The Study Of All-sky Cloud Classification Based On Deep Learning

Cloud is an import part of water and energy cycle in the atmosphere and plays an essential role in climate regulating.Thus,the cloud of real-time observation has significant effect on the production and living of mankind.Historically,the ground-based cloud observation was mainly done by a professional meteorological observer.Manual classification is not only an expensive undertaking,but also time-consuming and heavily relies on the experience of the weatherman.In recent years,with the development of digital image acquisition equipment,all-sky cloud images are increasingly large.Therefore,an efficient automatic image classification method is strongly required.Based on the traditional content-based image classification technology and the bag-of-words model theory,this paper focuses on the feature representation and classification method of all-sky cloud images.In this paper,a cloud classification method based on micro-structure model is proposed.The micro-structure of cloud images is a basic feature map which is mapped from a visual dictionary.And the dictionary is generated from a bag of local features by unsupervised clustering algorithm.The proposed method treats the all-sky cloud image as a set of micro-structures instead of a collection of traditional pixels.The weight histogram of the micro-structure is used as the feature vector of the image,and finally is fed into a SVM classifier to do the classification.Due to the small number of samples and the large resolution of the image,two kinds of deep network structures and different training methods are proposed in this paper.Firstly,according to the deep learning theory and its experiments,it is shown that the filter template of the first layer is mainly used to extract color and texture features,and the latter is regarded as a more advanced combination of the underlying features.Based on transfer learning,the network is trained to identify the cloud by fine-tuning operations.Secondly,an unsupervised self-learning method that combines with supervised-oriented learning is proposed to solve the images classification task with few labeled samples and high resolution.Using a large amount of unlabeled data,the unsupervised process can play a role of dimension reduction and feature learning.Meanwhile,most parametric learning is transferred to the unsupervised network.Finally,the deep model is trained to identify all-sky cloud images by using the label data.An all-sky cloud image dataset,jointly published by the Chinese Academy of Meteorological Sciences and Beijing Jiaotong University,is used in this paper to verify the performance of the proposed method.The experimental results show that the classification accuracy based on the microstructure model and deep learning classification method is 90.9%and 98.4%,respectively,which is 20%higher than state-of-the-art.