Research On Cosmic Ray Identification Algorithm Based On Deep Learnin

The China Space Station Telescope(CSST)is a major scientific project planned for China’s manned space flight project,and is planned to complete a survey mission of 17,500 square degrees of multicolour imaging observation and 400 square degrees of multicolour imaging deep field observation within 10 years after its completion.However,as the space telescope is not protected by the Earth’s atmosphere during the survey,it suffers from severe cosmic ray interference,which has a great impact on the quality of CCD astronomical images and subsequent scientific analysis.The existing cosmic ray identification methods rely too much on manually adjusting parameters and cannot accurately identify cosmic rays in different images,and there are shortcomings in operational speed.Therefore,researching an effective cosmic ray identification method is an important part of the CSST scientific data processing.In this paper,we first systematically investigate and analyze the characteristics of cosmic rays on CCD astronomical images.Since the CSST is currently under construction,we select real observation data from the Hubble Space Telescope(HST)to carry out a deep learning-based cosmic ray identification study in the research process,and the main work includes the following aspects:(1)We select calibrated scientific images from HST to build the dataset and test them with Astro Drizzle data processing pipeline to obtain more reliable cosmic ray annotations,and finally construct a total of 10710 256×256 size image datasets(more than 500,000 cosmic rays)for model training and evaluation.The training set consists of 8190 extragalactic field images and globular cluster images(over 400,000 cosmic rays),and the test set consists of extragalactic field test set(630 extragalactic field images,over 60,000 cosmic rays)and globular cluster test set(1890 globular cluster images,over 60,000 cosmic rays).(2)A convolutional neural network model CRUSNET based on the codingdecoding structure is proposed.For the problem that cosmic rays are easy to lose the shallow features in the downsampling process due to the small pixel scale,a feature extraction network based on the U-shaped structure combined with the residual connection is designed,and a feature fusion network is constructed to retain the shape and size information of cosmic rays.The experimental results show that the model achieves a recall rate of 91.3% and an F2 score of 91.6% on the extragalactic field test set,which is 4.7% and 3.2% higher than the mainstream deep CR method,and a recall rate of 91.8% and an F2 score of 92.3% on the globular cluster test set,which is 2.7%and 1.9% higher than the mainstream deep CR method.(3)We quantitatively assess the usability and stability of CRUSNET using deep learning evaluation metrics and photometric accuracy,respectively.The test results on20 sets of observation data from HST between 2010 and 2019 show that the stability of the model is good and can be applied for a long time after modeling;meanwhile,the relationship between cosmic ray density and photometric accuracy in images with a field-of-view size of 202 × 101 square arcseconds from more than 300 sources is analyzed through cosmic ray density simulation,and the results show that only about6% of the stars have abnormal photometric results when the cosmic ray density is 2%(i.e.,the image exposure time is short).Overall,the deep learning model proposed in this paper can meet the cosmic ray identification requirements of HST short exposure observation data.At the same time,the paper provides quantitative assessment of the model’s availability and stability,and gives the effect of the model recognition effect on the photometric accuracy of astronomical sources under different cosmic ray densities,which has important reference value for future CSST scientific research.