Hyperspectral Remote Sensing Images Classification Based On Sparse Auto-encoder Network

Hyperspectral remote sensing technology uses sensors to survey surface features,the information of the ground objects can be obtained by different target objects reflection ability in electromagnetic waves and light.Remote sensing is play a crucial role in the fields of object detection,medical,military,geological exploration,etc.Hyperspectral image represents the categories of ground objects accurately as it has abundant spatial and spectral information.At the same time,it also has lots of problems,for example,“dimensional disasters”,obtained difficultly label samples,small target recognition and insufficient utilization of spatial features.In order to improve the classification accuracy of hyperspectral remote sensing image,in this thesis,two kinds of hyperspectral image classification algorithms based on sparse auto-encoder network are presented in combination with spectral and spatial information.Traditional hyperspectral image classification approaches can’t effectively use spatial features and solve the question of dimensional disasters.To solve the problems,a TSNE and multiscale sparse auto-encoder neural network hyperspectral image classification algorithm is proposed.T-distribution stochastic neighborhoods have great advantages in visualization and dimensionality reduction.On the basis of dimensionality reduction,the target information can be retained effectively.T-distributed stochastic neighborhood embedding is reduced the hyperspectral image dimension,solve effectively the “dimensional disaster” problem,reduce the calculation complexity and redundancy.Each pixel’s multiscale spatial information is extracted.Using the spatial spectrum joint information trains a sparse auto-encoder network and classifies target by softmax classifier.In terms of the problem the difficulty in obtaining training samples and identifying tiny samples,a hyperspectral image classification algorithm based on multiple features and improved stack sparse auto encoder network is proposed.Manifold learning is used to obtain low-dimensional data structures of hyperspectral images.On this basis of dimensionality reduction,the circular local binary pattern features,extend multi attribute profiles features of the image are extracted.The spectral features,circular local binary pattern features and extend multi attribute profiles features are multiple fused.Then the samples of the fusion spatial spectrum joint information are used to train the stack active sparse auto-encoder neural network and classified by the softmax classifier.In view of the two algorithms are proposed in this paper,two hyperspectral remote sensing data from Indian Pines and Pavia U images are simulated.The overall classification accuracy,Kappa index and average accuracy are used to evaluate the experimental results and compared with the other eight algorithms.The conclusions indicated the effectiveness of the two algorithms method in this paper,classification accuracy,Kappa index and average accuracy of hyperspectral images has been increased.