Research On Remote Sensing Image Classification Based On Convolution Neural Network

With the updating of sensing technology and the rapid development of Aerospace Science and technology,the resolution of satellite remote sensing image is increasing,and the data of remote sensing image becomes more and more complex.The characteristics of remote sensing image data,such as multivariate,temporal,multi spatial resolution and complexity of ground objects,make image processing difficult to extract effective information from mass data,and lead to high redundancy of data and difficulty in adding remote sensing scene tags.Therefore,how to classify the scene more efficiently and accurately affects the development of remote sensing technology.Remote sensing scene classification is a process of learning to map images to semantic content labels.Traditional image feature extraction methods,such as support vector machine,K mean and Gauss mixed classifier,and the algorithm based on principal component analysis can only extract the image data features of the shallow layer.The artificial neural network can abstract the features and get the deep features of the image.In recent years,convolutional neural networks have made great progress in image recognition and classification tasks.Compared with the traditional classification method,the neural network must not assume the probability model in advance,and has very strong learning ability and fault tolerance.It is suitable for the various problems of spatial pattern image recognition.In order to effectively utilize the superiority of convolution neural network in image processing,this paper proposes a remote sensing image classification method for multi-scale maximum output convolution neural network and a remote sensing image classification method based on cyclic learning rate and migration learning.In this paper,the structure and training process of the perceptron,the single layer neural network and the multilayer neural network are briefly introduced.Then the characteristics of the convolution neural network and the process of updating the weight by the positive backward propagation method are briefly described.Then the sparse network structure of the traditional convolution neural network can not retain the shortcoming of the high efficiency of the full connection network,and the determination of the activation function is selected by experience or the training results of different experimental functions in the traditional case,which reduces the accuracy rate and increases the amount of calculation outside the amount.Remote sensing image classification method based on maximum output convolution neural network.The method first inserts the Inception module before the third convolution layers,so that the convolution kernel of different scales can extend the network width and improve the ability of network feature extraction.Then the Maxout network is connected before the full connection layer,so that the network can improve the image classification accuracy on the basis of fitting any active function under the operation of dropout.Finally,the training features are sent to the softmax classifier for classification.The second methods want to shorten the experiment time on the basis of the proposed method,let the network extract the characteristics of the remote sensing image and improve the classification accuracy in a relatively short time,so that the network which has been pre trained has acquired the related image features to classify the remote sensing scene image,and then proposes a kind of migration based on the migration.A learning convolution neural network algorithm:a remote sensing image classification method based on cyclic learning rate and transfer learning.This method can accelerate the training process and alleviate the difficult problem of remote sensing image data set to a certain extent.It can also effectively prevent the local minimum vibration caused by adaptive selection learning rate.The most important thing is that this method greatly improves the accuracy of the remote sensing image scene classification.Finally,experiments on UCM_LandUse₂1 remote sensing images verify the feasibility of the above two methods and the effectiveness of each improvement.