Research On Target Classification Method Of Remote Sensing Image Based On CNN

In this paper,the research significance of remote sensing image classification and the development status of remote sensing image classification in recent years are briefly summarized,and the traditional classification methods and recent hotspot deep learning methods are analyzed.The deep learning methods,especially the convolutional neural network model,are carried out.On the basis of in-depth research,the classical convolutional neural network model was used as a classifier to carry out remote sensing image classification experiments.In the research of traditional remote sensing image classification methods such as bag of visual word and artificial neural network,it is found that such methods usually need to manually select features and extract when performing classification tasks.The accuracy of this classification method depends on the extraction features.The quality is not high enough to cope with the challenges of big data.To solve this problem,this paper proposes to use large-scale deep neural network to improve the automation degree and classification accuracy of remote sensing image classification.The image classification model DenseNet was introduced,which was improved by the feature recalibration strategy to realize remote sensing image classification.For the problem of less experimental data,the weight migration strategy was adopted to greatly shorten the training time and avoid over-fitting.Finally,in the Tensorflow environment of Linux system,four types of experiments based on DenseNet model,based on SE-DenseNet model,based on SE-DenseNet of improved training method model,based on weight migration for remote sensing image classification are carried on the NWPU-RESISC45 remote sensing image dataset.And finally comparing the experimental results with traditional methods and deep learning methods.The comparison with traditional methods verify the feasibility of improving the degree of automation and classification accuracy,and the comparison with the deep learning method verify the effectiveness of the proposed algorithm.In the study of large-scale deep neural networks,it is found that large-scale deep neural networks are easy to extract the essential features of the samples,but at the same time,the requirements for the number of samples are correspondingly improved;in some areas of classification,the establishment of target data sets is usually subject to various subjective and objective conditions.It is difficult for individuals to obtain large-scale data or the data categories and quantities obtained are quite different.The remote sensing image classificationmethod based on weight migration set by training through large-scale data.In response to this problem,this paper proposes to use small-scale deep neural networks to reduce the model's requirement for sample size.The image classification model VGG is introduced,which is improved by activation function,batch normalization layer and global average pooling to realize remote sensing image classification.In the training method,the over-fitting is prevented by the Dropout method,and the classification accuracy is improved.Finally,in the Tensorflow environment of Linux system,the NWPU-RESISC45 remote sensing image dataset was classified based on VGG-16-ReLU remote sensing image classification,based on VGG-16-Leaky remote sensing image classification,and based on VGG-14 remote sensing image classification.Finally,the experimental results are compared with the remote sensing image classification algorithm based on the improved training method SE-DenseNet model and other people's algorithms.The comparison results verify the feasibility of reducing the model's requirement for sample data and improving the accuracy of classification accuracy through the small-scale deep neural network.The results of comparison with other people's algorithms verify the effectiveness of the proposed algorithm.