海洋环流与波动重点实验室知识库

台风是热带海洋上生成的强烈天气过程，登陆后会因其携带的强风和降雨给人民生命财产安全带来巨大的危害。我国所处的西北太平洋是全球台风发生频率最高、强度最强的区域，对台风信息的准确提取和预报显得尤为重要。本文基于1979-2021年地球静止红外卫星、大气海洋再分析数据、台风最佳路径数据集开展了西北太平洋和全球的台风信息提取和预报研究。研究内容包括：（1）基于深度学习和红外卫星图像，研发不对称台风风圈半径反演算法；（2）利用迁移学习和红外卫星图像，探索跨领域知识指导下的台风中心定位技术；（3）结合大气海洋再分析数据和红外卫星图像，展开模型知识引导的台风强度预报研究；（4）基于大气海洋再分析环境场数据和红外卫星图像，解决快速增强台风样本不均衡问题，开展快速增强台风预报研究。最终建立了一套集台风风圈半径反演、中心定位、强度预报和快速增强台风预报于一体的台风监测、预报方法。研究发现：

（1）在台风不对称风圈半径反演研究中，结果表明在VGGNet模型中引入注意力层和对应象限子图像，能有效提升不对称台风风圈半径的反演精度。不同海域的台风风圈半径具有差异性，通过分别建立北大西洋、西北太平洋、东北太平洋和南半球的DL-TCR 3（deep learning based tropical cyclone wind radius 3）模型，构成了全球台风风圈半径反演模型，其在NE、SE、SW和NW象限的R34台风风圈半径MAE（mean absolute error）分别为18.5、19.1、18.2和18.3 n mi（海里）；R50的MAE分别为11.1、10.8、10.9和10.6 n mi；R64的MAE分别为8.9、9.7、9.0和8.6 n mi。DL-TCR 3模型性能比现有台风风圈半径反演方法高出12.1 ~ 35.5%。此外，采用改进后的MAE-weight损失函数能有效减少大尺寸台风风圈半径的低估，尤其在SW象限，大于210 n mi的R34的低估被改善了51.2%。DL-TCR 3模型的可视化结果表明其能够准确学习台风的不对称结构，其中靠近台风中心的区域、低亮温的云区以及台风的螺旋结构是影响模型性能的主要因素。

（2）在台风中心定位研究中，在ResNet-TCL（ResNet based tropical cyclone location）模型中引入残差全连接模块能提升1.2%的定位精度。使用迁移学习可将定位精度进一步提高14.1%。基于迁移学习的ResNet-TCL模型对测试台风的定位误差为29.3 km，对H1-H5强度台风的平均定位误差为20 km，对H2-H5强度台风的定位误差均小于20 km，定位精度比现有台风中心定位方法高出15 ~ 45%。随着训练数据数量的增加ResNet-TCL模型性能呈对数增长。在样本较少时增加训练样本数量可以显著提高定位精度，但在训练样本较多时，使用迁移学习的收益优于增加样本数量。可视化结果显示，迁移学习模型能准确提取与台风中心相关的特征，包括台风眼、纹理和轮廓。

（3）在台风强度预报研究中，相比只输入大气风速U、V、W分量和历史台风信息，加入海表面温度和红外卫星图像可将深度学习模型的预报性能提升8.0%。结果表明，在特定等压高度下台风的时空变化特征对强度预报尤为重要。通过引入两个特殊模型分支，可以增强模型提取、融合预报因子间时空关联性的能力，有效减小预报误差。此外，通过使用模型可解释性方法可以提取模型知识（热图），利用模型知识指导建模，可以减轻环境“噪声”对模型的干扰，提高模型收敛速度和性能。MK-TCIF-fusion（model knowledge based tropical cyclone intensity forecast）模型对2020-2021年西北太平洋的24小时台风强度预报误差为3.56 m/s，与传统基于深度学习的方法相当，甚至更优。

（4）在快速增强台风预报研究中，将大气海洋三维环境场作为预报因子，能更好地捕获台风的空间结构和时间特征，进而更好的提取与快速增强台风相关的特征。使用对比学习的RITCF-contrastive-RI（contrastive-learning based rapidly intensifying tropical cyclone forecast）模型可以简化快速增强台风预报这一复杂问题，有效改善样本不均衡问题。相比其他深度学习方法，RITCF-contrastive-RI模型在维持高快速增强台风预报准确率的同时，保持较低的其他台风误报率。台风动力（U、V、PV）、热力（SST）和结构（SAT、HIS、TWR）预报因子，是RITCF-contrastive-RI模型的最佳预报因子组合。此外，使用持续学习和调整学习率的策略可以缓解模型过拟合，使其能够更好地保存学过的知识，得到更好的预报结果。RITCF-contrastive-RI模型对2020-2021年西北太平洋的快速增强台风预报准确率为85.9%，其他台风误报率为6.2%，明显优于传统预报方法。与其他深度学习方法相比，RITCF-contrastive-RI模型在相同快速增强台风预报准确率的水平下，将其他台风误报率降低了3倍。

