Dvorak 等人提出的基于卫星云图的云型和云系特征的热带气旋强度估计方法已被世界气象组织推荐给全球使用。本文尝试从历史数据中自动发现典型的云型模式,实现Dvorak 模板图像的自动选取、匹配和识别。采用SOM 对12000 多幅红外云图进行聚类,采用提出的局部统计信息等特征,分析了某些距离作为相似性度量存在而只能发现球形簇的缺点;对不同的特征和相似性度量进行了对比试验,并分析了SOM 拓扑误差和量化误差。从实验结果可以看出,局部精特征有着最小的量化误差,但聚类准确度较低。原始图像作为输入特征时,有着较高的聚类准确度及拓扑保持度。局部统计信息特征比局部情特征量化误差大,但有着更高的聚类准确度。这些结论为采用无监督聚类方法来发现云型模式并找到最佳的特征和较好的相似性度量以取得更好的结果提供了重要的参考,也有助于避免目前云图自动化分析研究中对特征和度量选取的随意性。

The tropical cyclone intensity estimation method based on satellite cloud pattern and cloud system features proposed by Dvorak et al has been popularized globally by the World Meterological Organiztion(WMO).An attmept at discovering typical cloud patterns in the historical cloud images automatically is made and the automatically selection, and matching and recognization of Dvorak template imagesare realized. More than 12000 historical satellite infrared images are clustered by using SOM. The shortcoming that Euclidean distance and other similar meaSures can only be used to detect point clusters is analyzed. A comparative experiment is made on different features and similarity measures and the quantization and topology errors of SOl\!I results are given. The experimental result shows that the local entropy feature has the minimal quantization error, but its clustering accuracy is lower. When the original images are used as input features, higher clustering accuracy and topology retention can be obtained. The local statistical information has a quantization error greater than the local optimal features, but its accuracy is higher. These conclusions are of significance to the discovering of cloud patterns, optimal features and better similarity measures by using unsupervised clustering and are also helpful to the avoidance of the subjectivity in feature and similarity selection for current automatical cloud image analysis.