Research On Hyperspectral Image Classification Based On Extreme Learning Machine

A hyperspectral remote sensing image is a collection of pixels representing a given scene or object.It has hundreds of spectral bands and contains rich data information.The biggest advantage of hyperspectral images is the high spectral resolution,but it also has the disadvantage of high feature dimensions.With the research and development of image processing technology,how to extract large amounts of information from hyperspectral remote sensing images and successfully apply them to classification tasks is a huge challenge in the field of remote sensing images.At present,many machine learning methods have an important position in the research and development of many classification methods.The Extreme Learning Machine(ELM)is frequently used to process hyperspectral image classification tasks with its simple,fast,and good generalization ability.Although ELM can perform well in many studies,there are still some problems that need to be improved.For example,the data has noise,insufficient learning,lack of spatial information extraction,and insufficient spatial information extraction.This article will further research and innovate the ELM-based hyperspectral image classification method for the above problems.The results are as follows:1)In the learning process of hyperspectral remote sensing images,there is a certain amount of noise in the hyperspectral image data,and the extreme learning machine lacks the effective extraction and combination of spatial and spectral information,and cannot provide good classification results in classification.In order to solve this problem,by introducing a weighted spatial spectral algorithm,a weighted spatial spectral locality information preserving extreme learning machine(WSSLPKELM)is proposed.The algorithm obtains hyperspectral images with good spatial continuity by weighting the spatial information and spectral information,which effectively suppresses the influence of pepper and salt noise and improves the classification accuracy of the image.In this paper,two sets of commonly used hyperspectral data sets,Indian Pines and University of Pavia,are used as comparative experiments.By comparing with different algorithms,the results of the proposed algorithm show that it has more classification advantages.2)Some pixel classifiers used in hyperspectral image classification tasks only use spectral information,without taking into account the rich spatial information of the image or the insufficient extraction of spatial information.In response to this shortcoming,a new localbinary patterns information preserving extreme learning machine(LLPKELM)algorithm is proposed.It makes full use of HSI's rich texture information,and mainly uses local binary patterns(LBP)to better extract the local features of hyperspectral images,such as edges,corner positions,and spots,in order to better improve the classification effect.This article chooses to conduct experiments on two sets of commonly used hyperspectral data sets.By comparing with the traditional SVM classification algorithm and the currently popular KELM,KCRT-CK,MLR and LPKELM algorithms,the results prove that the proposed algorithm has a good classification effect.