Remote Sensing Images Classification Based On Convolutional Neural Network

Classification of remote sensing images plays a significant role in many applications.For instance,it is a key step in natural hazards detection,LULC determination,geospatial object detection,geographic image retrieval,vegetation mapping,environment monitoring,and urban planning.In the past couple of years,remote sensing technology has been greatly advanced and due to which,many techniques to acquire more and more airborne or satellite remote sensing images with different resolutions(spatial resolution,spectral resolution,and temporal resolution)came into existence.The consequence of this phenomenon is that a large number of remote sensing scene image datasets(labeled)became publicly available.However,most of them possess higher intra-class variations and smaller inter-class dissimilarities.Scene classification on such high-resolution datasets is highly challenging and more importantly,most of the conventional methods often fail to achieve optimal classification accuracy.Hence,the methods to classify such datasets quickly and effectively are highly indispensable.The specific content and work arrangements are as follows:This thesis focuses on the research of remote sensing image classification based on convolutional neural network model.Firstly,the remote sensing images classification algorithms and research status of existing are introduced.Then,an efficient convolutional neural network model was proposed for the problem of high-resolution SAR images classification.The research of remote sensing scene images classification was investigated in two aspects.The specific contents and working arrangements were as follows:(1)In order to improve the classification accuracy of high-resolution SAR images,we proposed a convolutional neural network model which performs remarkably in the problem of high-resolution SAR images classification,named M-Conv Net.M-Con Net model consists of four convolution layers,three pooling layers layers and two full-connected layers.We composed details of the algorithm and the training settings.Finally,MSTAR high-resolution SAR image dataset is experimentally verified and summarized.(2)To improve the accuracy of remote sensing scene images classification,the method based on pre-training Res Net model in remote sensing scene classification was explored.We used three different Res Net pre-training models as feature extractors to extract feature vectors which is more representational.Besides,the algorithm of classification based on feature extraction and feature fusion about pre-trained Res Net models is proposed.Three different types of feature fusion are used in the experiments.Compared with traditional classification algorithms and other network models,classification accuracy got improved a lot.(3)In order to further improve the accuracy of remote sensing scene classification,the method of combining the pre-training network model and the transfer learning is described.The parameters of the pre-training network are used as the initial values of the parameters,and then fine-tuning the Res Net-50 network structure that was more consistent with remote sensing scene data using the images augmented previously.Finally,methods of extracting feature vectors using fine-tuned model and feature confusion is experimented between UCMD and NWPU-RESISC45 datasets.