Application Of Deep Learning In Hyperspectral Image Classification

Hyperspectral remote sensing has been developed rapidly in recent years and has been applied in many fields,which gradually becomes an important method to quantitative research of the objects.Because of its high spectral resolution and the characteristics of the pattern of unity,in the surface of the fine identification of materials and classification has an unparalleled advantage.Therefore,One of the main part of the study in hyperspectral remote sensing image hyperspectral remote sensing image classification.The high-dimensional feature of hyperspectral image also brings difficulties to image classification,that is,for high-dimensional feature data,a large number of labeled training samples are needed to train the classification model to ensure the accuracy of the classifier,that is,the occurrence of the Hughes phenomenon.Due to the Hughes phenomenon,the dimensionality of the hyperspectral image data is usually widely applied to the hyperspectral image classification by the dimension reduction algorithm,and the Hughes problem can be solved better by the dimension reduction method.However,the commonly used dimensionality reduction method is often limited to the shallow feature of the extracted pixel,which can not get the deep feature of the pixel,which limits the performance of the classifier,and the robust deep feature often contains the abstract of the pixel Structural information,which is more conducive to the improvement of classification accuracy.In this paper,the deep learning is applied to the hyperspectral image classification,try to extract the more favorable pixel deep features,and analyze the advantages and characteristics of the three commonly used deep learning algorithms,focusing on the unsupervised learning algorithm stacked auto-encoder and the deep belief network is applied to the feature extraction of hyperspectral image pixels to solve the Hughes problem.The research process is as follows: Firstly,the nonlinear characteristics of the pixels in the hyperspectral image and the occurrence of the Hughes phenomenon in the hyperspectral image are proved to demonstrate the applicability of the deep learning in the hyperspectral image classification.Secondly,by comparing with the traditional dimensionality reduction method,The results show that the performance ofthe auto-encoder and the restricted bozmann machine in the deep learning theory is superior to that of the traditional algorithm.Then,the method is used to analyze the parameters,such as parameter optimization,visual analysis and precision analysis under different classifiers.The optimal classification model of the stacked auto-encoder and the deep belief network is obtained.Finally,the stacked auto-encoder with better classification performance is further optimized,that is,adding the sparsely constrained restriction condition and introducing GPU parallel computing,to further enhance the classification accuracy and classification speed.In this paper,the deep learning theory is introduced into the hyperspectral image classification,through unsupervised learning method can use a large number of unlabeled data,and can extract the deep characteristics of the pixel.Experiments show that the deep learning algorithm is superior to the traditional feature extraction algorithm and obtains the optimal classification model based on the deep learning,that is,the stacked sparse auto-encoder,the classification accuracy can reach 93.41%and 94.92% under the two experimental data.Model training time is long,the use of parallel computing method can make the model training speed increased more than 7 times.