2026-10-12T13:50:36Z https://www.scidb.cn/oai

10.57760/sciencedb.07177 2023-02-03T19:59:30Z

2023-02-03 Galaxy, star, quasar dataset doi:10.57760/sciencedb.07177 en The data used in this paper is from the 16th issue of SDSS. SDSS-DR16 contains a total of 930,268 photometric images, with 1.2 billion observation sources and tens of millions of spectra. The data obtained in this paper is downloaded from the official website of SDSS. Specifically, the data is obtained through the SkyServerAPI structure by using SQL query statements in the subwebsite CasJobs. As the current SDSS photometric table PhotoObj can only classify all observed sources as point sources and surface sources, the target sources can be better classified as galaxies, stars and quasars through spectra. Therefore, we obtain calibrated sources in CasJobs by crossing SpecPhoto with the PhotoObj star list, and obtain target position information (right ascension and declination). Calibrated sources can tell them apart precisely and quickly. Each calibrated source is labeled with the parameter "Class" as "galaxy", "star", or "quasar". In this paper, observation day area 3462, 3478, 3530 and other 4 areas in SDSS-DR16 are selected as experimental data, because a large number of sources can be obtained in these areas to provide rich sample data for the experiment. For example, there are 9891 sources in the 3462-day area, including 2790 galactic sources, 2378 stellar sources and 4723 quasar sources. There are 3862 sources in the 3478 day area, including 1759 galactic sources, 577 stellar sources and 1526 quasar sources. FITS files are a commonly used data format in the astronomical community. By cross-matching the star list and FITS files in the local celestial region, we obtained images of 5 bands of u, g, r, i and z of 12499 galaxy sources, 16914 quasar sources and 16908 star sources as training and testing data.1.1 Image SynthesisSDSS photometric data includes photometric images of five bands u, g, r, i and z, and these photometric image data are respectively packaged in single-band format in FITS files. Images of different bands contain different information. Since the three bands g, r and i contain more feature information and less noise, Astronomical researchers typically use the g, r, and i bands corresponding to the R, G, and B channels of the image to synthesize photometric images. Generally, different bands cannot be directly synthesized. If three bands are directly synthesized, the image of different bands may not be aligned. Therefore, this paper adopts the RGB multi-band image synthesis software written by He Zhendong et al. to synthesize images in g, r and i bands. This method effectively avoids the problem that images in different bands cannot be aligned. The pixel of each photometry image in this paper is 2048×1489.1.2 Data tailoringThis paper first clipped the target image, image clipping can use image segmentation tools to solve this problem, this paper uses Python to achieve this process. In the process of clipping, we convert the right ascension and declination of the source in the star list into pixel coordinates on the photometric image through the coordinate conversion formula, and determine the specific position of the source through the pixel coordinates. The coordinates are regarded as the center point and clipping is carried out in the form of a rectangular box. We found that the input image size affects the experimental results. Therefore, according to the target size of the source, we selected three different cutting sizes, 40×40, 60×60 and 80×80 respectively. Through experiment and analysis, we find that convolutional neural network has better learning ability and higher accuracy for data with small image size. In the end, we chose to divide the surface source galaxies, point source quasars, and stars into 40×40 sizes.1.3 Division of training and test dataIn order to make the algorithm have more accurate recognition performance, we need enough image samples. The selection of training set, verification set and test set is an important factor affecting the final recognition accuracy. In this paper, the training set, verification set and test set are set according to the ratio of 8:1:1. The purpose of verification set is used to revise the algorithm, and the purpose of test set is used to evaluate the generalization ability of the final algorithm. Table 1 shows the specific data partitioning information. The total sample size is 34,000 source images, including 11543 galaxy sources, 11967 star sources, and 10490 quasar sources.1.4 Data preprocessingIn this experiment, the training set and test set can be used as the training and test input of the algorithm after data preprocessing. The data quantity and quality largely determine the recognition performance of the algorithm. The pre-processing of the training set and the test set are different. In the training set, we first perform vertical flip, horizontal flip and scale on the cropped image to enrich the data samples and enhance the generalization ability of the algorithm. Since the features in the celestial object source have the flip invariability, the labels of galaxies, stars and quasars will not change after rotation. In the test set, our preprocessing process is relatively simple compared with the training set. We carry out simple scaling processing on the input image and test input the obtained image. Galaxy; star; quasar Li Xin PUBLIC https://creativecommons.org/publicdomain/zero/1.0/ dataset Science Data Bank