Hyperspectral Image Classification Based On Deep Learning

In recent years,the classification HSI has gradually become an important subject in the field of remote sensing,and has been applied more and more in the field of commerce and defense.Many researchers apply the latest technologies,such as deep learning,to the field of HSI.Compared with traditional multispectral data,hyperspectral data a large number of bands,and each band has strong correlation,and a stronger ability to distinguish features.However,the traditional classification methods are often faced with the problems of poor generalization ability,large amount of calculation,time consuming and difficulty in classification accuracy when extracting feature information from remote sensing images.In addition,conventional remote sensing data are based on small samples,and it is difficult,expensive and time-consuming to collect real feature tags.At the same time,using a small number of tagged training samples for classification also faces many challenges.In recent years,the deep learning based on neural network has achieved great success in extracting the further features of the data.Therefore,in order to improve the above mentioned problems,this thesis focuses on feature extraction of hyperspectral remote sensing images combined with depth learning.The major completion of work in the thesis is as follows:(1)A model based on CNN structure is studied,and the model is running on the hyperspectral remote sensing image dataset of high-dimensional small sample.It's showed based on the experimental results that compared with traditional learning methods,depth learning has advantages in high-dimensional and small-sample data.(2)In this thesis,a HSI classification algorithm derived from the GAN is proposed.Unlike the general supervisory algorithms,the HSI classification algorithm we proposed is based on semi-supervised learning,which can make full use of a limited number of labeled samples and a large number of unlabeled samples and the core idea of the algorithm is twofold.First of all,spatial and spectral(space-spectrum)features are extracted by using a 3DBF,which naturally treats HSI as a cubic dataset.Spatial information is integrated into the extracted features by 3DBF,which is beneficial to the subsequent image classification.In addition,in semi-supervised learning,the integrated space-spectrum features are used to train GAN.GAN consists of two competing neural networks(generating networks and discriminant networks).Semi-supervised learning is achieved by adding samples from the generating network to the feature and increasing the dimension of the classifier output.The experimental results on two data sets show that the proposed method is effective.