Study Of Hyperspectral Image Classification And Parameter Setting Based On Deep Learning

One of the most important parts of hyperspectral remote sensing theory and application research is accurate classification of target,and feature extraction is the key of classication.The traditional manual feature extraction methods are based on professional knowledge and rich experience,which is time-consuming and laborious,and it can only extract shallow features.Hyperspectral image provides richer spectral-spatial information and it's a three-dimensional tensor data which has a large number of spectral dimensions and large amount of data,but traditional feature extraction methods are difficult to fulley express the complex spectral-spatial characteristics of hyperspectral data.In recent years,as an important branch of machine learning,deep learning has attracted wide attention due to its strong analysis capabilities and feature extraction abbilities.The deep learning model can extract features of the input data from the bottom layer to the top level,and finally form the high-level abstract features suitable for the pattern classification,which help improve the accuracy of the target classification.Therefore,deep learning is introduced into hyperspectral image classification in this thesis to fully explore the spectral-spatial information and improve the classification accuracy by establishing deep learning models and stdying the parameter setting.The main reseach in thisthesis focuses on the following aspects:First,data preprocessing is carried out according to the data characteristics of hyperspectral image.Hyperspectral image is a three-dimensional tensor data,which adds one-dimensional spectral information to ordinary two-dimensional image.Due to the influence of water molecules and noise in the atmosphere,the information contained in some bands of hyperspectral images is very low.Therefore,these spectal bands can be removed according to the Frobenius norm of the image,and the hyperspectral data is preprocessed by reducing dimension.Then the hyperspectral image after dimensionality reduction is normalized,which lays a solid foundation for the subsequent target classification.Next,considering the spectral-spatial characteristics of hyperspectral image,different classification models of deep belief network and different two-dimensional convolution neural network are constructed based on spectral information,space characteristic and spectral-spatial information,respectively,and the classification results of different classification models are compared and analyzed.The experimental results show that the classification models based on spectral-spatial characteristics performance better.Then,the network performance is greatly influenced by the parameter settings of the deep learning classification model.For example,convolutional neural network has low network complexity and a relatively small number of weights owing to local connection and weight sharing,but its performance is still influenced by the parameter settings.To solve this problem,a classification method based on two-dimensional convolutional neural network with parameter tuning is proposed.The network parameters are tuned in turn according to the unique variable principle based on experimental results,and the optimal parameters are selected based on the overall accuracy of target classification.Furthermore,dropout is introduced into the process of parameter tuning to reduce overfitting.Simulation results show that the proposed two-dimensional convolutional neural network with parameter tuning method has considerable potential for hyperspectral image classification.Finally,spectral-spatial features extraction and the limited trainging samples are two difficult problems that must be faced in hyperspectral image classification.On the one hand,to extract high-level invariant features and improve classification performance,three-dimensional convolutional neural network is used to fully extract the spectral-spatial features of hyperspectral image.In this classification model,the three-dimensional hyperspectral data including the spectral-spatial characteristics can be directly input into the network,which greatly preserves the original spatial structure of the target and avoids the complex data reconstruction.On the other hand,to eliminate the problem of limited samples,virtual samples which are created from the existing samples are combined with the original samples to form training samples to achieve the full training of the network.Experimental results show that the proposed method perfomes better compared to other considered methods methods and has promising prospect in the field of hyperspectral image classification.