The DSAT structure analysis dataset for 1981−2020 global TCs

Climate change has been linked to tropical cyclone (TC) poleward migration, TCextreme precipitation, and an increased proportion of major hurricanes. Understandingpast TC trends and variability is critical for projecting future TC impacts on humansociety. However, past trends in TC structure and energy remain uncertain due tolimited observations; subjection analyses and inherent spatiotemporal heterogeneity ofbest-track data have led to reduced confidence in the assessed TC responses to thechanging climate. The DSAT uses deep learning to reconstruct past "observations" andyield an objective global TC wind profile dataset during 1981‒2020, facilitating acomprehensive examination of TC intensity (Vmax), structure, and integrated kineticenergy (IKE). The proposed model converts multichannel satellite imagery to a 0-750-km wind profile of axisymmetric surface winds; it is trained based on labeled dataintegrating best tracks and numerical model reanalysis of 2004–2018 TCs.The DSAT model (Chen et al. 2025) features a hybrid GAN-CNN architecture (Chen B et al., 2021a), incorporating a generative adversarial network (GAN, Goodfellow et al., 2014) module and a polar convolutional neural network (CNN, Krizhevsky et al., 2012; Chen B et al., 2021b) regressor. The DSAT GAN module is trained to generate synthetic VIS images (VIS*) based on all-time available infrared (IR) and water vapor (WV) images, enabling the indirect take advantage of original VIS images to train the CNN regressor. On the other hand, the CNN regressor, employing a "polar-convolution filter" design (Chen B et al., 2021b), extracts essential features from satellite images and regresses these features to labeled wind profiles of azimuthally averaged winds.Chen, B. F., Chen, B., Hsiao, C. M., Teng, H. F., Lee, C. S., &amp; Kuo. H. C. (2025). Constructing Deep Learning Datasets to Reveal Climate Trends of Tropical Cyclone Intensity and Structure Extremes. Artificial Intelligence for the Earth Systems (accepted with revision).<br>

气候变化已被证实与热带气旋（Tropical Cyclone, TC）向极迁移、极端降水以及强飓风占比提升存在关联。厘清热带气旋过往的变化趋势与变异性特征，对于预估其未来对人类社会的影响至关重要。然而，由于观测资料有限，加之主观分析以及最佳路径数据（best-track data）固有的时空异质性，学界对热带气旋响应气候变化的评估结果可信度有所降低。本数据集DSAT采用深度学习方法重构历史"观测数据"，并构建了1981–2020年间全球热带气旋风场廓线的客观数据集，可支撑对热带气旋强度（Vmax）、结构以及综合动能（integrated kinetic energy, IKE）的全面分析。所提出的模型可将多通道卫星影像转换为0–750公里范围内的轴对称地表风场廓线，其训练基于整合了2004–2018年热带气旋最佳路径资料与数值模式再分析数据的标注数据集。DSAT模型（Chen等人，2025）采用混合生成对抗网络-卷积神经网络（GAN-CNN）架构（Chen B等人，2021a），整合了生成对抗网络（Generative Adversarial Network, GAN，Goodfellow等人，2014）模块与极坐标卷积神经网络（Convolutional Neural Network, CNN，Krizhevsky等人，2012；Chen B等人，2021b）回归器。DSAT的生成对抗网络模块以全时段可用的红外（infrared, IR）与水汽（water vapor, WV）影像为输入，生成合成可见光（VIS）影像（VIS*），从而可间接利用原始可见光影像训练卷积神经网络回归器。另一方面，该卷积神经网络回归器采用"极坐标卷积滤波器"设计（Chen B等人，2021b），可从卫星影像中提取关键特征，并将这些特征映射为经方位角平均后的标注风场廓线。Chen, B. F., Chen, B., Hsiao, C. M., Teng, H. F., Lee, C. S., & Kuo, H. C. (2025). 《构建深度学习数据集以揭示热带气旋强度与结构极端性的气候趋势》. 《地球系统人工智能》（修回后录用）。

搜集汇总

背景与挑战

背景概述

该数据集通过深度学习技术重建了1981-2020年全球热带气旋的风剖面数据，用于研究热带气旋的强度、结构和动能。数据集采用混合GAN-CNN架构，结合卫星图像和数值模型再分析数据，提供了一个客观且全面的热带气旋分析工具。