Research On Remote Sensing Image Scene Classification Based On Convolutional Neural Network

The fast development of instruments and technologies in remote sensing,provide more the opportunities to deep observe earth,by multi/hyperspectral images and synthetic aperture radar.Remote sensing instruments help us to capture more different type of airborne or satellite images acquired with different resolution(spatial resolution,spectral resolution,and temporal resolution).The high dimension of these images are the most challenging acts for high resolution(HR)scene classification.This asks more efficient methods for Land Cover and Land Use image scene classification,which become an important task in remote sensing image community.The most critical step in HR image scene classification is the features extraction,which describes and characterizes the image scene with features vector.The existing scene description methods for HR images classification can be classified as three categories of techniques depended on the pixel-/object-level images representation,where HR image scene classification methods depend directly on fully representation of the image scene.In this thesis,we interest on the features extraction approaches for HR image scene classification,to allow us to propose and develop some techniques,which lead to precisely categorizing the different area on the ground and geometrical proprieties from HR image scene such as(airport,building,forest,agriculture,etc).The HR images contains very useful information for several applications related to monitoring the natural environment under the results of human activities.HR image scene classification is devoted to extracting the features that represent the object area.However,the huge volume of data associated with HR 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 principal objective of this thesis is to propose novel techniques for the classification and analysis the of HR images,in order to perform the ability to automatically extract useful and informative features from HR images scene.In particularize the following specific issues are considered in this work.Labeling HR image scene according to a set of semantic categories is a very critical point,we propose a framework to explore CNN for HR Image Semantic classification,because land covers characterizing a given class can present a large variability and objects can appear at different scales and orientations.First,the pretrained Visual Geometry Group Network(VGG-Net)model is based as deep feature extractors from the original HR images.Second,we select the fully connected layers built by VGG-Net in which each layer is considered as separated feature descriptors.And then we put them into auto-encoder to make final representation of the HR image scenes.For the HR image scene classification problem,a deep study of the literature is performed and the limitations of many published techniques are highlighted in section 2.1,starting from this analysis,new approach is theoretically proposed,implanted and used for real remote sensing images,in order to verify their effectiveness.The achieved results confirm the efficiency of all the proposed techniques.