Hyperspectral Image Clustering Algorithms And Applications Based On Generative Adversarial Network

Hyperspectral remote sensing is one of the most attractive topics in remote sensing area and it has been applied widely on agriculture,geology,grassland monitoring,forest research,marine research,environmental monitoring and disaster reduction.Classification is an important problem for hyperspectral image processing.However,because of the high dimensionality of hyperspectral images and expensive cost of artificial labelling,the number of training samples with labels is pretty small,and the classification accuracy is often low as“Hughes”phenomenon shows.Clustering needs no label information from hyperspectral images and hence becomes more and more important as the preprocessing step of classification.Generative adversarial network（GAN）is one of the newest and hottest generative models.It utilizes game theory and trains two networks simultaneously which are generator and discriminator respectively.The generator tries to simulate the data distribution of training samples and fool the discriminator so that the discriminator cannot recognize the generated samples.In the meantime,the discriminator tries to maximize its ability of distinguishing real samples from generated ones.In this thesis,from the view of data augmentation and feature extraction,several novel clustering algorithms for hyperspectral images were proposed which utilized the ability of generating fake samples and the ability of feature extraction from generative adversarial network.The main contributions of this thesis are as follows.（1）The pros and cons of traditional clustering algorithms（including K-Means and its variants,hierarchical clustering,density-based clustering and so on）were analyzed.Five traditional clustering algorithms were implemented using the C++programming language and their performances were evaluated on the Indian Pines AVIRIS dataset.The experiment results showed that the ISODATA algorithm（Iterative Self-Organizing Data Analysis Technique）achieved high clustering accuracy and low time consumption,and thus it had the best comprehensive performance and practicality.（2）HSGAN^[85]was a GAN model especially for classification of 1-D spectral data.Researches were taken on clustering analysis of hyperspectral images from following three points based on the HSGAN model:firstly,from the viewpoint of data augmentation,the generator in the HSGAN was utilized to generate fake spectral samples,and the fake samples were added into the real dataset.The clustering algorithm was performed on the augmented dataset.Secondly,from the viewpoint of feature extraction,the discriminator was utilized to extract features and the clustering algorithm was carried on the extracted features.Thirdly,by combining data augmentation and feature extraction,the discriminator was utilized to extract features of generated fake samples and the clustering algorithm was performed on the augmented features which were joined features of real and fake samples.The experiment on the Indian Pines AVIRIS dataset showed that the above three methods improved clustering measurement of K-Means algorithm.The experiment results suggested that shallow features performed better in the second method and deep features had better performance in the third method.（3）Based on the above model,several architecture extensions of the HSGAN model were made and applied to clustering analysis of hyperspectral images.The first one was to substitute the original generator with the VAE（Variational Auto Encoder）model such that it generated a VAE/GAN model.The second one is to add another convolution block including two convolution layers and a maxpooling layer into the discriminator and to add a batch normalization layer after each convolution layer so as to make the training phase of GAN model more stable.The experiments on the Indian Pines AVIRIS dataset showed that the above two measures improved clustering performance of K-Means algorithm and in both the second and the third method,deep abstract features worked better than features of shallow convolution layers.（4）A clustering analysis software of hyperspectral images was developed.This software implemented reading,fake spectral data generation,spectral feature extraction,clustering and visualization functions for hyperspectral images.