Research On Scene Classification Of Remote Sensing Image Based On Transfer Learning And Random Erasing

Remote sensing image scene classification,as an important research content in the field of optical remote sensing image,is widely used in the fields of natural disaster processing,urban and rural planning,and geological survey.Because the background of high-resolution remote sensing images is complex and contains rich information about the geometry and details of the features,and the remote sensing images are easily affected by environmental factors during the imaging process,the resulting remote sensing images may have shadows,which may cause some parts of the image features to be lost.All these have brought great difficulties to the study of remote sensing image scene classification.The classification performance of the currently known remote sensing image scene classification algorithms still has some room for improvement.How to improve the feature extraction ability and generalization ability of the classification model is the key in scene classification research.This thesis uses deep neural networks for transfer learning,and is committed to improving the feature extraction capabilities of the classification model,and for the characteristics of the remote sensing image itself,using data enhancement methods suitable for remote sensing images to improve the generalization performance of the classification model,aiming to further improve the accuracy of scene classification of remote sensing images.The main research contents include:(1)The research status of remote sensing image scene classification at home and abroad is introduced,and the performance and shortcomings of various remote sensing image scene classification algorithms are analyzed.By analyzing the characteristics of the high-resolution remote sensing image itself,the difficulties in the task of remote sensing image scene classification are summarized.It is pointed out that because remote sensing images are easily affected by environmental factors during the imaging process,the resulting remote sensing images may be shaded,which makes the image lose some of its features,and the image background is complex and rich in information,resulting in the existing classification model has the problem of insufficient feature extraction ability and generalization ability.Finally,it is pointed out that using the method of transfer learning to select an appropriate network structure to improve the feature extraction capability of the model,and according to the characteristics of the remote sensing image itself,adopting a data enhancement method suitable for remote sensing image to improve the generalization ability of the model can further improve the remote sensing image scene classification accuracy.(2)By implementing a deep neural network(Res Net-50 and Inception V3)with excellent results on the Image Net dataset,and then using the method of transfer learning,combined with the remote sensing image dataset used to improve the network structure,a remote sensing image scene classification algorithm based on multi-network deep transfer learning is proposed.By conducting multiple experiments on three public datasets(UC-Merced,AID,NWPU-RESISC),the effectiveness of the multi-network transfer learning classification model is verified,and the effects of two different network models on transfer learning are qualitatively compared.Finally,through comparison and analysis with several latest remote sensing image scene classification algorithms known,and combining the characteristics of the remote sensing image itself,the direction of further optimization of the algorithm is pointed out.(3)In order to further improve the accuracy of remote sensing image scene classification,this thesis proposes a RE-Res Ne Xt remote sensing image scene classification algorithm based on deep transfer learning.Through the use of Res Ne Xt-50 for migration learning,combined with remote sensing image data sets to improve the network structure,the feature extraction capability of the classification model is improved.And finally using random erasure data enhancement methods,the classification model has the ability to identify the image from the overall structure of the image,reducing the interference of the remote sensing image itself to the scene classification,thereby improving the generalization performance of the classification model.By conducting multiple experiments on three public data sets(UC-Merced,AID,NWPU-RESISC),the classification accuracy is significantly higher than the classification accuracy of the remote network image scene classification algorithm based on multi-network deep transfer learning,which proves the effectiveness of the proposed algorithm.Finally,it is compared with several latest remote sensing image scene classification algorithms currently known.The experimental results show that classification accuracy of the RE-Res Ne Xt remote sensing image scene classification algorithm based on deep transfer learning on three public datasets(UC-Merced,AID,NWPU-RESISC)is higher than most of the currently known remote sensing image scene classification algorithms.