Research On Hyperspectral Remote Sensing Image Classification Method Based On Convolutional Neural Network

Hyperspectral remote sensing images have been widely used in agriculture,military and many other fie lds because of their rich spectral and spatial information.Image classification is an important part of remote sensing image applications and lots of labeled data are needed to achieve good image classification performance.However,hyperspectral remote s ensing images has large spectral information dimension,and image labeling is a high cost and time-consuming work,which all affect the performance of image classification.This paper introduces deep learning theory and develops h yperspectral remote sensing image classification method to overcome the shortage problem of labeled training samples.The main work is summarized as follows:1.According to the characteristics of hyperspectral remote sensing images and different image preprocessing,two different image classification models based on convolution neural network are proposed,and experiment al results on different data sets show that the proposed methods can achieve good image classification performance.Meanwhile,this work also discusses the influence of the number of training samples and the size of sample space.2.Since deep network usually needs a large number of labeled samples to train the network,this paper integrates the idea of semi-supervised learning into deep learning framework,and develops a semi-supervised deep learning method for a hyperspectral remote sensing image classification.In which,t he labeled samples are used to provide the cluster center for unlabeled samples,and then the labeled samples and unlabeled samples are both used to train deep network,so that the feature extraction ability and classification performance of the network are greatly improved.Experimental results show that the proposed semi-supervised deep learning method is feasible and efficient.3.By introducing the idea of transfer learning into deep learning,this paper proposes a hyperspectral remote sensing image classification method based on deep sample transfer learning.It transfers the samples of other hyperspectral remote sensing images to the target im age to increase the number of training samples,so that it can improve the problem caused by the lack of labeled training data.Experimental results sho w that the proposed sample transfer learning method can effectively improve the classification performance of the target network and improve the classification accurac y of the image.