Research On Machine Learning Application In The Photometric Redshift For Massive Data

Redshift of galaxies is an important parameter for the research of the large-scale structure and evolution of the universe.With the continuous development of advanced telescopes,galaxy photometric data has entered an ear with exponential growth.In this background,it has been an inevitable trend to processing massive photometric data by using efficient machine learning algorithms.In this thesis,the theory of photometric redshift and its development status are introduced first,and then the galaxy samples are classified according to their data characteristics.After the data preprocessing,the error back propagation(Error Back Propagation,BP)neural network optimized by genetic algorithm,convolutional neural network and mixed density neural network are used to predict the photometric redshift of galaxies.After that,it has analysis the error of the results of these algorithms.The main results of this thesis are as follows:(1)Data filtering and cluster analysis of galaxy samples.Based on the complex characteristics of the nonlinearity of the massive galaxy metering data,this thesis uses the extinction and threshold method to filter the data.And then,the galaxy samples are divided into early type galaxies and late type galaxies by using the self-organizing neural network.The results show that classification can effectively reduce the complexity of sample data and the error of regression prediction was also reduced.(2)BP neural network model optimized by genetic estimates photometric redshift of galaxies.In this thesis,the BP neural network is used to the prediction of the galaxy photometric redshift.And for defect of the BP neural network which is easy to fall into the defect of the local optimal solution during the gradient descent process,the genetic algorithm is used to optimize the parameters of the BP algorithm.Experimental results show that compared with traditional photometric estimation algorithms such as K-Nearest Neighbor,BP algorithm has great advantages in terms of accuracy and efficiency.The optimized BP neural network can predict the photometric redshift better for both galaxy samples than the latest K-Nearest Neighbor algorithm.(3)Improved convolutional neural network algorithm estimates the photometric redshift of galaxy photometry.In this thesis,the convolutional neural network which is commonly used in image classification was used to estimate the photometric redshift.The Tensorflow framework is used to construct the convolutional neural network model,and the CNN model is properly modified.By using the ReLU activation function instead of the traditional activation function Sigmoid,the output is directly fitted to realize the prediction of the continuous nonlinear photometric redshift.The experimental results confirm that the convolutional neural network has been successfully applied to the prediction of galaxy photometric redshift,and the efficiency has been improved compared with the BP algorithm.(4)Mixed density neural network model estimates photometric redshift of galaxies.In this thesis,the conditional probability distribution of photometric redshift is fitted according to the idea of mixed density neural network model of probability density kernel estimation,so as to realize the non-linear estimation of galaxy photometric redshift.The use of mixed density neural network model for the prediction of photometric redshift is an attempted innovation.The experimental results show that the prediction results are better than the BP neural network model.And it opens up new ideas for the prediction of photometric redshift.