Remote Sensing Image Scene Classification Based On ResNet And Dual Attention Mechanism

With the rapid development of remote sensing technology,more and more scholars have explored and studied the classification of multiple scenes in remote sensing images.In this kind of research,the ground object targets contained in the pictures are generally called scenes.Because manual classification has certain limitations in speed and accuracy,relying on computers to complete scene classification of remote sensing images has become the main technical means.Because manual classification has certain limitations in speed and accuracy,relying on computers to complete scene classification of remote sensing images has become the main technical means.Aiming at the problem that image features cannot be extracted quickly and effectively in remote sensing image scene classification based on traditional machine learning,resulting in inaccurate classification results,this paper proposes a remote sensing image scene classification method based on attention residual mechanism.The main contents are as follows:(1)The theoretical knowledge of each layer of convolutional neural network is described concretely,and several effective solutions to the overfitting phenomenon are given.Five structures of different depths of Res Net were studied and analyzed.The five structural models of Res Net were used to classify the UC Merced Land-Use data set,and the time and accuracy were compared respectively.Finally,the best Res Net101 basic model was selected for the experiment,and the full connection layer of Res Net was modified.(2)Data enhancement processing is carried out on UC Merced Land-Use data set.The data set is expanded through operations such as color dither,translation,and flip,and images containing complete ground object targets are selected and retained to expand the sample size of the data set.The classification results of data set s with different sample sizes are compared and analyzed.The experimental results show that the accuracy of model training and classification of the data set with large sample size is 0.7% higher than that of the original data set.Taking Resnet residual network as the benchmark model,the attention mechanism was introduced to construct the residual attention module for remote sensing image scene classification.According to the training results,constantly adjust the learning rate,batch size and other super parameters;Using Dropout technology to improve the model generalization ability,and finally according to the classification accuracy,precision and recall rate,confusion matrix,F1 score,six indicators to assess the Kappa coefficient model,residual attention mechanism in remote sensing images is verified on the effectiveness of the classification of the scene,and with other convolution neural network model and the accuracy analysis of machine learning algorithm,prove that the proposed method has many advantages.(3)To UC Merced Land-Use dataset and SIRI-WHU dataset and WHU-RS19 explore respectively mixed experiment data sets,The experimental investigation shows that the training of mixed data sets does not significantly promote the classification performance of the model,but in the scene classification of small sample data sets,combined the experiment the application migration study of hybrid data set method to choose the appropriate data set training model to combine with small sample data set.