Automatic Classification And Recognition Research About Cloud On Meteorological Satellite System

The meteorological satellite realized the observation of meteorological elements from the space to the Earth and its atmosphere,and generated satellite cloud images 24 hours a day continuously.The observation range is wide and cover the radiation and distribution information about various clouds.At this stage,meteorologists mainly rely on artificial experience to judge cloud-type information in satellite cloud images,and then analyze possible weather phenomena.This method is subjective and requires high professionalism for meteorologists.At the same time,the amount of cloud images is large,the analyze is inefficient Therefore,the use of computers contribute to perform automatic classification about cloud-type in satellite cloud images efficiently and accurately,this way has practical research significance and is also the trend of meteorological work in the futureThis paper relies on the needs of aeronautical meteorological service to carry out classification research of cloud-type in satellite cloud images.The data set was selected from the satellite images of Himawari-8 and FY-2.In order to avoid the influence of "Homology heterographic homology" and similar color to the classification results,a cloud classification method based on multi-texture features is proposed.The fusion of Gabor transform and gray level co-occurrence matrix algorithm is used to realize.For the cumuliform cloud,stratiform cloud,and cirriform cloud,the accuracy is 93.33%,and the classification accuracy for the five cloud types of cumulus,cumulonimbus,stratiform cloud,cirriform cloud,and clear sky is 69.2%In order to improve the classification accuracy,a method based on deep learning is proposed.Based on the data enhancement operation,the amount of training set was effectively expanded.Then,with the GoogLeNet network,the classification accuracy of three cloud types is improved to 95.67%,and the five types accuracy is improved to 96.8%.However GoogLeNet still did not meet the requirements,and it also has a serious over-fitting phenomenon.Therefore,by reducing the number of inception modules in the GoogLeNet network,adjusting the parameters,simplifying the network layers to modulate the GoogLeNet In the end,with the improved network,the accuracy of the three types can achieve 98%,the classification accuracy of five types can reach 98.4%,which meet the expected accuracy,and alleviate the problems of over-fitting and computational resource consumption in the GoogLeNet cloud classification network effectivelyFinally,to facilitate meteorologist to monitor the disaster weather caused by the cumuliform cloud,the YOLO deep convolutional neural network is used to detect the cumuliform cloud in the satellite cloud image,the result is basically consistent with the artificial discrimination result,and meets the requirements of the aviation meteorological service.