Research On Deep Learning-based Method For Hyperspectral Remote Sensing Imagery Classification

Hyperspectral imaging widely used in geological exploration,ecological protection,agricultural estimates and many other areas,it is an important part of remote sensing systems in many countries.Hyperspectral image classification refers to the process of recognizing the ground objects categories contained in hyperspectral image;classification is the prerequisite for understanding and interpreting hyperspectral image.There are problems like the high dimension data,a large amount of data,and a small number of marked samples in the hyperspectral image classification process;Effective combination of unsupervised learning and supervised learning is one of the ideal solutions.The deep belief networks is a deep learning model that includes unsupervised pre-training and supervised fine-tuning,it has the theoretical advantage of application on hyperspectral image classification tasks.At present,the application of the deep belief networks to hyperspectral image classification is in the initial stage,for the existing research examples,lacking of specific basis for the value of structural parameters those used in the deep belief networks,and lacking an efficient introduction method to introduce spatial information of hyperspectral images to classification.For hyperspectral image classification tasks,the specific work of this thesis is summarized below:(1)This thesis presents a new deep belief networks structured for hyperspectral image feature extraction and ground objects classification.We stack the restricted boltzmann machine to build a deep belief networks.The key to build a deep belief networks is to determine the parameters of its layers and the basic composition of model,including the layers of the restricted boltzmann machine,the number of neurons in hidden layer,the maxepoch of unsupervised training,and the layer learning rate.In this thesis,detailed comparison experiments are carried out to determine the optimal value of each parameter.We apply the built deep belief networks to the real hyperspectral image classification,results shows that deep belief networks proposed in this thesis has achieved higher classification accuracy than RBFN and SVM algorithm in the same running platform.(2)This thesis presents a hyperspectral image texture information enhancement model for the introduction of spatial information to hyperspectral image classification.To utilize the advantages of the hyperspectral image having so many bands,we use the Guided Image Filtering algorithm for band images to raise their edge information,and enhance the texture information of hyperspectral image.We combined the proposed texture information enhancement model and the proposed deep belief networks,and propose a hyperspectral image classification framework which including texture enhancement.The classification results of the proposed framework on true hyperspectral image shows that,the proposed framework improved the accuracy of hyperspectral image classification significantly.(3)The spatial filtering method is used to reduce the redundancy and denoising of the hyperspectral image to optimize the texture information model.The redundancy and noise may come from the process of texture enhancement.So as to realize the spatial information optimization of hyperspectral images after texture enhancing.The classification results shows that the optimizing of the hyperspectral image which treated by the proposed texture information model can further improve the classification accuracy,and optimize the classification performance.