基于生成对抗网络模型的热带和亚热带海洋中尺度涡预报研究
刘爽(1996—), 女, 黑龙江省七台河市人, 硕士研究生, 从事海洋过程与大数据研究。email: |
Copy editor: 殷波
收稿日期: 2021-09-07
修回日期: 2021-12-27
网络出版日期: 2021-12-28
基金资助
国家自然科学基金(92058201)
国家自然科学基金(41776040)
国家自然科学基金(41949907)
国家自然科学基金(42149907)
中国科学院基础前沿科学研究计划原始创新项目(ZDBS-LY-DQC011)
广州市科学研究计划(201904010420)
Predicting the mesoscale eddy in the tropical and subtropical ocean based on generative adversarial network model
Copy editor: YIN Bo
Received date: 2021-09-07
Revised date: 2021-12-27
Online published: 2021-12-28
Supported by
National Natural Science Foundation of China(92058201)
National Natural Science Foundation of China(41776040)
National Natural Science Foundation of China(41949907)
National Natural Science Foundation of China(42149907)
Original Innovation Project of Basic Frontier Scientific Research Program of CAS(ZDBS-LY-DQC011)
Guangzhou Science and Technology Project(201904010420)
中尺度涡蕴含海洋超过90%的动能, 显著影响海洋物质能量循环。对中尺度涡的预报是目前物理海洋学研究的热点和难点。文章基于卫星高度计观测的近30年海表面高度异常数据(sea level anomaly, SLA), 采用基于博弈思想的生成对抗网络方法(generative adversarial networks, GAN), 构建了中尺度涡预报模型, 进行了28天预报, 并采用独立样本分析了预报涡旋的空间分布、时间分布、能量强度等特征参数, 探讨影响预报结果准确性和时效性的主要因素。结果表明, 半径为100~200km的涡旋在15天左右的预报时长仍能保持较好的准确性及时效性, 误差在20%以内。该区域的平均涡动能约为0.875m2·s-2, 其预报的均方根误差(root mean square error, RMSE)普遍介于0.02~0.04m2·s-2。且涡旋预报结果受异常天气影响较小, 在正常天气条件和台风娜基莉条件下具有相似的预报能力。这些结果对进一步理解并应用生成对抗网络这一新方法预报海洋中尺度涡提供了参考。
刘爽 , 经志友 , 詹海刚 . 基于生成对抗网络模型的热带和亚热带海洋中尺度涡预报研究[J]. 热带海洋学报, 2022 , 41(5) : 1 -16 . DOI: 10.11978/2021121
Mesoscale eddies occupy more than 90 % of the kinetic energy in oceans, which significantly impact on the oceanic mass and energy cycle. The prediction of mesoscale eddies remains a very essential, but difficult research topic in the current physical oceanography field. Based upon the sea level anomaly (SLA) data measured by satellite altimeter in the last 30 years, this study develops a model for the mesoscale eddy prediction according to the generative adversarial networks (GAN) method and game theory. The results of 28-day prediction, their spatial-temporal distributions, and energy intensities are analyzed with independent samples. The main factors that affect the spatial and temporal accuracies are discussed. The results show that the temporal accuracy of this method can be accepted in about 15 days. For the mesoscale eddies with a radius of 100 ~ 200 km, the prediction error is generally less than 20 % by this method. The mean eddy kinetic energy in the study domain is about 0.875 m2·s-2 and the root mean square error (RMSE) is roughly between 0.02 ~ 0.04 m2·s-2. Furthermore, the results are suggestive of that the prediction is less affected by abnormal weather, and has similar forecasting ability under normal weather conditions and typhoon Nakri. These results provide a reference for further understanding and applying the new method of generative adversarial networks to predict ocean mesoscale eddies.
图2 模型示意图a. 训练过程。从训练集中随机选取不同长度的时间序列, 并对每一个时间序列进行随机子区域裁剪, 然后在G中训练并生成数据, 在D中判别数据的真实性; b. 验证过程。与训练过程同步进行, 将验证集中完整的研究区域数据输入到模型进行预报, 输出预报数据; c. 测试过程。待模型训练完成之后进行, 用去掉陆架部分的测试集数据输入模型中进行预报, 得到最终的预报数据。SLA为海表面高度异常; G为生成卷积模型; D为鉴别卷积模型 Fig. 2 Diagram of how the model works. (a) The training process, time series of different lengths are randomly selected from the training set, and each time series is cropped with the random subregion, then trained in G and generated data, the authenticity of the data is evaluated in D; (b) The verification process is carried out synchronously with the training process. The complete research data in the verification set is input into the model for prediction; (c) The test process is carried out after the model training is completed. The test set data with removed the continental shelf is input into the model for prediction to obtain the final prediction data. SLA refers to sea level anomaly; G is the generator; D is the discriminator |
