Cloud segmentation is one of the hot tasks in the field of weather forecast, environmental monitoring, site selection for observatory, and other areas. In this letter, we proposed a new deep convolutional neural network architecture called CloudU-Net for daytime and nighttime cloud images' segmentation. The net consists of dilated convolution, activation, batch normalization (BN), max pooling, upsampling, skip connection, and fully connected conditional random field (CRF) layers. The benefits of the net architecture are four aspects: First, the dilated convolution increases the receptive field of the filters to obtain more information of the context without increasing the extra amount of computation and the extra number of parameters. Second, the BN layer increases the speed of network training and prevents over-fitting. Third, the fully connected CRF optimizes the output of the front end of the architecture, and finally gets better segmentation results. Finally, the enhanced optimizer Lookahead improves the learning stability and speeds up model convergence. Compared with the current deep-learning-based state-of-the-art cloud images' segmentation algorithms, the CloudU-Net demonstrates better segmentation performance for daytime and nighttime cloud images.

Most astronomical source classification algorithms based on photometric data struggle to classify sources as quasars, stars, and galaxies reliably. To achieve this goal and build a new Sloan Digital Sky Survey photometric catalogue in the future, we apply a deep learning source detection network built on YOLO v4 object detection framework to detect sources and design a new deep learning classification network named APSCnet (astronomy photometric source classification network) to classify sources. In addition, a photometric background image generation network is applied to generate background images in the process of data sets synthesis. Our detection network obtains a mean average precision score of 88.02 when IOU = 0.5. As for APSCnet, in a magnitude range with 14–25, we achieve a precision of 84.1 ${{\ \rm per\ cent}}$ at 93.2 ${{\ \rm per\ cent}}$ recall for quasars, a precision of 94.5 ${{\ \rm per\ cent}}$ at 84.6 ${{\ \rm per\ cent}}$ recall for stars, and a precision of 95.8 ${{\ \rm per\ cent}}$ at 95.1 ${{\ \rm per\ cent}}$ recall for galaxies; and in a magnitude range with less than 20, we achieve a precision of 96.6 ${{\ \rm per\ cent}}$ at 94.7${{\ \rm per\ cent}}$ recall for quasars, a precision of 95.7${{\ \rm per\ cent}}$ at 97.4${{\ \rm per\ cent}}$ recall for stars, and a precision of 98.9 ${{\ \rm per\ cent}}$ at 99.2 ${{\ \rm per\ cent}}$ recall for galaxies. We have proved the superiority of our algorithm in the classification of astronomical sources through comparative experiments between multiple sets of methods. In addition, we also analysed the impact of point spread function on the classification results. These technologies may be applied to data mining of the next generation sky surveys, such as LSST, WFIRST, and CSST etc.

In recent years, convolutional neural networks (CNNs) have achieved competitive performance in the field of ground-based cloud image (GCI) classification. Proposed CNN-based methods can fully extract the local features of images. However, due to the locality of the convolution operation, they cannot well establish the long-range dependencies between the images, and thus they cannot extract the global features of images. Transformer has been applied to computer vision with great success due to its powerful global modeling capability. Inspired by it, we propose a Transformer-based GCI classification method that combines the advantages of the CNN and Transformer models. Firstly, the CNN model acts as a low-level feature extraction tool to generate local feature sequences of images. Then, the Transformer model is used to learn the global features of the images by efficiently extracting the long-range dependencies between the sequences. Finally, a linear classifier is used for GCI classification. In addition, we introduce a center loss function to address the problem of the simple cross-entropy loss not adequately supervising feature learning. Our method is evaluated on three commonly used datasets: ASGC, CCSN, and GCD. The experimental results show that the method achieves 94.24%, 92.73%, and 93.57% accuracy, respectively, outperforming other state-of-the-art methods. It proves that Transformer has great potential to be applied to GCI classification tasks.

The all-sky camera (ASC) images can reflect the local cloud cover information, and the cloud cover is one of the first factors considered for astronomical observatory site selection. Therefore, the realization of automatic classification of the ASC images plays an important role in astronomical observatory site selection. In this paper, three cloud cover features are proposed for the TMT (Thirty Meter Telescope) classification criteria, namely cloud weight, cloud area ratio and cloud dispersion. After the features are quantified, four classifiers are used to recognize the classes of the images. Four classes of ASC images are identified: “clear”, “inner”, “outer” and “covered”. The proposed method is evaluated on a large dataset, which contains 5000 ASC images taken by an all-sky camera located in Xinjiang (38.19^∘ N, 74.53^∘ E). In the end, the method achieves an accuracy of 96.58 % and F1_score of 96.24 % by a random forest (RF) classifier, which greatly improves the efficiency of automatic processing of the ASC images.

Stellar classification is a central topic in astronomical research that relies mostly on the use of spectra. However, with the development of large sky surveys, spectra are becoming increasingly scarce compared to photometric images. Numerous observed stars lack spectral types. In Sloan Digital Sky Survey (SDSS), there are more than hundreds of millions of such stars. In this paper, we propose a convolutional neural network-based stellar classification network (SCNet) in an attempt to solve the stellar classification task from photometric images alone, distinguishing between seven classes, i.e. O, B, A, F, G, K, and M. A total of 46 245 identified stellar objects were collected from the SDSS as the training samples for our network. Compared to many typical classification networks in deep learning, SCNet achieves the best classification accuracy of 0.861. When we allow an error to be within three neighbouring subtypes for SCNet, the accuracy even reaches 0.907. We apply the final SCNet model to 50 245 638 SDSS stars without corresponding spectra and present a new star classification catalogue, containing 7438 O-type stars, 31 433 B-type stars, 201 189 A-type stars, 910 007 F-type stars, 10 986 055 G-type stars, 18 941 155 K-type stars, and 19 168 361 M-type stars.