| With the large number of high-resolution images now being acquired and the rapid development of high spatial resolution(HSR)remote sensing image processing technology,the classification methods for HSR images have evolved from per-pixel-oriented methods to object-oriented methods,taking advange of clear details and rich spatial and texture information in HSR images for classification on the object level,greatly improving the feature classification results.However,current object-oriented classification can only reach the ground object level,such as buildings and roads.It is still difficult to obtain the high-level scene semantics,such as industrial areas and residential areas.In order to obtain high-level scene semantic information,how to cross the “semantic gap” between low-level features and high-level semantic information and realize the mapping from HSR to high-level scene semantic is a hot issue.To overcome the semantic gap,some scholars have proposed scene classification methods based on object recognition and middle-level features.However,these methods all require handcrafted features and rely on expert knowledge and data prior.Due to the excellent performance on the ability of automatically learning essential features from image data,deep convolutional neural network(CNN)has draw great attentation in remote sensing image processing field and has been successfully applied in scene classification for HSR remote sensing images.However,the scene classification based on CNN still has the following problems: 1)Decaied classification result with limited labeled training data.The scene classification based on CNN often requires a large amount of annotated data for training model.When the training data set is limited,the feature generalization ability of the model is poor.2)Weak robust to scale change.The object scale changed in image due to the altitude and angle change of the sensor or scale diversified characteristics of the object.The traditional CNN is trained with fixed single scale,which makes the learned features weak robust to object scale change and leads to wrongly classified scene images containing objects with a changed scale.3)Classifier suffering from the intraclass diversity and the interclass similarity.In remote sensing scene,the problem of within-class diversity and between-class similarity is still a big challenge.However,the common classifier SoftMax does not explicitly maximize the intraclass similarity and minimize the interclass similarity.The problem caused by the intraclass diversity and the interclass similarity is not effectively alleviated.To overcome the challenges of CNN for HSR scene classification,this thesis develops scene classification methods based on CNN model for HSR images from three aspects of data,feature extraction and scene semantic classification.The major works and contributions of this thesis are listed below:(1)The characteristics of HSR image scenes,the current state and the problems of HSR image scene classificatin are systematically summarized in this thesis.The basic theory of CNN is introduced,and this thesis also specifically analyzes the advantages and practical potential of CNN for HSR scene classification.(2)In the aspect of data,a deep semi-supervised framework based on local label propagation is proposed for remote sensing scene classification.In general,the large of labeled data is usually required to train the network.To train the CNN with limited labeled data,the local label propagation is proporsed for label propagation between the unlabeled data and labeled data,and the cross-entropy loss objective function is computed to maximize the label propagation weight between the samples with the same label and minimize the label propagation weight between the samples with the different labels.The proposed semi-supervised framework can effectively improve the remote sensing scene classification accuracy of the CNN with limited training data.(3)In the aspect of feature extraction,the multi-scale CNN for HSR remote sensing scene classification is proposed.The scale of the objects can change greatly between images and the methods based on CNN mostly train the network at a fixed scale,which makes the features weak robust to scale.To solve this problem,this thesis models the scale change of the objects in image to generate multiple-scale samples to train the CNN model,forcing the trained CNN model to correctly classify the same image with different scales.To fuse the information from different-scale patches with different locations in the image,thereby further improving the performance,this thesis applies multiple views of one image and conducts fusion of the multiple views.The patches with a random position and scale are cropped and then stretched to a fixed scale to be classified by the trained CNN model.Finally,the whole image's label is decided by selecting the label that occurs the most.(4)In the aspect of scene semantic classification,an end-to-end framework for the extraction of compact deep features for scene classification is proposed.To solve the problem of the high intraclass diversity and interclass similarity in scene classification,a new cluster loss objective function is proposed,to ensure that the training samples are close to their corresponding class center and that the class centers of all the classes move away from each other,which promotes the discriminative ability of the CNN.(5)HSR image scene understanding framework based on CNN is built.Combined with HSR scene scene classification method based on CNN proposed from multiple perspectives,the HSR image scene classification framework based on CNN is built for different demands.This thesis conducts the research of CNN in data,feature extraction and scene semantic classification for HSR image scene classification.This is able to improve the performance of HSR scene classification,and further promotes the practical application of HSR scene classification,therefor have significant scientific and social values for many fields,such as image retrieval and environmental monitoring. |
PDF Full Text Request