Hyperspectral Image Clustering Based On Autoencoder Network

Hyperspectral images carry rich spectral information and are widely used in the field of earth observation.Their high-dimensional characteristics and nonlinear structure bring "dimension disaster" and linear inseparability problem to clustering tasks.Previous work separated feature extraction process from clustering process,making it difficult to optimize them simultaneously.In order to solve the above problems and extract more effective deep features,this paper jointly optimized the feature extraction process and clustering process of hyperspectral image clustering task,and proposed a new joint learning depth self-encoder network fuzzy C-mean clustering method(CDFCC).First of all,using the depth of the encoder network dimensionality of strong nonlinear transformation ability the hyperspectral image raw data is mapped to the potential in the feature space,complete hyperspectral image spectrum feature extraction process,depth of the encoder network by minimizing reconstruction error learning each pixel of hyperspectral image characteristics to achieve the purpose of keeping local structure.Secondly,in order to cluster hyperspectral images better,a clustering algorithm is constructed on the hidden layer of the deep self-encoder network,that is,the feature extraction process is constrained by the fuzzy C-means clustering algorithm.Finally,in order to adaptive learn the spectral features suitable for clustering algorithm,CDFCC algorithm model is trained by combining reconstruction error and clustering error to obtain more effective deep spectral features of hyperspectral images.CDFCC algorithm has the advantage of joint optimization of two tasks.It can adaptively learn the deep spectral features of hyperspectral images suitable for clustering tasks and dynamically adjust the clustering indication matrix.In the experimental part,four hyperspectral data sets of Indian Pines,Pavia University,Salinas and Kennedy Space Center were used to verify the effectiveness of CDFCC algorithm.The clustering accuracy of CDFCC algorithm on the four data sets reached 42.95%,60.59%,66.29% and 62.64%,respectively.The experimental results show that CDFCC algorithm can extract more effective deep features from the high-dimensional spectral information of hyperspectral images,improve the clustering accuracy,and greatly improve the clustering efficiency because CDFCC algorithm does not need additional training process.This paper has 26 Figureures,11 tables and 83 references.