Research On Hyperspectral Image Classification Based On Representation Learning

In recent years,with the rapid development of hyperspectral remote sensing technology,classification techniques based on hyperspectral images plays an extremely important role in target detection,environmental management and mineral mapping.These applications usually need to classify images in specific scenes.Some scholars have applied representation learning to hyperspectral image classification,but there are still some challenges and limitations in traditional hyperspectral image classification:(1)The dimension of hyperspectral image is much larger than multispectral image,while traditional representation learning technique is designed for multispectral image.The processing effect of hyperspectral image using traditional technology will be restricted to some extent.(2)There are obvious problems in hyperspectral image,such as high correlation between bands and limited samples.The traditional dictionary learning algorithm is used to process hyperspectral image,and the classification performance is affected by serious spatial homogeneity and heterogeneity.To improve the classification performance of hyperspectral image,this paper mainly conducts in-depth research from the following two aspects:(1)Directly using representation learning for hyperspectral image classification will suffer from a serious curse of dimensionality.In this paper,a spatial aware collaborative representation based on augmented spatial spectral features network is proposed.This algorithm first constructs a hierarchical network based on spatial spectral features according to the inherent low-dimensional manifold of hyperspectral image.Secondly,the trained network is used for feature extraction of high-dimensional data and spatial aware collaborative representation algorithm is used for classification.Experiments on two hyperspectral datasets Indian Pines and Pavia University prove the effectiveness of the proposed algorithm.(2)Directly using transform learning to select feature dictionary will be seriously affected by spatial homogeneity and heterogeneity.In this paper,a discriminative transform learning algorithm based on spatial spectral features is proposed.Firstly,the spatial features of hyperspectral image are extracted by using the local binary pattern,and the spatial spectral features are obtained by stacking.Secondly,extreme learning machine is fused in transform learning algorithm with regularization term.Finally,the discriminative transform matrix is used to recover the test samples and the hyperspectral images are classified by the extreme learning machine classification model.Traditional representation learning algorithms use training samples as "dictionary" for classification,and the proposed algorithm uses the learned optimal transform matrix as a "dictionary" for classification,which can make good use of and deal with the problems of high correlation between bands and limited samples in hyperspectral image.The experimental results show that the proposed algorithm can effectively avoid the influence of spatial homogeneity and heterogeneity,and the classification performance is improved.