Hyperspectral Remote Sensing Image Classification Based On Deep Learning

Hyperspectral remote sensing images are very narrow continuous spectral images formed by electromagnetic wave imaging.The resolution can reach nanometer range,and the imaging data contains rich information.In recent years,the hyperspectral remote sensing image has received the attention of various countries in the world due to the uniqueness of its own contained rich information.However,there is a strong correlation between hyperspectral bands,and the characteristics of spectral mixing and high dimensions increase the difficulty of classification.In this paper,through in-depth analysis of hyperspectral remote sensing data,based on deep learning networks,the feature extraction and classification algorithms of hyperspectral remote sensing images are improved.The main research is as follows:1.Combining stack sparse auto-encoding networks in deep learning networks to improve the extraction of spatial information,first using PCA to reduce the dimensionality of image data,the first principal component difference between different categories after dimensionality reduction is large,These differences sort,delete,reorganize,and stack the neighborhood information of the central cell.Finally,the obtained space-spectrum information is input to a stack sparse auto-encoding network combined with the SoftMax classifier for classification,and a good classification effect is obtained.2.Based on the current situation that hyperspectral remote sensing tag data is difficult to obtain,a secondary classification algorithm based on non-local information feature fusion is proposed.First,reduce the dimensionality of the spectral data,and find a certain amount of pixel information most similar to the central pixel through the spectral angle matching algorithm,and then form a new spectral vector with the original central spectral stack of these pixels and input directly into the SoftMax classification In order to improve the accuracy of the imported samples,a certain threshold is set for the result of the classification,and a new training data set composed of the qualified spectrum and the original training samples is input into the stack sparse automatic encoding machine.Training and classification,and finally further use spatial information to modify the classification results,so that the classification results tend to be smoother.3.In Hyperspectral image classification process,in order to more fully take advantage of its spatial information,hyperspectral imagery proposed classification algorithm based on feature fusion.First use the principal component analysis algorithm and the local linear embedding algorithm to reduce the dimensionality of hyperspectral remote sensing data in different ways,and then use the Gabor filter to extract the information from the dimensionalized data,and stack the extracted information with the spectral information,Fusion forms a new vector and inputs it to a stack sparse auto-encoding network for classification,and a good classification effect is obtained.By the result of the classification algorithm improved algorithm compared with other results demonstrate the feasibility and validity of the proposed algorithm.