Research And Application Of Remote Sensing Image Scene Classification Technology Under Multiple Tricks

As the technology of remote sensing imaging grows mature gradually,research on classification methods of remote sensing image scene is receiving widespread attention.Due to its characteristics of large diversity within class and high similarity among classes,classification task of high-resolution remote sensing image scene is facing great challenges.Combined with deep approach of learning,the dissertation tries to improve the performance of remote sensing image scene classification by adopting three research tricks including single training trick,fusion training trick and model ensemble trick.A large number of experiments verify the effectiveness of the above strategy in high-resolution remote sensing scene classification tricks.The main research content and achievements of this dissertation are as follows:(1)The research on remote sensing scene classification mainly focuses on data features and neural network structure,but the influence exerted by neural network training tricks on performance of remote sensing image scene classification is rarely mentioned.To solve the problems above,this dissertation selects seven neural network training tricks frequently used in scene image classification for experiments,and selects the neural network training tricks suitable for remote sensing image scene classification.The dissertation detailedly evaluate several neural network training tricks about remote sensing image scene classification through ablation experiment and superposed training.The valid neural network training tricks are obtained from the analyses of overall accuracy,confusion matrix and kappa further.The result proves the validity and applicability of the neural network training tricks on the performance of the remote sensing image scene classification.(2)The research on remote sensing image scene classification based on deep learning technology ignores the impact of method ensemble tricks on the performance of remote sensing image scene classification,which leads to low efficiency when capturing hierarchical structures of image features and limited improvement of remote sensing image classification performance.Based on Res Net,this dissertation proposes a valid remote sensing image scene classification framework CNN-Ensenet by analyzing the performance of method ensemble tricks in remote sensing image scene classification.CNN-Ensenet is able to make full use of the advantages of varying classification methods.Additionally,the methods involved in this dissertation are also extended to models such as VGG-16,googlenet and Res Net models for experiments,verifying CNN-Ensenet has better classification performance and presents good migration ability and applicability in various convolution models.The ablation study of the single training tricks shows that it performs best on the Alex Net and UC Merced Land-Use dataset(training set is 80%),and the overall accuracy is increased by about 1.24%.The experimental results of the superposed training tricks show that the overlay strategy is generally effective in improving the classification accuracy on different models and data sets.It performs best on the Goog Le Net model and the NWPU-RESISC45(training set is 10%)data set,and the overall classification accuracy is improved by approximately 6.2%.The experimental results of CNN-Ense Net based on the model ensemble tricks show that the overall accuracy on UC Merced Land-Use(training sets are 50% and 80%)is 98.76% and 99.97%.The overall accuracy on AID(training sets are 20%and 50%)is 95.36% and 96.98%,and the overall accuracy on NWPU-RESISC45(training sets are 10% and 20%)is 93.13% and 95.10%.CNN-Ense Net performs well on different datasets.In addition,the results of the transfer experiment of CNN-Ense Net show that the extended model structure(VGG-Ense Net and Google-Ense Net)based on the CNN-Ense Net framework shows excellent classification performance on each dataset.