【作者】 王继光；

【导师】 郁文贤；

【作者基本信息】 国防科学技术大学 ， 信息与通信工程， 2007， 博士

【摘要】 军事气象保障的复杂性促进了先进、实用的预报理论和保障技术的涌现。卫星云图蕴含着丰富的气象信息。客观、准确、实时地识别卫星云图中的云类特征、开展云团短时活动预测是军事气象保障中急待解决的重点和难点问题。本文针对云分类和云团短时移动的特殊性和复杂性,基于学科和方法交叉的思想,研究探索了卫星云图云类识别和云团短时预测的研究方法和技术途径。本文研究内容大致分两部分。第一部分为多光谱卫星云图的云分类判别研究,主要内容包括:(1)资料收集整理、误差订正和样本采集。对收集的GMS-5静止气象卫星的红外1、红外2、可见光和水汽四个通道的云图资料进行了灰度与分辨率的标准化转换和太阳高度角订正;从中采集多种典型云类和陆地、水体的若干样本,建立了云类样本库,为随后的云分类研究提供建模和实验的数据信息。(2)季节内的云分类判别。季节内云的特征识别与分类是日常气象保障的基本内容,针对当前云分类研究中广泛采用的特征空间映射分类法调控性差、云类样本要求高、噪声误差需要人工判断消除等缺点以及常规人工神经网络分类模型对云图梯度和纹理特征描述不足等问题,通过计算云图样本灰度-梯度共生矩阵及其若干特征统计量,用BP神经网络建立了云分类模型。该模型可综合利用云图样本的多种特征信息(如灰度、梯度和纹理),在特征量输入与分类输出间建立起良好的非线性映射关系;通过在网络训练中引入模糊化处理方法,减小了网络复杂程度、加速了网络收敛,分类结果更为准确、合理。所建BP网络分类器模型能够对云-地之间、云-水之间和季节内的多种典型云类进行有效的分类判别。(3)年际内的云分类判别。在实际业务应用中,需要考虑全年各个季节的云分类问题,即年际尺度和季节间的云分类判别。由于云生成和维持的复杂性和多样性,不同季节的云类主体的形态特征存在较大差异,因此笼统采用全年资料序列建模的模型训练难度大、误差收敛慢、泛化性能差,尤其对季节转换期间的云分类判别效果更差。针对上述问题,本文基于自组织特征映射网络(SOMF)和概率神经网络(PNN)各自优势,提出SOMF网络与PNN网络结合的二级分离云分类方法。该方法将云分类过程分为两个阶段:首先用自组织特征映射网络对样本特征库进行无监督聚类(一级分类);其后再针对分离出的各类样本进行有监督训练和目标修正,分别建立其各自的概率神经网络的云分类模型(二级分类)。上述处理方法不是机械地按照季节划分,而是遵循云类样本特征空间结构的自然相似性进行客观合理的分类建模,既考虑了不同季节云类主体的差异,又简化了云分类过程复杂性,使得季节过渡和转换期间的云分类模型建立和选择有了客观依据,云分类准确率得到了有效提高。(4)云类过渡区的多属性判别。针对各云类过渡区间存在的多属性(同时具有多种云类特征)问题和云类样本中的误差问题,基于云类样本的红外-可见光二维灰度特征空间,提出了用模糊C均值聚类(FCM)方法调整优化云类样本特征映射集的技术途径。该方法以隶属度为度量,较为合理地描述了云图空间点(尤其是云类过渡区空间点)的多云类属性和云类特征的“模糊”性,其“柔性”分类结果更贴近客观实际和人们的视觉理解,同时既有效地降低了采样误差,又保持了云类样本的基本结构特征。针对常规FCM方法对聚类初值的敏感性,提出用样本特征均值替代常规FCM方法中随机的初始中心的改进办法,改善了常规模糊C均值聚类方法难以检测不规则云结构特征子集的不足。(5)非典型云类和多层云系的分类判别。实际天气中,除了特征较为明显的“典型”云类外,还存在大量特征较为“模糊”的云系(如天气系统生消演变期间和多种天气系统相互共存等情况下的云类)。针对实际天气中上述非典型云类和多层云系的“模糊性”分类问题,以及常规模糊聚类方法对初始聚类中心选取的随机性和聚类结果的“局部优化”等问题,本文提出遗传算法(GA)全局寻优与模糊C均值聚类(FCM)局部寻优以及模糊减法聚类(FSC)客观估算聚类数等优势互补的研究思想和技术途径。首先用FSC方法客观确定云图中的云类数,其后基于估算的云类数,采用GA进行全局搜索,确定初始聚类中心,最后用FCM方法对GA提供的聚类结果再进行局部的调整优化。试验结果表明,综合云分类方法(FSC-GA-FCM)的分类效果明显优于单一的FCM或GA算法,表现出较好的创新特色和技术优势,并被有效应用于相关科研项目的云分类算法设计之中。第二部分为卫星云图中云团移动的短时预测研究,主要内容包括:(1)确定了基于局部小波矩方法的云团特征匹配和移动矢量计算方案。首先系统阐述了云团移动矢量计算的基本思想,重点引入和分析了图像处理中四种重要的匹配方法:傅立叶位相法、交叉相关系数法、局部信息熵方法和局部小波矩方法在云团特征匹配中的应用,并针对云团移动过程中强度变化和旋转等非平稳运动特点,设计了上述四种云团匹配计算方案,通过仿真模拟,分析、评估了各种算法的优缺点,发现局部小波矩方法能够更为合理恰当地表现云团环境的变化和特征匹配,进而为云团预测的运动矢量计算提供了可靠的计算方案。此外,还提出了移动矢量平滑处理和质量控制的改进方法。(2)建立了基于云团移动矢量线性外推的云团短时预测模型。基于上述优选的云团运动矢量算法和Cressman多次连续平滑方法质量控制,结合后向轨迹方法,设计了云图1～3小时的短时预测模型。试验结果表明,该方法的云图1-3小时的预测结果可有效改善主观目测卫星云图判断云团移动的偏差,并可在一定程度上表现云的旋转等非平稳运动特性,进而为云团移动预测提供了一个客观、定量的技术手段。(3)提出了经验正交函数分解与动力系统重构相结合的云图非线性预测思想。基于经验正交函数(EOF)方法对卫星云图的样本序列进行时、空分解;在EOF分解的空间场相对稳定的基础上,引入遗传算法对EOF分解的时间系数序列进行动力系统重构和模型参数反演,建立了云团演变的非线性动力预报模型。通过对EOF时间系数预测结果与空间结构模态的重构合成,实现了云图非平稳运动的中、长时效预测,为云图预测提供了一种创新思想方法和实用技术途径。更多还原

【Abstract】 To compete with hardness of military meteorological service , it is forced to produce lots of advanced theories and practical technologies which can serve the purpose of weather forecast. A satellite cloud image contains a lot of weather information. It is hard and important task to distinguish the characteristics and short-time movement forecast of satellite cloud image timely、correctly、objectively and automatically. To meet the challenge of particularity and complexity of cloud classification and short-time movement forecast of satellite cloud image, it is studied to improve used theory and find a new