Research On Hyperspectral Image Classification Based On Deep Learning

Hyperspectral image classification is one of the most important tasks in the analysis of hyperspectral data.Accurate classification results are the prerequisite for the wider application of hyperspectral remote sensing.In recent years,deep learning has made great success in the field of computer vision due to its powerful feature representation ability.Applying deep learning to hyperspectral image classification problems is a hot research topic in remote sensing domain.However,due to the imbalance between dimensions of hyperspectral data and the number of available training samples,deep learning networks often face overfitting problems.The building blocks inherent in various deep learning networks limit their joint utilization and learning of spectral and spatial information.In view of the above problems,two hyperspectral image classification methods based on deep learning are proposed in this paper combining with the characteristics of hyperspectral data.The details are as follows:A classification method based on spectral angular distance weighted fusion and deformable Convolutional Neural Network(CNN)is proposed.Firstly,the low-level spatial features such as edges are extracted by using hierarchical guidance filter.Then,the cosine angle distance between pixels is calculated to to determine the weight of feature fusion.Finally,a deep deformable convolutional neural network is constructed to abstract the fusion features and implement end-to-end output.This algorithm considers the spectral correlation in the process of acquiring the joint spatial-spectral feature,which is ignored by the two-dimensional convolution.In addition,a convolutional neural network with the deformable structure is used to adaptively summarize various transformations of scene imaging in hyperspectral images.A classification method of Stacked Sparse Auto-Encoder(SSAE)network with graph regularization based on multi-scale shape is proposed.Firstly,a superpixel segmentation method is used to construct uniform homogeneous regions of different shapes.Next,the covariance representation of the shape adaptive region block is calculated for each pixel.Then,the obtained joint spatial-spectral features are fed into SSAE network,and the preliminary classification results are obtained through high-level feature abstraction.Finally,the network output are modified by message transfer under graph mapping.The algorithm solves the limitation of fixed window input.By using multi-scale strategy,the samples are expanded and the problem of small samples in hyperspectral images can be solved.In addition,covariance description operator can utilize spectrum correlation information which is not considered by other feature extraction methods.Finally,the graph-based spatial regularization method corrects the classification result output from the SSAE network,which makes up for the defect that the SSAE network completely loses the spatial information during feature extraction and classification.Experiments show that two proposed algorithms above both have good classification performance.