Research On Remote Sensing Image Scene Classification Based On Deep Learning

With the rapid development of core technologies in the field of remote sensing,The application of remote sensing image is more and more extensive,at the same time,the classification of remote sensing image scene is an important part.So the accuracy of remote sensing image scene classification is improved plays a key role in agricultural production,military identification and environmental detection.Based on the current research status at home and abroad,this paper starts from the purpose of comprehensive and effective use of feature extraction methods and neural network structures,in order to improve the accuracy of classification.deep learning theory and neural network models were researched,meanwhile,the effective spatial features contained in remote sensing scene images and suppresses interference features in training models were improved.Reach the goal of improving classification accuracy.The main research contents and results of this paper are as follows:1.The methods for scene classification of remote sensing images at home and abroad were studied,and the traditional methods and deep learning methods were compared,and their advantages and disadvantages were fully analyzed.2.Aiming at the problems of rich spatial information and redundant geographic features in remote sensing image scenes that would cause interference during network training,a method of remote sensing image scene classification using re-calibrated feature fusion dense neural networks was proposed.The SE block was established through the condensation and excitation mechanism to improve the correlation between channels and adaptively recalibrate the feature mapping between channels in the network model,moreover,the SE block and its multi-scale branches were embedded in DenseNet-121 for Feature recalibration uses dense connection methods,to enhance the transfer of information flow and enabling the network model to selectively extract effective remote sensing scene image features and suppress interference features through global information.This method allowed the overall model to obtain a robust feature representation of the global receptive field,and reduced redundant mapping of remote sensing scene features.Through experiments on two public remote sensing image data sets,the experimental results obtained are superior to other traditional classification methods,which proves the effectiveness of the method.3.In the course of research,it was found that the scene data of remote sensing images were difficult to obtain,the cost of manual labeling was large,and redundant geographic features will reduce the generalization ability of the model.Unsupervised remote sensing image scene classification method for adversarial domain adaptation.First,by embedding the squeeze and incentive mechanism SE block in VggNet16,it was used to improve the interdependence between channels to adaptively recalibrate the characteristic response between channels,so that the network model selectively extracts effective remote sensing through global information Scene image features and suppress interference features.Then,a large source domain data set was established and the adversarial domain adaptation method was used to reduce the feature differences between the source and target remote sensing image scenes,so that the generalization ability of the model was enhanced.Experiments show that the classification effect of this method was better than the current unsupervised method,and the accuracy was effectively improved in the classification of remote sensing image scenes.