Scene Classification Of High Resolution Remote Sensing Images Based On Transfer Learning

In recent years,with the rapid development of high-resolution remote sensing image acquisition technology,people can easily obtain a large number of high-resolution remote sensing images with sufficient information and detailed features.The recognition and classification of complex scenes of remote sensing image has become one of the research hotspots.It can extract and analyze the information in the image,which is widely used in surface coverage,image archiving,natural disaster prevention and so on.However,the traditional classification methods based on image features can’t obtain detailed information from high-resolution remote sensing images,and the ground objects in the images have the characteristics of large differences among the same class and high similarity between some classes.Therefore,it is of great significance to study how to automatically and effectively classify scenes.The classification method based on low-level image features has insufficient ability to describe the scene,and the classification accuracy is not high.With the development of deep learning and transfer learning,neural network has made continuous breakthrough in the field of image.Applying transfer learning method to high-resolution remote sensing image field classification task has become one of the popular research methods,which makes the classification accuracy significantly improved.On the basis of transfer learning theory,this paper studies how to use neural network model to extract image features for scene classification,and to identify a variety of scene categories in real complex scene images.The research contents of this paper include: firstly,collecting and sorting the commonly used scene classification data sets of high-resolution remote sensing images,and making a new scene classification data set suitable for high-resolution remote sensing images,which includes 18 categories of scene categories,600 of each category,a total of10800.Secondly,the remote sensing image scene classification model is built by using transfer learning method,and the pre training vggnet16 and resnet50 network models are improved.In the training process,multiple groups of comparative experiments are set up,and the optimizer with better effect is selected to optimize the model.The data expansion method is used to improve the classification accuracy of the model.It makes the model achieve a good effect of scene classification in the case of a small number of training samples and training time.Finally,a real large-scale remote sensing image scene clipping strategy is proposed,and the adjusted and trained vggnet16 scene classification model is applied to the actual large-scale remote sensing image,and the scene categories are identified.The experimental results show that the improved vggnet16 and resnet50 network model proposed in this paper,using transfer learning method to classify high-resolution remote sensing image scene,can achieve good results,using Adam optimizer and data augmentation method can effectively improve the classification accuracy and model performance,and the application in real large-scale remote sensing image has also achieved good results.