Automatic Recognition And Long-term Trend Statistics Of Atmospheric Gravity Waves Based On Machine Learning

The atmospheric gravity wave(AGW)is an important participant in the atmospheric activities.As a carrier of momentum and energy,AGW take the leading role in the atmospheric circulation.And AGW also affects the structure of mesosphere and lower thermosphere region.The propagation of AGW in the airglow causes the wave structure of airglow,which is used as tracer for the all-sky airglow imager(ASAI)to observe the AGW.The continuous observation of ASAI can track the perturbation in the airglow,providing both spatial the temporal information of the evaluation of the AGW.With the support of the Meridian Project,15 ground-based airglow observation stations are established.The ASAIs in the stations record the airglow data every night,accumulating a large number of airglow images over the years.In order to make full use of these data,this paper develops an automatic extraction algorithm to extract the AGW parameters automatically based on machine learning.With the help of the algorithm,the airglow data are statistically analyzed to obtain the long-term trend of AGW.The main research includes the classification of airglow images based on the convolution neural network,the location of wave structures based on object detection algorithm,and the automatic extraction of gravity wave parameters based on Two-Dimensional Fast Fourier Transform.According to the three algorithms,the long-term trends of AGW are summarized based on the observation of Linqu and Lhasa stations,respectively.In addition,an algorithm of the discrete wavelet transform is developed to separate the AGWs of different wavelength in an airglow image containing multi-scale AGW to denoise for better auto extraction and try to find the effect of wavelength on other propagation characters.The whole thesis consists as following:The first chapter briefly reviewed the research on AGW,including the influence of atmospheric gravity wave on the middle and upper atmosphere environment,observations from the space and ground,the progress and problems in the AGW research,the purpose and main content of this thesis.The second chapter introduces the ASAI observation instrument and its data inversion,as well as the airglow observation network of the Meridian Project which provides the airglow data used in this thesis.In the third chapter,the method of machine learning is used to automatically identify the atmospheric gravity wave in the airglow data.The data used for machine learning model training in this paper comes from the observation data of Linqu station in 2013.In the research,a machine learning model based on convolutional neural network is built to distinguish the original data in the clear night sky from the unclear night sky.Then,on the basis of image classification,Faster R-CNN(fast regions with revolutionary neuron network)is built based on the tensorflow object detection API.Through the training of manually labeled sample images,a wave structure recognition and location model(atmospheric gravity wave recognition)is generated,which can effectively identify and locate the wave.Combined with the wavelength correlation of the image sequence,the model recognition error caused by wave like structure cloud or fog is corrected.Finally,a program of automatic recognition of atmospheric gravity wave based on machine learning and traditional data processing is obtained.In the fourth chapter,the data of Linqu station in 2014 and Lhasa station from 2015 to 2017 are analyzed by the automatic extraction program of atmospheric gravity wave parameters developed in the third chapter.The propagation characteristics and long-term variation trend of atmospheric gravity wave in these areas is studied based on the analysis.In addition,by comparing with the artificial statistical data of Linqu station and Qujing station near Lhasa station in other references,the effectiveness of the automatic extraction program of AGW parameters is further confirmed.In Chapter five,discrete wavelet transform is used to reconstruct the atmospheric gravity wave in the airglow image to separate the waves of different scales.Finally,the above work is summarized.