Spectral-spatial Classification Of Hyperspectral Image Based On Stack Sparse Autoencoder Network

Hyperspectral remote sensing technology has a wide range of applications and plays an important role in many applications,such as precision agriculture、environmental monitoring and military reconnaissance.When applying hyperspectral data,classification is the premise of image understanding and interpretation.With the development of remote sensing technology,spectral resolution and spatial resolution are continuously refined,so that the acquired hyperspectral image contains rich spectral and spatial information,which brings opportunity and challenge for hyperspectral image classification at the same time.The main issues are:the data presents complex nonlinear characteristics;lack of labeled samples;noise interference;"same class has different spectral information" and "different classes have same spectral information".In general,the traditional spectral information-based classification algorithms ignore the cooperative effect of spatial information,and cannot obtain high classification accuracy when solving these problems simultaneously.Additionally,compared with the traditional classification algorithms,deep learning network has great potential in dealing with nonlinear data and mining deep information,which has been widely used in the field of hyperspectral image classification.Therefore,the purpose of this paperis to reduce spectral variability、adaptively extract high-level features of data and mine spatial domain features that promote classification task,different spectral-spatial information-based classification algorithms based on stack sparse autoencoder are proposed.The main research contents of the thesis are as follows:Firstly,the traditional shallow learning models cannot adaptively extract more abstract and useful high-level features of hyperspectral data in performing the classification task,and thus a hyperspectral image classification algorithm based on stack sparse autoencoder and random forest classifier(SSARF)is proposed.The algorithm uses stack sparse autoencoder(SSA)to adaptively extract abstract and useful high-level features from the original data,and then random forest(RF)classifier is used to fine tune the whole network and classify the acquired high-level features.This paper introduces RF into the SSA model,and SSA provides high-level features by characterizing the intrinsic property of the data,at the same time,RF can compromise the generalization ability,prediction accuracy and operation speed of the model.Experiments show that the proposed SSARF algorithm has the ability to obtain higher classification accuracy than some traditional shallow learning models.Secondly,the traditional spectral information-based classification algorithms do not consider the spatial resolution has two-sidedness on the classification accuracy of hyperspectral data.In fact,the classification accuracy of hyperspectral images depends on the net effect between spectral variability within the class and the number of mixed pixels at the edge.To solve this problem,a spectral-spatial information-based classification algorithm based on scaling up and SSARF is proposed(SU-SSARF).The algorithm first uses the cubic convolution method to obtain hyperspectral images with different spatial resolutions;then,the class separability criteria is adopted to select the most suitable spatial resolution image,and the spatial-spectral features of the each pixel are cascaded;finally,the cascaded features are sent to the SSARF model for learning and classifying.Compared to the initial hyperspectral image,the class separability becomes stronger among most of the land cover classes in the most suitable spatial resolution image,and thus the phenomenon of "same class has different spectral information" and "different classes have same spectral information" is improved to some extent,thereby improving the classification accuracy of the algorithm.Thirdly,compared to the spectral information-based classification algorithm,the proposed SU-SSARF achieves higher classification accuracy by considering the spectral variability of pixels under a fixed window,but the algorithm cannot further integrate the geometric proximity of pixels in a neighborhood window and cannot deal with the noise interference phenomenon exist in the image.To solve this problem,a spectral-spatial information-based classification algorithm based on joint bilateral filtering(JBF)and SSARF is proposed(JBF-SSARF).The algorithm first uses JBF to attenuate the image noise of the damaged image while extracting the spectral-spatial features with strong separability,and then SSARF is used to learn the high-level features and perform the classification task.Experiments show that JBF-SSARF obtains higher classification accuracy than some current mainstream classification algorithms by using only single-scale edge-preserving filtering features.In addition,comprehensive consider the positive impact of multi-scale filtering features in classification task,a spectral-spatial information-based classification algorithm based on multi-scale adaptive guided filtering(MAGF)and SSARF is proposed(MAGF-SSARF)at the same time.Experiments show that the collaborative effect of multi-scale filtering features facilitates the classier obtains higher classification accuracy.Finally,the backpropagation(BP)algorithm is adoped to optimize the objective function of a sparse autoencoder(SA),and the procedure exists some shortcomings such as the convergence speed is slow and it is easy to fall into the local extremum.In addition,some existing spectral-spatial information-based classification algorithms do not consider the spatial structure information of unlabeled samples.To solve these problems,a spectral-spatial information-based classification algorithm based on the improved SSARF and SUSAN is proposed(ISSARF-SUSAN).The algorithm first introduces the improved artificial fish swarm algorithm(IAFSA)into the SSARF model,that is,IAFSA is used to solve the shortcomings of BP algorithm in optimizing the objective function of SA;then,the improved SSARF model is used to perform high-level feature learning and obtain the classification map;in addition,a post-processing step is used to integrate the spatial information of unlabeled samples into the initial classification map,and the purpose is to reduce some misclassified sample points in the initial classification map.At last,to verify the proposed ISSARF-SUSAN algorithm has strong scalable,a spectral-spatial information-based classification algorithm based on multi-strategy fusion mechanism and ISSARF is proposed(MSF-ISSARF).Experiments show that the proposed MSF-ISSARF algorithm obtains highest classification accuracy compared to some current mainstream classification algorithms when the training samples are limited or sufficient.