Abstract: Global observation of atmospheric wind fields are potentially the most impotent data in the analysis for numerical prediction. This paper describes the computer image recognition technique of deriving cloud motion winds (CMW) from the IR images of GMS. The first step in the processing is objects classification analysis in an image segment corresponding to an area of 32×32 IR pixels. In this paper, we use direct classification and K-means algorithm to distinguish the high cloud, meddle cloud, low cloud and earth's surface. At the second step we adopt IR cross-correlation technique in the calculation of CMW, the correlation values between the brightness of the template window and the search window are calculated in the search area of the later time image, a cross-correlation coefficient surface is obtained. For the determination of CMW height, the average weight value of IR brightness temperature of object cloud is compared with the climatic statistic vertical temperature profile in this season, and CMW height, i.e. high cloud, meddle cloud or low cloud, is determined.

Key words: classification, IR cross-correlation coefficient, cloud motion wind

摘要： GMS静止气象卫星探测所得到的红外云图资料，用计算机图象识别技术来反演云迹风。将整幅云图划分为若干个象素点数为32×32的小块，对每个小块进行目标物的分类运算。应用模式识别中区域聚类法即最近邻简单试探法和K-均值聚类算法来完成高云、中云、低云和地表的区分。云迹风的计算采用红外亮温交叉相关法，在设定的搜索区内通过计算相邻两个时次云图目标区与搜索区的红外亮温交叉相关系数，可以得到一个交叉相关系数匹配面。对匹配面上的主极大峰值和次极大峰值应用匹配面检测和连续性检测选出合适的峰值，进而得到云迹风矢量。云迹风的高度推算用选定目标云的红外亮温加权平均值与该季节气候统计温度垂直分布值相比较，推算出目标云为高云、中云或低云。

关键词: 聚类, 红外亮温交叉相关系数, 云迹风