The Research On The Intelligent Classification Technology Of Ground-based Visible-light Cloud Image Based On Transferred Convolutional Neural Network

The formation and evolution of clouds are the macroscopic manifestations of physical processes in the atmosphere.Ground-based cloud observation plays an important role in the observation,recording and research of weather phenomena.In recent years,the ground-cloud automatic observation system has been developing continuously,but the cloud classification has always been a difficulty to be solved.People have been trying to figure out how to describe different kinds of clouds differently from visual representations,and most of the existing methods use hand-crafted visual descriptors,and the results are not satisfactory enough.Inspired by the great success of convolutional neural network(CNN)in the largescale image classification task,this paper first proposed a basic convolutional neural network model and explored whether CNN can effectively capture the characteristics of ground-based cloud image.Experimental results on two different databases showed that compare with several traditional feature extraction methods,CNN achieves better results.Secondly,in the case of insufficient data being flagged and additional categories and task complexity being required,We proposed a ground-based visible-light cloud image classification method based on transferred convolutional neural network,in which the sample base was expanded by preprocessing technology and the optimal fine-tuning scheme was found by layer-by-layer fine-tuning method.The results showed that a more satisfactory classification accuracy can be obtained by using the migrated convolutional neural network than the conventional one trained from scratch.This paper also analyzed the basic features and key parameters involved in CNN.Several cloud image databases were established and their characteristics were analyzed.The operation process of CNN was demonstrated more intuitively by using visualization technology.The conclusions obtained in this paper are helpful to further optimize the network structure and relevant parameters,and provide a reference for the practical application of CNN in the classification of ground-based visible-light cloud images.