| 基于SOM聚类的台风云型模式的发现 |
| 投稿时间：2009-06-25 修订日期：2009-07-12 点此下载全文 |
| 引用本文： |
| 摘要点击次数: 4087 |
| 全文下载次数: 144 |
| 作者 | 单位 | 地址 | | 魏坤* | 上海交通大学 | 上海市东川路800号上海交通大学闵行校区A0603291班 |
|
|
| 中文摘要:Dvorak等提出了基于卫星云图的云型和云系特征实现热带气旋强度估计方法，并被世界气象组织推荐全球使用。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。对12000多幅红外云图采用SOM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的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，并给出了SOM拓扑误差和量化误差的分析。从实验结果可以看出，局部熵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熵特征量化误差大，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进行云型模式发现时，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，也有助于避免目前已有的云图自动化分析研究中的特征和度量的选取的随意性。 |
| 中文关键词:聚类分析 台风云型模式 知识发现 自组织网络 局部统计信息 |
| |
| Typhoon Cloud Pattern Discovery by SOM Clustering |
|
|
| Abstract:The Dvorak technique which estimates tropical cyclone intensities using satellite cloud pattern and cloud system features, has been popularized globally by World Meteorological Organization (WMO). This paper tries discovering prototypical cloud patterns in a mass of history cloud images automatically and selecting, matching and recognizing automaticly the Dvorak template images. SOM clustering is implemented on more than 12000 history satellite infrared images, with the proposed feature of local statistical information and so on being used. We analyze the shortcoming that Euclidean distance and other similarity measures can only detect point sets of datasets while clustering. Contrast experiments of clustering using various features and similarity measures are carried out, and then the quantization and topology errors of the SOM results are given. From the experimental result, we can see that: local entropy feature has the minimal quantization error, but its clustering accuracy is lower; higher clustering accuracy and topology retention are obtained when using original images as input features. The quantization error of local statistical information is greater than that of local entropy feature, but its accuracy is higher. These conclusions are meaningful for searching optimal features and similarity measures to improve the performance of typhoon cloud pattern discovery using unsupervised clustering, and for avoiding the subjectivity of feature and similarity measure selection in existing researches of satellite image automatic analysis. |
| keywords:Clustering Analysis Typhoon Cloud Patterns Knowledge Discovery SOM Local Statistic Information |
| HTML 查看/发表评论 下载PDF阅读器 |