Surface Object Classification Of Optical Remote Sensing Image Based On Joint-Net

Remote sensing image classification has great research significance because it is the basic work of intelligent interpretation of remote sensing images.Among them,hyperspectral images are widely studied and used because of their rich spectral information,especially in the field of ground object classification.However,by observing a large amount of data,we found that the spectra of pixels in the same class of hyperspectral images have large variation.The intra-class difference caused by the interference of spectral variation and other factors is usually an important reason that hinders the improvement of the accuracy of hyperspectral image classification.In addition,Convolutional Neural Network（CNN）has shown significant advantages in the field of remote sensing image classification.Compared with earlier feature engineering methods,CNN can extract deeper semantic information of images,and achieving better classification results.However,the CNN-based method ignores the problems caused by large differences within the class,which makes it impossible to obtain more accurate classification results.Therefore,this dissertation focuses on following work:First,to solve the problem of large intra-class differences in hyperspectral images,we clustered the image and divided it into more sub-classes.Specifically,by using an automatic clustering method based on density peaks as preprocess module,the handcraft classes with larger intra-class difference in the image can be automatically clustered into smaller sub-classes,thereby a new ground truth map can be obtained to supervise the network and improve the discrimination ability of the network.Second,this dissertation designed a spatial-spectral joint network（JointNet）to extract the spatial and spectral information in the image and capture more valuable high and low-order features.The spatial-spectral joint network includes two branch networks.The first one is the Double Strip Convolutional Network（DS-CNN）,which is composed of double directional strip convolution kernels and used to extract spatial features.In addition,the Sinc Function Network（SincNet）uses Sinc function filters as convolution kernels,which is introduced to extract image spectral information.Finally,the different features from two networks are merged through the fully connected layer,and then sent to the log-softmax classifier to classify the samples.This dissertation is based on the three hyperspectral image data sets of Houston,PaviaU and Xiong-An to complete the ground object classification experiment.The experimental results show that the network model proposed in this dissertation has better classification performance than other methods,especially for classes with large difference within the class.