Research On Spectral-Spatial Clustering And Classification For Hyperspectral Image

Hyperspectral remote sensing technology can simultaneously image the target region through hundreds of continuously subdivided bands and obtain 3-D hyperspectral images containing abundant spectral and spatial information.It has been widely used in many areas such as astronomical exploration,military target detection,geological exploration.Hyperspectral image has high dimensionality and redundant information,while subspace clustering refers to clustering in different subspaces under the same data set,which can effectively reduce redundant information.On the other hand,due to the large amount of hyperspectral data,limited training samples;it is hard to select an appropriate dimensionality reduce method for complex hyperspectral images,thus classification algorithms are facing challenges.Recently,image classification based on deep learning has progressed both in methods and performance.The main research work and innovations of this thesis are listed as follows:Firstly,this thesis proposes a new algorithm called Spectral-Spatial Sparse Subspace Clustering Based on Three-dimensional Edge-preserving Filtering(SSC-3DEPF).According to the special structure of hyperspectral images,a 3-D edge preserving filter is constructed.In the sparse representation of the original hyperspectral image,we constrain the representation coefficients by three-dimensional edge-preserving filtering,which integrates relatively stable spatial information and edge(intensity)information into the sparse subspace clustering process to obtain a more precise representation coefficient matrix.The filtering operation realizes local optimization of the sparse representation coefficient matrix and better maintain the intrinsic structure of the original data.This paper optimize the objective function by using the alternating direction method of multipliers(ADMM)optimization framework.Experiments demonstrate that compared with the existing clustering methods,SSC-3DEPF has higher clustering accuracy on three typical hyperspectral datasets.Secondly,a novel classification method called Spectral-Spatial Pseudo-3D Dense Network(SSP3DNet)is proposed for hyperspectral image.Firstly,we implemented a simple data augmentation algorithm,including edge mirroring and adding gaussian noise,to alleviate the problem of limited training samples.SSP3 DNet use pseudo-three-dimensional convolution to learn spectral and spatial features economically and effectively,so that the weight of the network is sparse,the computational cost is reduced,and the convergence speed is faster.Densely-connected structures can alleviate the gradient disappearance phenomenon caused by deep network and thus constructs a deeper network,the receptive field is increased to obtain the global features.Experiments demonstrate that the proposed method can use deeper spectral-spatial features to obtain higher classification accuracy and smoother classification maps.Edge region pixels can be better divided.Compared with other algorithms,the training process achieves the best accuracy within 80 epochs which reduce the computational cost and memory demand.