大/小模型性能越好；加粗数据为各列最优结果。

下载: 导出CSV

表 3 模型运行时间

Tab. 3 Model execution time

	训练阶段每轮运行时间/s	推理阶段每步长运行时间/s
SRCNN	5.557 8	0.009 7
SRGAN	13.934 9	0.010 6
DiffIR	264.650 9	0.047 7

下载: 导出CSV

表 4 各种模型的高分辨率重构性能评估

Tab. 4 Evaluation of high-resolution reconstruction performance across various models

模型	MAE/m ↓	R² ↑	COR ↑	RMSE/m ↓
Bilinear	0.084 7	0.953 4	0.977 6	0.190 5
SRCNN	0.027 2	0.990 0	0.994 9	0.089 7
SRGAN	0.016 0	0.994 2	0.997 1	0.086 5
DiffIR	0.011 5	0.995 4	0.997 7	0.063 1
注：↑ /↓ 表示该指标值越大/小模型性能越好；加粗数据为各列最优结果。

下载: 导出CSV

参考文献(19)

[1]	The Wamdi Group. The WAM model-a third generation ocean wave prediction model[J]. Journal of Physical Oceanography, 1988, 18(12): 1775−1810. doi: 10.1175/1520-0485(1988)018<1775:TWMTGO>2.0.CO;2
[2]	Tolman H L. Effects of numerics on the physics in a third-generation wind-wave model[J]. Journal of Physical Oceanography, 1992, 22(10): 1095−1111. doi: 10.1175/1520-0485(1992)022<1095:EONOTP>2.0.CO;2
[3]	Booij N, Ris R C, Holthuijsen L H. A third-generation wave model for coastal regions: 1. Model description and validation[J]. Journal of Geophysical Research: Oceans, 1999, 104(C4): 7649−7666. doi: 10.1029/98JC02622
[4]	袁业立, 潘增弟, 华锋, 等. LAGFD-WAM海浪数值模式——Ⅰ. 基本物理模型[J]. 海洋学报, 1992, 14(5): 1−7. Yuan Yeli, Pan Zengdi, Hua Feng, et al. LAGDF-WAM numerical wave model—I. basic physical model[J]. Haiyang Xuebao, 1992, 14(5): 1−7.
[5]	袁业立, 华锋, 潘增弟, 等. LAGFD-WAM海浪数值模式——Ⅱ. 区域性特征线嵌入格式及其应用[J]. 海洋学报, 1992, 14(6): 12−24. Yuan Yeli, Hua Feng, Pan Zengdi, et al. LAGFD-WAM numerical wave model—II. Characteristics inlaid scheme and its application[J]. Haiyang Xuebao, 1992, 14(6): 12−24.
[6]	杨永增, 乔方利, 赵伟, 等. 球坐标系下MASNUM海浪数值模式的建立及其应用[J]. 海洋学报, 2005, 27(2): 1−7. doi: 10.3321/j.issn:0253-4193.2005.02.001 Yang Yongzeng, Qiao Fangli, Zhao Wei, et al. MASNUM ocean wave numerical model in spherical coordinates and its application[J]. Haiyang Xuebao, 2005, 27(2): 1−7. doi: 10.3321/j.issn:0253-4193.2005.02.001
[7]	Park S C, Park M K, Kang M G. Super-resolution image reconstruction: a technical overview[J]. IEEE Signal Processing Magazine, 2003, 20(3): 21−36. doi: 10.1109/MSP.2003.1203207
[8]	Patanavijit V. Super-resolution reconstruction and its future research direction[J]. AU Journal of Technology (AU JT), 2009, 12(3): 149−163.
[9]	Merizzi F, Asperti A, Colamonaco S. Wind speed super-resolution and validation: from ERA5 to CERRA via diffusion models[J]. Neural Computing and Applications, 2024, 36(34): 21899−21921. doi: 10.1007/s00521-024-10139-9
[10]	Wang Shuo, Li Xiaoyan, Zhu Xueming, et al. Spatial downscaling of sea surface temperature using diffusion model[J]. Remote Sensing, 2024, 16(20): 3843. doi: 10.3390/rs16203843
[11]	Wan Xianci, Liu Baojian, Guo Zhizhou, et al. Super resolution mapping of scatterometer ocean surface wind speed using generative adversarial network: experiments in the southern China sea[J]. Journal of Marine Science and Engineering, 2024, 12(2): 228. doi: 10.3390/jmse12020228
[12]	Michel M, Obakrim S, Raillard N, et al. Deep learning for statistical downscaling of sea states[J]. Advances in Statistical Climatology, Meteorology and Oceanography, 2022, 8(1): 83−95. doi: 10.5194/ascmo-8-83-2022
[13]	Zhu Xiaowen, Wu Kejian, Huang Weinan. Deep learning approach for downscaling of significant wave height data from wave models[J]. Ocean Modelling, 2023, 185: 102257. doi: 10.1016/j.ocemod.2023.102257
[14]	Yu Miao, Wang Zhifeng, Song Dalei, et al. Deep learning approach for downscaling the significant wave height based on CBAM_CGAN[J]. Ocean Engineering, 2024, 312: 119169. doi: 10.1016/j.oceaneng.2024.119169
[15]	Wu Xiaoyu, Zhao Rui, Chen Hongyi, et al. GSDNet: a deep learning model for downscaling the significant wave height based on NAFNet[J]. Journal of Sea Research, 2024, 198: 102482. doi: 10.1016/j.seares.2024.102482
[16]	金权. 基于机器学习算法对海浪波高的预测及优化研究[D]. 青岛: 自然资源部第一海洋研究所, 2019. Jin Quan. Prediction and optimization of wave height based on machine learning algorithm[D]. Qingdao: First Institute of Oceanography, Ministry of Natural Resources, 2019.
[17]	Dong Chao, Loy C C, He Kaiming, et al. Image super-resolution using deep convolutional networks[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2016, 38(2): 295−307. doi: 10.1109/TPAMI.2015.2439281
[18]	Ledig C, Theis L, Huszár F, et al. Photo-realistic single image super-resolution using a generative adversarial network[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Honolulu: IEEE, 2017: 105−114.
[19]	Xia Bin, Zhang Yulun, Wang Shiyin, et al. DiffIR: efficient diffusion model for image restoration[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision. Paris: IEEE, 2023: 13049−13059.