arithmetic which can be easily put into use based on idea of crossing of subjects and methods in this paper.The paper contains two parts. In part one, it is mainly aimed at distinguishing and classification of satellite cloud image. There are five subjects:(1) After standard conversion of gray and recognizing ratio and as well as rectifying to the height angle of the Sun from satellite cloud image of GMS-5 through four bands of IRK IR2、VS and WV , it is built of a sample base which collects numbers of samples of land and water and seven types of typical clouds as well, which is used for study of cloud classification;(2) Distinguishing and classification of cloud characteristics inner season is basic part of ordinary weather services. On the basis of above works, the samples of cloud classification are reflected to two dimension gray space of IR and VS by the means of Characteristic Space Projection (i.e. CSP). It is found the criteria of classification of cloud pixel with determination of falling area of the samples. The test results show that the method of CSP is simple and easy, especially suited for uneven division of unregulated clustering data, the result is better than the regular criterion method .(3) Due to complexity and variety of clouds presentation, it is necessary to distinguish and classify clouds’ characteristics within a year or inter-seasons. On the basis of each advantage of SOMF (i.e. Self-Organized Mapping Reflection) and PNN (i.e .Probability Neural Network), an improving method of two times classification of clouds is introduced in combination of SOMF and PNN.. The method makes process of cloud classification into two stages: first, it is clustered to the base of samples characteristics without supervision by the mean of SOMF, then, models of clouds classification with their own PNN are build after kinds of classified samples have been trained under supervision and objects’ rectification. The method can both consider differentiae of clouds in different seasons and simplify process of cloud classification .Therefore, it can effectively reduce the sampling errors, its classification results conform to reality .(4) It is hard to distinguish attribute of clouds in transitional area. On the basis of IR-VS two dimension gray space projection of the samples of cloud classification , the characteristic area of the samples is rectified and optimized by the means of Fuzzy C Mean clustering method(i.e. FCM). An improving method that characteristic mean values of the samples replace with random initial centre values in FCM can both avoid defection of that FCM is too sensitive to the initial values and correct the clustering results’ distortion to characteristic structure of the samples. Therefore, it can effectively reduce the sampling error and keep basic characteristic structure of cloud samples .