Research On Automatic Classification Algorithms For All-sky Cloud Image

Capturing cloud images using all-sky cameras is now widely used in the astronomical world to monitor cloud cover.The estimation of cloud cover will have an important impact on telescope observation and site selection.At present,the National observatory according to the interpretation of Thirty Meters Telescope to interpret cloud cover manually divides cloud images into four categories,including clear,covered,inner and outer.Because the manual interpretation cloud image is time-consuming and the interpretation process depends entirely on personal experience,so the paper classifies 16387 all-sky camera cloud images based on six classification algorithms,including threshold method,support vector machine,k-nearest neighbor,random forest,logistic regression and convolutional neural network.In addition,we compare the accuracy of classification with the six classifiers,the experiment results show that the accuracy of convolutional neural network is 90.8%,which is improved by 11.8% compared with support vector machine algorithm with the highest classification accuracy in the other five algorithms.Therefore,convolutional neural network algorithm can effectively liberate the manpower,but also improve the accuracy of the all-sky camera cloud image classification.The main results of this paper are as follows:(1)All-sky camera cloud image data preprocessing.Because all-sky camera cloud image data is complex and changeable,so this paper first processes the cloud image data through cloud image enhancement and cloud image denoising.Among them,cloud enhancement uses morphological high-low hat method;cloud image denoising is to construct different models for four kinds of noise,such as Moon,Star point,lightning and reflection,so the noise removal is accomplished effectively.The results show that cloud image preprocessing effectively reduces the complexity of cloud image data samples.(2)Cloud images classification based on threshold method.In this paper,cloudy and less cloud images are used for cloud detection by using Otsu of adaptive threshold method,and then cloud cover is calculated and cloud classification is completed.Experimental results show that compared with the artificial interpretation of cloud cover,the method has higher efficiency.(3)Cloud images classification based on four traditional machine learning methods.In this paper,the texture and statistical characteristics of cloud image are extracted by Gray-Gradient Co-occurrence Matrix,and then using four traditional classifiers,including support vector machine,random forest,k-nearest neighbor and logistic regression to classify cloud images.Experimental results show that the classification results are superior to the threshold method in efficiency and accuracy.(4)Cloud images classification based on convolutional neural network.In this paper,the convolutional neural network model is constructed by Tensorflow frame,and the model is improved appropriately.Experimental results show that convolutional neural network has been successfully applied to cloud image classification,and the accuracy is more improved than the traditional classification algorithm.This study builds a complete cloud image automatic processing system,which can realize the automatic classification of cloud images,and fully corresponds to the TMT classification method of the current astronomical world,which can mostly replace the corresponding manual processing work.Finally,this study makes a detailed comparative analysis of the operation efficiency and classification results of the above six classification algorithms.Experimental results show that the convolutional neural network has the best effect in terms of operation efficiency and accuracy,and the automatic classification of all-sky camera cloud image is completed efficiently.