图3 SLA逐日平均绝对误差(MAE)空间分布图a. 预报第7天时的日平均MAE, 矩形框对应于 Fig. 3 Spatial distribution of the SLA daily MAE. (a) Daily average MAE on the 7th day of forecast. The rectangular box corresponds to each region in |
图6 反气旋涡各振幅区间的涡旋数量日分布柱高为真实数量, 其中蓝色部分为预报数量 Fig. 6 Daily distribution of eddy number in each amplitude range of anticyclone. The column height is the real quantity, and the blue part is the predicted quantity |
图8 反气旋涡各半径区间的涡旋数量日分布柱高为真实数量, 其中蓝色部分为预报数量 Fig. 8 Daily distribution of eddy number in each radius range of anticyclone. The column height is the real quantity, and the blue part is the predicted quantity |
图10 涡旋位置相关统计图a. 平均预报准确率; b. 被准确预报的涡旋中, 真实涡旋与预报涡旋的平均最小涡中心距离; c. 被准确预报的涡旋中, 不同振幅范围的涡旋数量, 柱高表示数量; d. 被准确预报的涡旋中, 不同半径范围的涡旋数量, 柱高表示数量 Fig. 10 Different statistics of the eddy position. (a) Average prediction accuracy; (b) The mean minimum eddy center distance between the real and predicted results in the accurately predicted eddies; (c) In the accurately predicted eddies, the number of eddies in different amplitude ranges, and the column height represents the number; (d) In the accurately predicted eddies, the number of eddies in different radius ranges, and the column height represents the number. The abscissa of all subgraphs stands for days |
图11 涡动能真实场与预报场的差值空间分布a. 预报第7天时的差值空间分布; b. 预报第14天时的差值空间分布; c. 预报第21天时的差值空间分布; d. 预报第28天时的差值空间分布。该图基于国家测绘地理信息局标准地图服务网站下载的审图号为GS(2016)1561的标准地图制作 Fig. 11 Spatial distribution of the difference between observation and the prediction field of the eddy kinetic energy. (a) Spatial distribution of the difference on the 7th day of prediction; (b) Spatial distribution of the difference on the 14th day of prediction; (c) Spatial distribution of the difference on the 21st day of prediction; (d) Spatial distribution of the difference on the 28th day of prediction |
图12 正常天气条件(a~e)与台风天气条件(f~j)的真实涡旋与预报涡旋的分布图a~e分别对应2019年10月1日至2019年10月5日正常天气条件; f~j分别对应2019年11月1日至2019年11月5日台风娜基莉(Nakri)条件。图中浅红色虚线、红色实线、浅蓝色虚线、蓝色实线分别表示真实反气旋、预报反气旋、真实气旋、预报气旋的分布情况, 粉色点线为Nakri中心位置及移动路径。该图基于国家测绘地理信息局标准地图服务网站下载的审图号为GS(2016)1561的标准地图制作 Fig. 12 Distribution of real eddies and predicted eddies under normal weather conditions and typhoon weather conditions. (a~e) correspond to normal weather conditions from October 1, 2019 to October 5, 2019 respectively; (f~j) correspond to typhoon Nakri conditions from November 1, 2019 to November 5, 2019 respectively. The light red dotted line, red solid line, light blue dotted line and blue solid line in the figure represent the distribution of real anticyclone, predicted anticyclone, real cyclone and predicted cyclone respectively. The pink dotted line is the central position and moving path of Nakri |
[1] |
白杨, 李威, 邵祺, 2020. 基于经验正交函数和机器学习的南海海面高度异常预测[J]. 海洋通报, 39(6): 678-688.
|
[2] |
陈敏, 侯一筠, 赵保仁, 2003. 冬季东中国海环流中的中尺度涡旋数值模拟[J]. 海洋科学, 27(1): 53-60.
|
[3] |
崔凤娟, 2015. 南海中尺度涡的识别及统计特征分析[D]. 青岛: 中国海洋大学.
|
[4] |
江璟瑜, 徐丹亚, 韩宁生, 等, 2021. 基于LSTM的海表面高度异常预测方法[J]. 舰船电子工程, 41(2): 97-99.
|
[5] |
李佳讯, 张韧, 陈奕德, 等, 2011. 海洋中尺度涡建模及其在水声传播影响研究中的应用[J]. 海洋通报, 30(1): 37-46.
|
[6] |
王桂华, 苏纪兰, 齐义泉, 2005. 南海中尺度涡研究进展[J]. 地球科学进展, 20(8): 882-886.
|
[7] |
吴新荣, 韩桂军, 张学峰, 等, 2012. 人工神经网络在南海近海面气温反演中的应用研究[J]. 热带海洋学报, 31(2): 7-14.
|
[8] |
肖汶斌, 刘巍, 程兴华, 等, 2018. 海洋中尺度涡水声场特性分析[C]// 2018年全国声学大会论文集:B水声物理. 北京: 中国声学学会: 12-13.
|
[9] |
张盟, 杨玉婷, 孙鑫, 等, 2020. 基于深度卷积网络的海洋涡旋检测模型[J]. 南京航空航天大学学报, 52(5): 708-713.
|
[10] |
郑全安, 谢玲玲, 郑志文, 等, 2017. 南海中尺度涡研究进展[J]. 海洋科学进展, 35(2): 131-158.
|
[11] |
周水华, 洪晓, 梁昌霞, 等, 2020. 基于人工神经网络的台风浪高快速计算方法[J]. 热带海洋学报, 39(4): 25-33.
|
[12] |
|
[13] |
|
[14] |
|
[15] |
|
[16] |
|
[17] |
|
[18] |
|
[19] |
|
[20] |
DYELAX, 2021. Adversarial video generation[CP/OL]. https://github.com/dyelax/Adversarial_Video_Generation.
|
[21] |
|
[22] |
|
[23] |
|
[24] |
|
[25] |
|
[26] |
|
[27] |
|
[28] |
|
[29] |
|
[30] |
|
[31] |
|
[32] |
|
[33] |
|
[34] |
|
[35] |
|
[36] |
|
[37] |
|
[38] |
|
[39] |
|
[40] |
|
[41] |
|
[42] |
|
[43] |
|
[44] |
|
[45] |
|
[46] |
|
[47] |
|
[48] |
|
[49] |
|
[50] |
|
[51] |
|
[52] |
|
[53] |
|
[54] |
|
/
〈 | 〉 |