Feature Learning Based High Resolution Remote Sensing Image Scene Classification

With the rapid development of the techniques of the satellite sensors,a considerable amount of high-resolution remote sensing(HRRS)images are now available,which brings a great challenge to the information extraction and interpretation of remote sensing images.The classification of HRRS images is a significant task for the interpretation of remote sensing images.However,the traditional pixel-level and object-oriented classification methods can only make a distinction between different ground objects,which has no access to the high-level semantic information.Therefore,scene classification of HRRS images has become an active research topic for the interpretation of remote sensing images.In recent years,several scene classification approaches have been proposed for HRRS images based on feature coding,which can extract the semantic information effectively.Nevertheless,there are still some problems and challenges in the existing methods:(1)The traditional feature coding-based methods need to first extract the low-level handcrafted local features of the scene images.However,the insufficient descriptive ability of low-level features limits the performances of the feature coding-based methods.(2)The complicated dictionary learning and feature coding processes used in the feature coding-based methods make the global feature representation complex and time-consuming.(3)Most of the existing methods simply extract the mid-level features of the scene,which has no access to the high-level semantic information,and thus the limited descriptive ability and discriminative ability of mid-level features affect the classification results.To overcome the existing problems in the traditional scene classification methods,this thesis studied the scene classification methods for HRRS images based on unsupervised feature learning and deep learning.The main contents and contributions of this thesis are as follows:(1)A scene classification method is proposed based on the unsupervised feature learning(UFL)via spectral clustering(SC).To overcome the dependency on the handcrafted low-level feature which cannot effectively describe the scenes,this method automatically learns the local features by the UFL algorithm from the original pixel information of the HRRS image patches,which can represent the scenes of the remote sensing images adaptively.Meanwhile,to solve the high dimension problem,a UFL-SC based feature extraction framework is proposed to increase the efficiency of the local feature learning by mapping the original image patches into a low-dimensional and intrinsic feature space.(2)A fast binary coding(FBC)scene classification method is proposed.To overcome the problem in the traditional methods that the low/mid-level feature extraction,dictionary learning,and feature coding steps are separated and time-consuming,this method can adaptively learn the filters that can effectively describe the local textures and structures by the UFL algorithm,and quickly obtain the local features by binarization of the filter responses.The FBC method unite the local feature extraction,coding and pooling processes by simple linear filtering and binarization,thus the efficiency of the feature representation is strongly improved without sacrificing the performance of classification results.Compared with the traditional methods,the FBC method has a great advantage in computation efficiency.(3)A scene classification method is proposed based on the transfer of deep convolutional neural networks(CNNs).Based on the idea of transfer learning,this method directly uses the pre-trained CNNs to extract the features of HRRS images.Two scenarios are proposed for generating image features via extracting CNN features from different layers.Considering the specific characteristics of the CNN structure,extracting multi-scale CNN features and encoding them can enable the global features of the scene with multi-scale information.The experimental results on the public remote sensing scene datasets demonstrates that the pre-trained CNNs can generalize well to remote sensing scenes,since the features extracted from the convolutional layers and fully-connected layers contain high-level semantic information and have great descriptive ability and discriminative ability of the scene.(4)A topic model based on deep CNN features is proposed for HRRS scene classification.This method combines the excellent transferring ability of the deep CNN features with the probabilistic latent semantic analysis(pLSA)model to mine the relationship between the deep CNN features and scene topics.Due to the different response forms among different layers of CNN,two methods are developed to generate semantic features,called multi-scale deep semantic representation(MSDS)and multi-level deep semantic representation(MLDS).The topic features generated by the above two methods can extract more abstract and discriminative information with lower dimension compared with the original deep features.By fully utilize the scale and spatial information,the descriptive ability of the scene is further imaproved.(5)A deep sparse representation(DSR)method is proposed for scene classification.Since the sparse coding algorithm can reduce the reconstruct error and mine the salient characteristics of features,this method builds scene-level features via directly mapping the features extracted from the convolutional layers of pre-trained CNNs on high-dimensional sparse feature spaces,and applying a simple max pooling.The experimental results demonstrates that the scene-level DSR features can obtain remarkable classification accuracies on the public remote sensing scene datasets even with a simple linear classifier.This thesis proposes various remote sensing scene classification methods under different conditions for different demands,which not only improve the efficiency of feature extraction,but also improve the performance of scene classification.Therefore,this thesis has important theoretical and practical significance.