摘要
采用模糊逻辑理论与神经网络技术方法,构建一种新的分类识别方法———模糊神经网络(FNN),运用10.5~12.5μm通道的红外卫星云图资料,对(70°N^70°S,70°E^150°W)范围不同的云类进行定量自动模式识别高云、中云、低云区和无云区以及相关的云量,并结合数字图像处理的相关知识和技术,将分类结果用直观的图像输出。
A new technical method of Fuzzy Neural Network （FNN） based on fuzzy logic and neural network is designed, which is used to construct a classified recognition of automatic pattern in cloud cover and height, using HIRS （High Resolution Infrared Radiation Sounder） data from 10. 5 - 12. 5/ma in the area （70°N- 70°S, 70°E- 150°W）. It automatically recognizes the interrelated cloud cover and sort in high-level cloud, mid-level cloud, low-level cloud and no cloud areas. The distribution characteristics of membership function and the influence of geographical distribution on cloud are discussed, and it is pointed out that： the height of clouds is influenced by the geographical distribution, at the same time there are different influence degrees of underlying surface such as the plateau, ocean, land, forest, desert, etc. which cause incompletely same standard height of the same kind of cloud in different geographical regions. On the other hand, it should be a fuzzy areas in the compartition of height between cloud layers, but it is almost divided with the measure of threshold in the past research [it includes ISCCP （International Satellite Cloud Climate Project）] and will bring the error margin on the method, and thus influences the application of quantificational identifying classes of cloud in the weather forecast. Therefore, according to the weather tradition, it is divided into four kinds of cloud, i. e. , the clear sky or no cloud, low-level cloud, mid-level cloud and high-level cloud, and directly using HIRS data and making sure respectively each parameter of membership function [-a;, bi, ci], i = 1,2, 3,……, No, the distributive curve of membership function is obtained. But the classification of cloud height is still influenced by the geographical distribution and the change of latitude. In order to correct the influence, it is assumed that： （1） The distribution of membership function is influenced by the change of geography from south to north （here the change in longitude is neglected）, and its parameter value is written as：a（y）, b（y）, c（y）, where y expresses the change of latitude and corrects the influence of the variation of geographical areas in the low-level cloud, mid-level cloud and high-level cloud, so the result matches perfectly with observation. In order to operate conveniently, a（y） =λ（y）×a is taken, where λ（y） is called the geographical influence factor of membership function, correcting the parameter value of fuzzy areas in the transition zone in the calculation. In the same way, b（y） = λ（y）×b and c（y） =λ（y）×c. （2） The demarcation between cloud layers is not actually one dot, but is a zonal area. The parameters of different cloud patterns in “no cloud or clear sky”, “low-level cloud”, “mid-level cloud”, “high-level cloud” are taken as [0, 52, 801, [64, 112, 1561, [136, 176, 2081, and [186, 224, 2561. Among them, [64, 80], [136, 156], and [186, 208] are fuzzy areas in the transition zone between cloud layers. The different influence of each cloud pattern is corrected using the geographical influence factor λ in the geographical distribution, then the FNN is built with the geographical influence factor λ, and its results and methods are analyzed and contrasted with ISCCP. A classified result of visual display image is inputted with digital image techniques. As a result, it makes known that the technical method has high intelligent discernment, good effect and calculative speed, it overcomes the influence of the artificial factor about the method of ISCCP.
出处
《大气科学》
CSCD
北大核心
2005年第5期837-844,共8页
Chinese Journal of Atmospheric Sciences
关键词
模糊神经网络
隶属度函数
卫星云图
fuzzy neural network, satellite images, membership function
