Remote Sensing Image Classification With Deep Contourlet Convolutional Neural Network

This thesis focuses on investigating the remote sensing image classification task based on the deep Contourlet convolutional neural network.Deep convolutional neural network is a kind of feedforward neural network which can process raw image patches,thus will consider the spatial correlations in pixels.This characteristic could reduce the effects of speckle noise and improve the final classification result further.In order to use the phase information,we extend the deep convolutional neural network in complex domain.This proposed complex neural network is named complex CNN.The Contourlet transform can capture the intrinsic geometric structure of images,therefore the discriminative features are obtained,as its filter banks allow the image to be approximated by several directional subbands from a coarse version to fine resolution.Introducing the conception of Contourlet transform to complex CNN is an effective solution for the classification of remote sensing images in complicated background.Above all,the main contributions of this thesis are listed as follows:1.A multiscale deep learning network named nonsubsampled Contourlet transform-based convolutional neural network is proposed.This network is designed to be used for remote sensing image classification task.It is constructed by nonsubsampled Contourlet transform,based upon the deep convolutional neural network.By replacing filters of the first convolutional layer with the filter banks of the nonsubsampled Contourlet transform,features with multidirection,multiscale,and multiresolution properties are obtained.Our proposed model achieves a better classification performance in remote sensing images than that of deep convolutional neural network,especially as the proportion of training data is small.2.Based on Contourlet-CNN,a remote sensing image classification method is proposed.The multiscale deep Contourlet filter banks are constructed by combining Gabor filters and the filter banks of nonsubsampled Contourlet transform,then be used to replace the filters of the first convolutional layer.Finally,this new proposed model named ‘Contourlet-CNN' is constructed.It is a double-channel fusion frame.In comparison with deep convolutional neural network,Contourlet-CNN is able to get more robust discriminative features and can efficiently adjust parameters.3.For the remote sensing image classification task,a complex convolutional neural network named complex CNN is proposed.Based on the deep convolutional neural network,we redefine the operation rules of convolutional layer,subsampling layer,normalization layer and fully-connected layer in complex domain.The complex Contourlet-CNN is formed by replacing filters of the first complex convolutional layer with the multiscale deep Contourlet filter banks.This proposed network could extract the amplitude features and phase features with multidirection,multiscale,and multiresolution properties,thus can effectively improve the classification accuracy.