(5) In actual atmosphere, there are some of "fuzzy" clouds which are hard to tell what kinds of clouds are. FCM is an advantage non-supervised clustering arithmetic. It can fairly good realize non-linear distinguishing of both high dimension complicated data and non-determinative cloud patterns with calculating and comparing what the subordination is each pixel of cloud image apart from every clustering center. Nevertheless, the normal FCM has three native defections : less capability of global optimization; clustering results dependent on the initial clustering center which produces in random way; the clustering numbers must be given manually. To counter the defections of FCM, a new comprehensive method that Genetic Arithmetic is used for global optimization and FCM is used for local optimization and FSC (Fuzzy Subtract Clustering) is used for objectively estimating clustering numbers as well is introduced on distinguishing satellite cloud patterns. The test results show that the comprehensive method of cloud classification is obvious advantage over any one of three methods and effectively remedy the defection of FCM and GA on cloud classification, and can be used to practice.In part two, it is researched on short time forecasting of cloud motion of satellite cloud image. The major content includes:(1) An arithmetic of cloud motion vectors—local wavelet quadrature is selected to objectively and effectively reflect characteristics of cloud motion . Firstly, the basic thought of computing the cloud motion vectors was given and the four important match methods of Fourier Phase, Cross Correlation, Local Entropy were introduced, analyzed . Then, the four schemes are designed according to the unsteady property of could motion during process of heavily changes and rotations . By evaluation of advantage and disadvantage of each scheme, and comparison and simulation tests, it is found that the local wavelet quadrature is more reasonable to reflect variation and characteristics of clouds . Moreover, an improved method of quality control and cloud motion vectors using for corrections scheme is presented.(2) A model of the short-time forecast is established based on linear extrapolation of cloud motion vectors. Then, quality control is implemented to computed cloud motion vectors with multi-smoothing successive corrections scheme (i.e. Cressman). After smoothing process, 1-3 hours forecast of cloud motion is tested with Backward Trajectory method. The test results show that the method can effectively decrease deviation of distinguishing cloud motion by range estimation and present unsteady characteristics of could rotating to some extent, and as well as provide forecasting of cloud motion in a more objective and quantity way.(3) A nonlinear idea of cloud prediction combined EOF decomposition with dynamical system restructure was brought up. Series of the samples of the satellite cloud image were made space-time decomposition by EOF. On basis of relative stability of space in EOF decomposition, Genetic algorithms were introduced to go on dynamical system restructure and parametric inversion of model by EOF time factors, and a nonlinear dynamical model forecasting cloud evolution was established. By means of prediction of EOF time factors mingled with restructure of space-structure modalities, a middle and long-time prediction on unsteady characteristics of cloud motion has been realized. An innovative idea and useful way was created for cloud motion prediction.更多还原