Image Processing Technology Of Multi-channel Data Of Satellite Cloud Image

With the development of digital image processing technology, deeply miningmulti-channel satellite cloud data has become an inevitable trend of development.Meteorological satellite data has brought a profound reformation to the application ofmeteorological routine business and the analytical methods of meteorological services will beexpanded. Based on the study on image processing, satellite meteorology, data fusion, andpattern recognition, this paper puts forward a new research idea of meteorological analysis.This study has the following main aspects:The fusion of infrared cloud image and visable cloud image has sovled the problem ofinsufficient application of visable image and an image fusion method is proposed for fusinginfrared image and visable image. The original image is first decomposed into the sub-bandsof different scales and different directions by nonsubsampled contourlet transform. In theprocess of image reconstruction, using pulse coupled neural network to decide the value ofhigh frequency sub-band according to sum of modified energy of laplacian. The method canobtain better direction information, removing aliasing effect, and has good noise immunity.Application of this method will lay a good foundation for segmenting typhoon andrecognizing different cloud types.By the analysis of traditional level set active contour models, we propose a modifiedgeodesic active region model for typhoon segmentation in the fused satellite cloud image. Astyphoon is ever-changing, one or several shape and texture features can not be used to identifytyphoon. Also whether in development, mature, or extinction period, the typhoon has vortexwith spiral feature. Based on the above characteristics, this paper presents a method to statisticthe angle of spiral by the angle of boundary cloud and center cloud, as the feature ofrecognizing typhoon. Based on the spiral feature, the threshold, morphological andmathematical statistical methods are used to identify the interest region of typhoon. At last,typhoon segmentation is implemented by the proposed model.Three methods which are neighborhood-based autocorrelation function, statisticalgeometrical features and gray level co-occurrence matrix are applied to analysis the texturefeatures of the fused cloud image. Texture detection experiments are carried out for differentsizes neighborhood. According the statistics of the experiment results, effective texturefeatures are extracted for cloud recognition.We put forward a semi-supervised training method of self learning. Linear constraints of iterative process is intrduced in this method for different training samples and stop conditionof iteration and then this method is used to train the support vector machine. One against allmethod based decision tree is proposed for identifying cirrus spissatus, altocumulustranslucidus, cumulus humilis and nimbostratus. Experiment results show that support vectormachine with self learning can improve the recognition rate of cloud classification. Also therecognition time of one against all method based decision tree is much less than traditionalrecognition method of multi-class based on support vector machine. If the category of cloud isincreased, this advantage will become more apparent. For the problem of small samplerecognition, this algorithm performs better than PNN algorithm and k-NN algorithm.The research of this thesis is expected to put the automation applications ofmeteorological satellite data to a new level.