Research On Feature Extraction And Selection For Very High Resolution Image Scene Classification

The fast development of remote sensing instruments and technologies,provide more opportunities to deeply observe the earth,by multi/hyperspectral images and synthetic aperture radar.Remote sensing devices help us to capture more and more different type of airborne or satellite images with different resolution(spatial resolution,spectral resolution,and temporal resolution).The high dimension and of these acquired images are the most challenging acts for Very High Resolution(VHR)scene classification.This asks more efficient methods for Land Use and Land Cover(LULC)image scene classification,which become a critical task in remote sensing image community.The most important step in VRH image scene classification is the features extraction,which describes and represents the image scene with features vector.The existing scene description methods for VHR images classification can be classified as three categories of techniques depended on the pixel-/object-level images representation,where VHR image scene classification techniques depend directly on global representation of the image scene.In this thesis,we focus on the features extraction approaches for VHR image scene classification,in order to propose and develop some techniques,which allows to precisely categorizing the different area on the ground and geometrical proprieties from VHR image scene such as(airport,building,forest,agriculture,etc).The VHR images provide very useful information for several applications related to monitoring the natural environment under the results of human activities.VHR image scene classification is devoted to extracting the features that represent the object area.However,the huge amount of data associated with VHR images makes the classification problem very complex and the available approaches are still inadequate to analyses this kind of remote sensing data.For this reason,the general objective of this thesis is to propose novel techniques for the analysis and the classification of VHR images,in order to improve the capability to automatically extract useful and informative features from VHR images scene.In particularize the following specific issues are considered in this work:1)How to achieve good feature representation from a large amount of VHR data,is still a critical task for scene analysis in VHR images.We present a new method for VHR scene classification based on sparse handcraft features selection.In order to extract more efficient and robust features to classify the image scenes based on handcraft features First,handcraft operators are used to extract local features from the original VHR images to construct a visual dictionary.The sparse principal component analysis(s PCA)is then adopted to learn a set of informative features from the visual dictionary for each category.Finally,the scenes are represented by sparse informative low-level features.2)We present a new method named Salience Patches Sampling for VHR Images Classification(SPSIC).s PCA is then adopted to select the corresponding informative salient patches for image scene representation.The proposed technique for selecting informative salient patches is efficient and robust for scene understanding.3)Labeling VHR image scene according to a set of semantic categories is a very critical point,we present a framework to explore CNN for VHR Image Semantic classification,because land covers characterizing a given class may present a large variability and objects may appear at different scales and orientations.First,the pre-trained Visual Geometry Group Network(VGG-Net)model is proposed as deep feature extractors to extract informative features from the original VHR images.Second,we select the fully connected layers constructed by VGG-Net in which each layer is regarded as separated feature descriptors.And then we combine between them to construct final representation of the VHR image scene s.Third,Discriminant Correlation Analysis(DCA)is adopted as feature fusion strategy to further refine the original features extracting from VGG-Net,which allows a more efficient fusion approach with small cost than the traditional feature fusion strategies.4)Even though some literature has made advantages of deep features for VHR image scene representation,how to optimize the transferability of CNN models for VHR image scene understanding is a very challenging problem.With this in mind,we present a new method which explores the strengths of different CNN layers in a simple and effective way.To represent both of semantic and background of the VHR image content,the pre-trained CNN models are used to extract informative features from the original VHR images scene.Then,we select both the convolutional layers constructed by CNN in which considered same as handcraft features.Where convolutional layers are used to generate visual words correspond to the input images,by using coding methods.For the VHR image scene classification problem a deep study of the literature is carried out and the limitations of currently published techniques are highlighted in section 2.1,starting from this analysis,novel approaches are theoretically proposed,implanted and applied to real remote sensing images,in order to verify their effectiveness.The achieved experimental results confirm the effectiveness of all the proposed techniques.