本文基于深度学习技术在台风信息提取和预报研究中取得一系列成果，提高了台风信息提取和预报的准确性，为台风监测和预报提供可靠技术支持，将对防灾工作产生深远影响。

Typhoon is an intense weather phenomenon generated in tropical oceans. It can pose significant threats to human lives and property due to strong winds and heavy rainfall. The Northwest Pacific, where China is situated, experiences the highest frequency and intensity of typhoons, making accurate extraction and forecasting of typhoon information crucial. The research, based on geostationary satellite infrared data, atmospheric and oceanic reanalysis data, and the typhoon best track data from 1979 to 2021, conducted research on typhoon information extraction and forecasting in the Northwest Pacific and globally. The research includes: (1) Developing an asymmetric typhoon wind radius estimation method based on deep learning and satellite infrared images. (2) Developing typhoon center location technology guided by cross-domain knowledge using transfer learning and satellite infrared images. (3) Conducting typhoon intensity forecasting research by combining atmospheric and oceanic reanalysis data with satellite infrared images, guided by model knowledge. (4) Solving the problem of sample imbalance in rapid intensification typhoons using atmospheric and oceanic reanalysis data and satellite infrared images, and researching rapid intensification typhoon forecasting. Ultimately, a comprehensive and robust typhoon monitoring and forecasting method including typhoon wind radius estimation, center location, intensity forecasting, and rapid intensification typhoon forecasting was established. The research findings reveal that:

(1) In the asymmetric typhoon wind radius estimation research, introducing an attention layer and corresponding quadrant sub-images into the VGGNet model effectively enhances the accuracy of estimating the asymmetric typhoon wind radius. Typhoon wind radii in different regions exhibit distinctive characteristics. By establishing DL-TCR 3 (deep learning based tropical cyclone wind radius) models for the North Atlantic, Northwest Pacific, Northeast Pacific, and Southern Hemisphere separately, the global typhoon wind radius estimation model yields R34 typhoon wind radius MAEs (mean absolute errors) of 18.5, 19.1, 18.2, and 18.3 n mi in the NE, SE, SW, and NW quadrants, respectively. The MAEs for R50 are 11.1, 10.8, 10.9, and 10.6 n mi, while for R64, the MAEs are 8.9, 9.7, 9.0, and 8.6 n mi. The performance of the DL-TCR 3 model exceeds existing typhoon wind radius estimation methods by 12.1% to 35.5%. Additionally, employing the modified MAE-weight loss function effectively reduces the underestimation of large-sized typhoon wind radii, particularly in the SW quadrant, where the underestimation of R34 typhoon wind radii greater than 210 n mi significantly improves by 51.2%. Visual results of the DL-TCR 3 model demonstrate its ability to accurately learn the asymmetric structure of typhoons, with regions near the typhoon center, low brightness temperature cloud areas, and the spiral structure of the typhoon being the main factors influencing the model's performance.

(2) In the typhoon center location research, introducing a residual fully connected module into the ResNet-TCL (ResNet based tropical cyclone location) model significantly improves the location accuracy by 1.2%. Further accuracy enhancement of 14.1% is achieved using transfer learning. The ResNet-TCL model based on transfer learning exhibits a location error of 29.3 km for test data, an average error of 20 km for H1-H5 typhoons, and errors below 20 km for H2-H5 typhoons. The location accuracy is 15% to 45% higher than the existing typhoon center location methods. The performance of the ResNet-TCL model shows logarithmic growth with an increase in training data quantity. While adding more training samples significantly improves location accuracy when the sample size is small, the benefits of transfer learning outweigh increasing the sample size when training samples are abundant. Visual results indicate that the transfer learning model accurately extracts features related to the typhoon center, including the typhoon's eyes, textures, and contours.

(3) In the typhoon intensity forecasting study, adding sea surface temperature and infrared satellite images can enhance the performance of deep learning models by 8.0%. The results indicate that the spatiotemporal variations of typhoons at specific pressure levels are particularly crucial for intensity forecasting. The introduction of two special model branches strengthens the model's ability to extract and fuse spatiotemporal correlations among forecast factors, effectively reducing prediction errors. Furthermore, utilizing model interpretability methods to extract model knowledge (MK) and guiding modeling with MK can alleviate environmental "noise" interference, improving model convergence speed and performance. The MK-TCIF-fusion (model knowledge based tropical cyclone intensity forecast) model achieves a 24-hour typhoon intensity forecast error of 3.56 m/s for the Northwest Pacific in 2020-2021, comparable to, or even superior to, other deep learning-based methods.

