Research On Hyperspectral Image Classification Based On Rolling Guidance Filtering

In recent years,the rapid progress of imaging remote sensing technology has promoted the vigorous development of hyperspectral remote sensing field.The hyperspectral remote sensing data not only has high spectral resolution and wide spectral range,but also has many bands and strong correlation between bands.This makes hyperspectral remote sensing data packages contain abundant terrain information,which makes it possible for hyperspectral image classification to be applied in many fields.However,the training sample data set used for classification in actual hyperspectral image classification is usually difficult to obtain,and the dimension disaster problem affects the accuracy and effectiveness of image classification.In addition,the spectrum of hyperspectral remote sensing data itself varies greatly,and the phenomenon of "homology and heteroscope" exists universally.If the spectral features of original objects in hyperspectral images are used to classify,the accuracy is low and there is "salt and pepper phenomenon" in the classification result map.In order to obtain better classification results,it is necessary to make full use of spectral and spatial information in hyperspectral remote sensing images,reduce the spectral variation differences within classes,and expand the spectral variation differences among different ground objects.The main research work of this paper is as follows:(1)In order to reduce the noise of interior texture and further improve the classification effect,a spectral-spatial classification method for hyperspectral images based on rolling guidance filtering is proposed in this paper.Firstly,the principal component analysis algorithm is used to reduce the dimension of hyperspectral remote sensing data.Then,the dimension-reduced image is blurred by Gauss filter,which is used to reduce the noise of small-scale structure in the image;Then,the blurred image is used as the initial guiding image,and the reduced-dimensional image is recursively filtered to maintain the edge structure of the ground object,and the output result is taken as a new guiding image,and the process is repeated until the large-scale edge structure in the image is restored.Finally,the feature information obtained is used to classify the objects in the image using support vector machine.The experiment is carried out on real hyperspectral remote sensing images.The results show that the method can obtain higher classification accuracy when the training samples are rare.(2)The purpose of edge-preserving filtering is to smooth the image while preserving the edge structure.Due to the complex spatial structure in real hyperspectral remote sensing images,traditional feature extraction based on edge-preserving filtering relies on gradients in the image,and many visually prominent edges do not correspond to image gradients.In order to make full use of edge information in the process of feature extraction of hyperspectral image data,this paper presents a method of hyperspectral image object classification based on semantic information and edge-preserving filtering.Firstly,the dimension reduction processing is performed on the hyperspectral data,and then the scale sensing based rolling guidance recursive filtering and the confidence of the edge of the image pixels obtained by edge detection are fused to obtain the useful feature information for the classification of hyperspectral images.Finally,the feature information is extracted using the support vector for the feature extraction information.The experimental results show that this method can also get good classification results when the number of training samples is scarce.