Research On Scene Classification Method Based On Depth Features Of High-resolution Remote Sensing Image

With the development of remote sensing technology,the spatial resolution of remote sensing images is getting higher and higher.The traditional pixel level remote sensing image classification or object-oriented classification methods can no longer meet the current classification needs.Scene classification is a hot direction of remote sensing image interpretation.The features of ground objects are more detailed and contain richer semantic information,and the spatial relationship between features constitutes scenes with different semantics in high-resolution remote sensing images.Nowadays,the application of remote sensing data lags far behind the ability of remote sensing data acquisition.How to classify the scene conveniently and quickly in the massive remote sensing image data is the key issue to improve the utilization efficiency.In recent years,the rapid development of deep learning algorithms provides a new idea for intelligent remote sensing data interpretation,which were used for remote sensing image scene classification.However,there are still some problems in remote sensing scene classification based on deep learning method due to the complexity of remote sensing images.This thesis studies the scene classification of remote sensing images with machine learning methods and the deep features of high-resolution remote sensing images as the research object.The main research contents and conclusions are as follows:(1)Multi-kernel Support Vector Machine(SVM)using features extracted from Low Dimensional Convolutional Neural Network(LDCNN)is proposed to solve the problem that convolution neural network has high feature dimension and cannot express complex semantic information.We verify the rationality of this method using two public remote sensing datasets by comparative experiments,and also analyze the classification results.Besides,the transfer learning method is verified.The results show that this method can effectively improve the scene classification accuracy when compared with other traditional methods.It has achieved more than 99% classification accuracy based on the two standard data sets and the classification accuracy of transfer learning test is 94.76% and 92.25%,respectively.(2)A method of urban scene classification based on complementary information learning model is proposed.The image pairs from the two views of aviation and ground are input into the model,and uses the unified loss to learn the features from different views.For the model trained,the deep features of remote sensing images from two views are extracted for fusion and classification,so as to achieve the purpose of information complementarity and improve the accuracy of classification.The experimental results show that the classification accuracy of this method is 93.56% and 84.32% respectively based on two multi view remote sensing data sets,which is about 8% and 4% higher than that of single view classification.(3)A remote sensing scene classification test platform based on deep features is designed and implemented.The platform integrates the functions of network training,feature extraction,classification method,result analysis,test and example display.The platform has simple and clear interface,convenient operation and complete functions.It assists users in network training,test and analysis,and supports data management and visual display.