Remote Sensing Scene Classification Based On Multi-layer Network

In recent years,with the rapid development of remote sensing imaging system,people can obtain more and more remote sensing images with different resolutions.Remote sensing image has many advantages,e.g.,wide observing range,short cycle and abundant surface information,so it has become one of important methods for exploring the earth's surface in related institutions.How to identify different scenes effectively from massive remote sensing images is the key technology of obtaining related information.Remote scene classification aims to divide the images into different categories according to their content.Because remote sensing image has a very complex spatial structure and one scene usually covers multiple ground objects or land covers,the traditional methods applied to natural image classification could not describe the complex pattern of remote sensing image,which have a low accuracy.In this thesis,based on bag of words,feature coding and deep learning,we made a deep research on effective feature extraction of remote sensing image to classification task.The specific work is summarized as follows: We used single-layer unsupervised network to extract features of image patches and replaced the traditional local descriptors.The single-layer network make an operation on pixel value and describe potential characteristics of images well.Meanwhile,in the process of pooling,we used sub-regional pooling to retain the spatial arrangement of remote sensing image,which makes the global feature more representative and improves the classification accuracy.In order to depict high level feature of remote sensing image,we used convolutional neural network to extract features from image.We extracted features based on sliding window because the convolutional neural network requires fixed input size,then we replaced pooling with convoluting to get the final image representation.Dense extraction keeps the specific information of every local region and avoids the loss of information in changing the size of image.The operation of convolution makes the final representation more discriminative by giving different weights to each local feature.Finally,we proposed a framework which called deep differential coding to extract feature.It combines convolutional neural network with feature coding,which encode the convolutional features using differential coding to add the invariance to the final feature.To achieve better classification performance,we used our framework to extract feature of object region and concatenated it with feature extracted from the whole image.The concatenated representation not only describes global alignment of the whole image,but also outstands object area,which could achieve good classification performance.The experimental results demonstrate that our proposed feature extraction methods can obtain more representative feature and get better classification performance.