Galaxy Morphology Classification With Deep Convolutional Neural Networks

Along with the development,of the observation technology and observation instrument,large scale surveys such as the Sloan Digital Sky Survey(SDSS),Cosmic Evolution Survey(COSMOS)have gradually carried out and resulted in the availability of very large collections of images,which show an explosive growth trend.The new observation means make,astronomy enter the era of big data.In the study of Galaxy.massive data makes it impossible to manually classify large-scale galaxies images.Currently,it is a significant and challenging task to mine hidden information from massive and high-dimensional data,even discover new scientific problems,while classifying galaxies automatically,quickly,efficiently and accurately.Galaxy morphology classification based on traditional machine learning,de-pends on complex feature engineering,which needs specific domain knowledge and human ingenuity to extract good data representation.Feature extraction directly determines the classification performance of the final classifier.While the most important value of neural network is the automatic extraction and abstraction of features,which eliminates the complexity of manual feature extraction and can automatially find complicated and effective higher level representations.This paper proposes a modified deep convolutional neural network(CNN)for galaxy morphology classification based on classical CNN models such as AlexNet,VGG.Inception and ResNets.We try to improve residual unit,add a dropout,reduce the depth of the network and wider the width of the network in order to achieve the best generation performance of model.The network is fed with raw galaxy images pixels directly and automatically extracts galaxy morphological features to identification and classification.A sam-ple of 25911 galaxy images from Galaxy Zoo 2 dataset are applied to classify galaxies into five classes and the remaining 2879 galaxy images are testing set to test the classification performance of our model.Through the selection of super-parameter,a 26-layers network is adopted and called ResNet-26.A variety of metrics,such as accuracy,precision,recall,F1 value and AUC,show that ResNet-26 achieves the state-of-the-art classification performance among 5 popular CNNs,namely,Dieleman,AlexNet,VGG-1G.Inception V3 and ResNet-50.The overall classification accuracy of our network on the testing set is 95.2083%,the average F1 value is 0.9515 and average AUC value is 0.9823.Further,this paper digs into the networks to analyse what features are learned and how the learned features can help better understanding the data itself.We ex-tract the activations from the last,fully connected layer or the last convolutional layer or the last,average pooling layer of CNNs to study the high-dimensional feature representations of galaxy images.Then we apply t-Distributed Stochastic Neighbor Embedding(t-SNE),a popular dimensionality reduction tech-nique.to visualize the high-dimensional galaxy feaure repreentations in two-dimensional scatter plots.From the visualization,We try to understand the galaxy images data itself and obtain some highly valuable insights.For instance,we try to discover the implied global and local structure information,explore the internal laws and relations of galaxy high-dimensional and abstract representation.ResNet-2G can be applied to large-scale galaxy classification in forthcoming surveys such as the Large Synoptic Survey Telescope(LSST).And the visual-ization of galaxy morphological high-dimensional representation can be applied to outlier detection,and look for abnormal galaxies in the process of subsequent analysis of galaxy morphological classification.