Hyperspectral Image Classification Based On PCANet

With the development of hyperspectral remote sensing imaging technology,various hyperspectral remote sensing processing problems have become a research hotspot in the field of remote sensing,and classification of hyperspectral images is one of the most important problems.Unlike natural or multispectral images,hyperspectral images have richer spectral and spatial information that can be used for classification,but the rich information is a "double-edged sword",which can not only help the classification of hyperspectral images,but also cause the information overload and redundancy.Therefore,how to make the full use of this information to improve image classification results has become a research hotspot in the field of hyperspectral images.At the same time,when using deep learning to obtain excellent hyperspectral image classification results,how to save time and cost by simplifying network complexity is also a problem that scholars must consider in the research process.Principle Component Analysis Network(PCANet)is a simplified deep learning model and can extract richer depth information from hyperspectral images.In order to study how to make full use of spectral and spatial information and use the PCANet to obtain good results.This thesis proposes several hyperspectral image classification methods based on PCANet.The main work is as follows:(1)A hyperspectral image classification method based on spatial coordinates and PCANet is proposed.This method uses spatial coordinates as low-dimensional features,and uses PCANet to further extract depth features from low-dimensional features,which makes full use of the features of hyperspectral images.The method adopts the Rolling Guidance Filter to solve the problem that the spatial coordinates are insufficient for the classification of the feature types with a small sample size or scattered sample distribution.Compared with other existing method,this method can obtain better classification results.For the feature types with small sample size or scattered sample distribution,the classification result of this method is also very ideal.(2)A hyperspectral image classification method based on multi-feature fusion and PCANet is proposed.This method uses both spatial coordinate features and extended morphological features to make full use of the spatial and spectral information of hyperspectral images.Different from the conventional direct stacking or stitching of multiple features,this method designs a branch-and-merge structure of PCANet,which effectively solves the problem of mutual interference between different features.The depth features finally extracted are of great help for the classification of hyperspectral images.The experimental results on two common hyperspectral data sets fully verify that the method can effectively fuse different features and obtain very good classification results.(3)An efficient spatial spectrum feature fusion method based on PCANet and iterative SVM for hyperspectral image classification is proposed.In order to make full and effective use of spatial spectrum information,for spectral information,the method directly uses the Support Vector Machine(SVM)to analyze the original spectrum.The computational complexity is very low.For spatial features,the method further extracts deep features through PCANet from extended morphology.The combination of sufficient deep features and simple spectral features effectively balances the computational complexity of spatial-spectrum feature extraction.In addition,the method borrows the idea of ensemble learning,and uses the iterative SVM as a learner to effectively integrate the spectral classification results with the spatial classification results.Experiments on two hyperspectral data sets verify that this method fully learns the spatial spectrum information of hyperspectral images by ensemble learning.And this method has a significant advantage in classification accuracy.