Research On Satellite Image Scene Classification Technology Based On Convolutional Neural Network

With the rapid development of earth observation technology,the number of satellite images has increased significantly,and the valuable information contained in a large number of satellite images urgently needs to be tapped and utilized.In this paper,the satellite images are divided into residential areas,vegetation areas,etc.The classification of satellite image scenes can accurately grasp the changes of ground cover.The acquisition of this information can assist relevant departments to manage the environment,rationally use land resources,and plan urban development.Satellite images are often rich in information,complex in background,and noisy.Due to the influence of these factors,the traditional satellite image scene classification method based on low-level visual features is difficult to obtain an ideal classification result.After that,shallow machine learning algorithms such as support vector machines(SVM)perform well in image recognition and classification tasks,but due to their simple structure,these algorithms are difficult to fit complex functions,and the generalization ability of satellite image scene classification problems is still insufficient.In recent years,with the rapid development of deep learning and artificial intelligence,the application scope of deep learning is getting wider and wider.Among them,the convolutional neural network performs well on image scene classification problems.This paper classifies satellite image scenes based on convolutional neural networks,and uses a variety of different convolutional neural networks for experiments.First,this paper summarizes the design methods and basic principles of convolutional neural networks,and designs three convolutional neural network structures with different complexity and two fully connected layer structures with different complexity accordingly.In order to optimize the designed neural network model,various methods such as batch regularization(BN)and dropout technology are introduced.In the experimental part,this paper selects the RSSCN7 data set,and designs experiments to compare the effects of different complexity neural network structures on the experimental results.After comparing the classification accuracy,training memory and training time and other factors,it is determined that the best neural network structure and fully connected layer structure.Then use the control variable method to determine the hyperparameters of the network model through multiple sets of experiments,including the learning ate,the number of epochs and the batch size.In the end,the designed convolutional neural network achieved a classification accuracy of 91.8% on the data used in this article.Then,in order to further explore the convolutional neural network model with better classification effect on the data set in this paper,this article selected three classic convolutional neural network models for experiments.The three models are: Le Net,vgg-f and vgg16.Among them,the Le Net model is relatively simple,you can use the RSSCN7 data set for direct training,and improve the classification accuracy of the network through optimization models and parameter adjustment methods.The vgg-f and vgg16 models are more complicated.In this paper,the transfer learning method is used to train the pre-trained models of the two models on the data set used in this article.The parameters in the pretrained model are used as the initialization parameters for the second training.The experimental results show that the classification accuracy of the Le Net model is low,73.5%.The vgg-f and vgg16 models based on transfer learning have achieved classification accuracy rates of 92.3% and 92.9%,respectively.Finally,it compares and analyzes the advantages and disadvantages of the convolutional neural network designed in this paper and three classic convolutional neural networks in the classification of satellite image scenes.