Classification Of Hyperspectral Remote Sensing Image Based On Multi-classifier Fusion

Hyperspectral remote sensing image classification has always been a popular research direction in the field of remote sensing.Hyperspectral remote sensing continuously images the target area within a certain range of wavelengths.The image contains rich spatial information and spectral information,which can be effectively used for classification and identification of the ground objects.Hyperspectral remote sensing image data has a large amount of data,a large number of bands,and a high feature dimension of the pixel,which brings sufficient conditions to the pixel classification,and also brings about an increase in information redundancy,a reduction in signal-to-noise ratio,and difficulty in improving the efficiency of the classification model,problem.If only use the spectral information and ignore the spatial information,which will cause the problem of insufficient representation of the pixel.For the above problems of hyperspectral remote sensing image pixel classification,the main work of this thesis is summarized as follows:In terms of data preprocessing,firstly,this thesis uses the feature importance of GBDT to select the band of the original image data,effectively removes a large number of redundant bands and improves classification efficiency..Secondly,combined with the characteristics of hyperspectral image,this thesis integrates the correlation coefficient into the bilateral filtering algorithm to filter the image.It achieves a good denoising effect and makes the features of each category more concentrated without affecting the image texture,effectively improve the classification accuracy.About the feature extraction,this thesis improves the LPP dimension reduction algorithm,which use of both spatial dimension information and prior knowledge of category of hyperspectral remote sensing images,and achieves better results than classification after dimension reduced by PCA and LDA.This thesis also defines a LBP operator to extract texture information from hyperspectral remote sensing images,which can effectively extract spatial dimension features of hyperspectral remote sensing images.In the aspect of model selection and optimization,this thesis uses LightGBM model which has advantages in accuracy and efficiency in classification tasks as the basic classification model in the research process.Finally,using Stacking fusion strategy,fuses the three models of RBF_SVM,LightGBM and Random Forest through Logistic Regression,which achieves better classification effect than single models.