(4) In the rapid intensification typhoon forecasting study, utilizing the three-dimensional atmospheric and oceanic as forecast factors enables better capturing of the spatial structure and temporal characteristics of typhoons, thereby extracting features related to rapid intensification more effectively. The RITCF-contrastive-RI (contrastive-learning based rapidly intensifying tropical cyclone forecast) model, using contrastive learning, simplifies the complex issue of rapid intensification typhoon forecasting and effectively addresses sample imbalance problems. Compared to other deep learning methods, the RITCF-contrastive-RI model maintains high accuracy in forecasting rapid intensification typhoons while keeping a lower false alarm rate for other typhoons. Forecast factors encompassing typhoon dynamics (U, V, PV), thermodynamics (SST), and structure (SAT, HIS, TWR) are identified as the optimal combination for the RITCF-contrastive-RI model. Additionally, employing continuous learning and adjusting learning rate strategies helps mitigate model overfitting, allowing for better retention of learned knowledge and yielding improved forecasting results. The RITCF-contrastive-RI model achieves an 85.9% accuracy in rapid intensification typhoon forecasting for the Northwest Pacific in 2020-2021, with a false alarm rate of 6.2% for other typhoons, significantly outperforming traditional forecasting methods. Furthermore, compared to other deep learning methods, the RITCF-contrastive-RI model reduces the false alarm rate for other typhoons by threefold at the same accuracy in rapid intensification typhoon forecasting.

The research, based on deep learning, has achieved a series of results in typhoon information extraction and forecasting. It has enhanced the accuracy of typhoon information extraction and forecasting, providing reliable technical support for typhoon monitoring and forecasting.

第1章 绪论

1.1 研究背景和意义

1.2 国内外研究进展

1.2.1 台风概述

1.2.2 风圈半径反演国内外研究现状和趋势

1.2.3 风圈中心定位国内外研究现状和趋势

1.2.4 台风强度预报国内外研究现状和趋势

1.2.5 快速增强台风预报国内外研究现状和趋势

1.3 科学问题的提出

1.4 研究内容与论文结构

第2章 台风概述和数据资料

2.1 数据介绍

2.1.1 飓风卫星数据（HurSat-B1）

2.1.2 葵花8号卫星数据（Himawari-8）

2.1.3 网格化卫星数据（GridSat-B1）

2.1.4 ERA5再分析数据（ERA5）

2.1.5 全球台风最佳路径数据集（IBTrACS）

2.1.6 中国气象局台风最佳路径数据集（CMA BT）

2.2 小结

第3章 台风风圈半径反演研究

3.1 台风风圈半径反演研究的数据预处理

3.2 台风风圈半径反演模型

3.2.1 台风风圈半径反演模型的构建

3.2.2 实验设置

3.3 不同模型框架的台风风圈半径反演结果

3.4 全球台风风圈半径反演结果

3.5 数据不平衡分布对台风风圈半径反演的影响

3.6 讨论

3.6.1 模型解释和可视化分析

3.6.2 结果分析

3.6.3 个例分析

3.7 小结

第4章 台风中心定位研究

4.1 台风中心定位研究的数据预处理

4.2 台风中心定位模型

4.2.1 台风中心定位模型的构建

4.2.2 实验设置

4.3 不同模型框架的台风中心定位结果

4.4 迁移学习对台风中心定位的影响

4.5 数据数量对台风中心定位的影响

4.6 讨论

4.6.1 结果分析

4.6.2 模型解释和可视化分析

4.7 小结

第5章 台风强度预报研究

5.1 台风强度预报研究的数据预处理

5.2 台风强度预报模型

5.2.1 台风强度预报模型预报因子筛选

5.2.2 台风强度预报模型框架设计

5.2.3 基于模型知识的台风强度预报（MK-TCIF）模型

5.3 预报因子对台风强度预报的影响

5.4 不同模型框架的台风强度预报结果

5.5 模型知识对台风强度预报的影响

5.6 讨论

5.6.1 结果分析

5.6.2 个例分析

5.7 小论

第6章 快速增强台风预报研究

6.1 快速增强台风预报研究的数据预处理

6.2 快速增强台风预报模型

6.2.1 快速增强台风预报模型的构建

6.2.2 快速增强台风预报模型预报因子筛选

6.2.3 实验设置

6.3 不同建模方案的快速增强台风预报结果

6.4 预报因子对台风强度预报的影响

6.5 持续学习策略对快速增强台风预报的影响

6.6 讨论

6.6.1 结果分析

6.6.2 误差分析

6.6.3 个例分析

6.7 小论

第7章 总结与展望

7.1 论文的主要研究成果

7.2 论文的创新点

7.3 未来工作展望

参考文献

附录  字母缩写和全称

致谢

作者简历及攻读学位期间发表的学术论文与其他相关学术成果