Research On Hyperspectral Image Spectral–Spatial Classification Based On Deep Autoencoder Network

Remote sensing exploits satellite-borne or airborne sensors to detect ground target.As an active imaging technology,it can emit electromagnetic wave of different wave length.When compared to general optical imaging equips,remote sensing has the advantages of all-time.Due to the diversity of material reflectivity,remote sensing images can be used to detect the ground targets.In recent years,the spectral resolution of remote sensing equips becomes higher and higher,which makes the hyperspectral image(HSI)contains rich information of ground scenes.Although the improvement of resolution contributes to the increase of detection accuracy,too much data may cause difficulties for the algorithm: how to remove redundant information of the HSI,how to fully exploit the spatial texture information,how to improve the robustness of algorithm under small samples.Deep neural network exploits nonlinear transform to fit the data distribution.Compared to traditional shallow machine learning methods,it can learn more abstract feature automatically,without the need of manual effort,and has been used in speech recognition and automatic driving areas successfully.Traditional hyperspectral image classification methods can not make full use of the spatial information,and the classifica tion accuracy is not satisfactory.To solve the existing problems,this paper utilizes the advantage of multiscale transform on edge detection,and studies the feature extraction and small sample problems based on the properties of HSIs.The research results are as follows:1.Since traditional methods cannot extract the spatial information effectively,a hyperspectral image classification method based on gabor filter and autoencoder network is proposed.Principle component analysis is first adopted for dimension reduction,then gabor filter is used to extract spatial texture information.Finally,after combining the spectral and the spatial feature,the deep autoencoder network is employed to predict the label of all pixels.The experiment results on three hyperspectral datasetsets indicate that the proposed method has higher classification accuracy than traditional ones.2.For gabor filter is not effective in extracting the spatial structure information,a hyperspectral image classification method based on multiscale geometric transform and autoencoder network is proposed.Multiscale geometric transform is firstly employed to enhance the edges,and the statistical characteristics of local image patch is used to describe the spatial structure information.The input spectral-spatial feature is compressed by the stacked autoencoder,and the dimension reduced feature is mapped into another data domain.Finally,the multinominal logistic regression classifier is used to predict the class of the processed pixel.Compared with some existed methods,our proposed method is more effective in feature extraction,and achieves excel ent classification accuracy.3.For the labeled samples are limited in HSI classification task,a method based on the active learning strategy is studied.Despite the great representation capacities of deep neural network,the number of model parameters increased rapidly.As a result,more samples are needed in training the model.The active learning mechanic is used to solve the problem.Instead of choosing samples randomly,the active learning strategy selects training samples iteratively.In each iteration,by calculating the information entropy of all candidate samples,the samples which make greater contribution to the decision boundary can be determined.As is indicated in the experiments,compared to the random sampling,the active learning strategy can improve the classification accuracy of the proposed method.