Digital Image Processing Of Meteorological Satellite Sensing Images And Severe Cloud Cluster Recognition

With the rapid development of the digital image processing technology, this makes it as an inevitable trend to dig out the information of satellite cloud images (SCI) and water vapor images (WVI) from the deep graduation and multi-angle. And the routine application of meteorological satellite data is undergoing a profound revolution. It is drastically changed that those data are analyzed qualitatively. By the means of the interdisciplinary cross-studies, such as digital image processing, data fusion, pattern recognition and satellite meteorology, et al, a novel research clue is presented in this thesis. The main research of this thesis is as below:By the means of fusion of SCI and WVI, an effective solution is proposed to deficiency of WVIs application. And a non-uniform fusion scheme is proposed into the fusion process of SCI and WVI in this present work. This scheme means that the weighted proportions are inconsistent between the discomposed approximation image and detail image according to their importance of original SCI and WVI. This is a meaningful improvement which gives prominence to the important features of original images.There are two detecting methods to perform the edge detecting of FCIs. The one is detecting operator method. The other is a method based on wavelet package decomposition. It is concluded that the compositive results of horizontal detail images and vertical detail images excel the approximation image and diagonal detail images in detecting effects.Three methods, i. e., grayscale statistics, correlation-self function, histogram statistics based on neighourhood of each pixel, are applied in the texture features detecting in the present work. The probability of adopting information entropies to describe the denseness-scale, an important notion in the analyses of SCIs, is firstly discussed in the histogram statistics method. And the situation of qualitative analysis of denseness-scale for a long time is broken in this thesis. Furthermore, a notion, adverse-direction correlation, is proposed in the correlation-self function method. This makes sure that the informations on FCIs are utilized suficiently. It is concluded that information entropies and correlation-self function have excellent detecting effects to the texture features within the cloud clusters. The probability of FCIs, edge detecting images and texture detecting images which are applied into routine work is discussed in this thesis. A new situation of meteorological satellite data application is started from a brand-new angle of view in the present work.Support Vector Machine algorithms based on soft margin on the class1 and class2 are introduced to the recognition researches in the present work. Twenty-four experimental schemes are designed to check the convergence of Support Vector Machines according to different parameter C, different iteration stopping condition and whether the linear restriction condition is introduced into the iteration process. Three sorts of sample sets, i. e., thermal convection cloud cluster set, systematic convection cloud cluster set and their mixture set, are acquired and trained to check the convergence of those schemes. It is concluded that the results of experiments are accorded with the actual situation of the recognition of cloud properties on SCIs. Furthermore, except for four schemes based on the soft margin on class2 whiling C is 100 are not convergent, the others can all tend to convengence. Especially, an improvement, which the linear restriction condition is introduced periodly into the iteration process, has been proposed in the itcrative algorithm in this thesis. The itcrativc cycles can be reduced to a large extent, from 11% to 35%.To sum up, the researches in the present work have contributed to promote the automatic applications of meteorology satellite data to a new height.