Celestial Object Classification And Parameter Prediction Based On Multimodal Machine Learning

The massive amounts of astronomical observation data pose a challenge for astronomers to classify celestial bodies efficiently and accurately,which can help them conduct further scientific research.Once classified,predicting the important parameters of these celestial bodies becomes a crucial research task.This thesis proposes a novel solution for the classification of stars and quasars,as well as the prediction of redshift,one of the most significant parameters of quasars,based on machine learning methods.Starting with single-modal data,this thesis analyzes and processes astronomical data to achieve automatic classification of stars.Next,multi-modal data is introduced to improve the classification accuracy of stars and quasars,which are two celestial bodies that are easily confused.Finally,a multi-modal machine learning method is employed to enhance the accuracy of quasar photometric redshift prediction.The main contents of this thesis are divided into the following three parts.First of all,in the face of high-dimensional stellar spectral data,this thesis performs normalization and other preprocessing,and constructs a Classification model of Stellar Spectral type and Luminosity type based on Convolution Neural Network(CSSL＿CNN).The experimental results show that,compared with the existing stellar classification models,CSSL＿CNN not only avoids the dependence on the standard stellar library,but also extracts the deep features in the data,and obtains a higher classification accuracy.Secondly,faced with easily confused images of stars and quasars,this thesis proposes a celestial object classification method based on multimodal machine learning.Due to the scarcity of features contained in the image data,the spectral data is introduced to enrich features.The two modalities are fully aligned and correlated to achieve feature fusion,and finally complete the task of classifying stars and quasars together.Compared with the classification by using image data alone,the multimodal classification method that introduces spectral data significantly improves the accuracy.Finally,in view of the scarcity of photometric data features and the difficulty in improving the accuracy of photometric redshift prediction,this thesis introduces the corresponding spectral data as supplementary features,constructs a photometry-spectra feature transformation model,and based on the converted data,proposes a Multimodal Machine Learning method for estimating the Photometric Redshifts of quasars(PhotoRedshift-MML).Experiments show that the PhotoRedshift-MML model improves the prediction accuracy compared with the single-modal quasar photometric redshift prediction method and verifies the effectiveness and generality of multimodal mothods.