Research On High Redshift Estimation Based On Deep Learning And Massive Photometric Data

The estimation of galaxies redshift has always been a difficult and hot topic in astronomy.At present,the development of various advanced sky survey observation systems have promoted a large amount of photometric data,which also lays a foundation for deep learning method to deal with astronomical data.Firstly,the galaxies photometric database of SDSS and TWOMASS is selected as the data source.Then according to the characteristics of the corresponding photometric data,the galaxies samples are clearly classified and preprocessed.Particle Swarm Optimization(PSO)is used to optimize the Error Back Propagation Neural Network(BPNN),Convolutional Neural Network(CNN)and the Generalized Regression Neural Network(GRNN)based on Fruit Fly Optimization Algorithm(FOA)are used to predict the photometric redshift and analyze the results.The main works and innovations of this thesis are as follows:(1)Data preprocessing and classification operations.Firstly,the characteristics of photometric data in SDSS and TWOMASS systems are described,and the homologous cross matching is carried out.According to the characteristics of the photometric data of different systems,this thesis uses the extinction and threshold method to filter the data.Then,the selforganizing neural network is used to classify the galaxies in a single SDSS observation system,and the early type galaxies with relatively large proportion are selected for subsequent redshift estimation to fully reduce the complexity of the samples.And then,the data after homologous cross matching are divided into low redshift galaxies and high redshift galaxies based on DNN classifier,so as to analyze the photometric redshift of high redshift galaxies more accurately.(2)PSO-BPNN algorithm is proposed to improve the high photometric redshift estimation in SDSS.Firstly,the BP neural network is used to prediction of the galaxies photometric redshift from SDSS-DR13,and the results are analyzed.Then,the PSO algorithm is used to optimize the parameters of the BPNN.Finally,the results of the two algorithms are compared by error analysis and visualized.Experimental results show that PSO-BPNN has achieved a high estimation accuracy for early type galaxies,which is superior to the estimation results of Genetic Algorithm(GA)optimized BPNN.(3)Improved CNN algorithm estimates the photometric redshift for the homologous cross data of SDSS and TWOMASS.By improving the activation function,the seven-layer CNN model is designed to estimate the photometric redshift based on the Tensorflow framework.Then the model is applied to the redshift estimation of low-redshift galaxies and high-redshift galaxies.The experimental results confirm that the CNN algorithm can effectively estimate the photometric redshift,and the estimation accuracy and the effective coverage of redshift are higher.(4)FOA-GRNN algorithm is proposed to estimate the high photometric redshift.The spread of key parameters in GRNN was determined by using FOA algorithm,and the highredshift estimation based on FOA-GRNN was realized according to the idea of nonlinear regression analysis of GRNN.By comparing and analyzing the estimation results of FOAGRNN with those obtained in current astronomical studies,the results show that the estimation based on FOA-GRNN is the best.It is a new attempt to apply FOA-GRNN to the estimation of photometric redshift,which open up new ideas for the estimation algorithm of galaxies photometric redshift.