Space-spectrum Joint Hyperspectral Image Classification Method And System Implementation Based On Nuclear Extreme Learning

Hyperspectral remote sensing technology can get image data whose spectral resolution is nanometer in the range of visible to short-wave infrared and even thermal infrared bands.It contains abundant spectral,radiation and spatial information.Hyperspectral remotesensing can be widely used in environmental monitoring,military reconnaissance,geological exploration and other fields.Hyperspectral image classification is the key step of hyperspectral image processing.aimed to assign a unique class label to all pixels to generate a land-cover thematic maps which can accurately reflect the distribution of the true land to helps researchers to discover and understand the laws and provide the basis for other tasks of hyperspectral image processing.However,there are several important challenges when performing hyperspectral image classification.For instances,unbalance between the limited training samples and high dimensionality,presence of mixed pixels,difficulty to integrate the spatial and spectral information to take advantage of the complementarities,quite complex geometry,many spectral bands and large computation time.In this paper,using Extreme Learning Machine(ELM)as the basic framework,we make use of its advantage in training speed,classification ability and generalization ability to perform the task about hyperspectral image classification.Based on analyzing and extracting spectral features and spatial features of hyperspectral images,we propose a joint spatial-spectral hyperspectral image classification model and algorithm based on KELM with joint spatial-spectral kernel and multiple features fusion.Then we design and realize a corresponding hyperspectral image classification and processing system.Experiments on real hyperspectral images data verify the validity of the proposed classification methods and system.The main tasks in this paper include:(1)We propose a joint spatial-spectral hyperspectral image classification based on neighboring filter kernel extreme learning machine.Based on the analysis of classical kernel methods such as Gaussian radial basis kernel,composite kernel which combines spatial and spectral information linearly,etc.,the paper employs joint spatial-spectral kernels to extract spatial and spectral information of hyperspectral image fully and proposes a joint spatial-spectral hyperspectral image classification method based on neighboring filter kernel extreme learning machine.Experiments on AVIRIS Indian Pines and ROSIS University of Pavia datasets show that this method is of high accuracy,especially in the case of small training samples.(2)We propose a spatial-spectral multiple features fusion method for hyperspectral image classification based on kernel extreme learning machine.In order to solve the problem about single feature on kernel extreme learning machine,we employ principal component analysis,local binary pattern,Gabor filter and extended multiattribute profiles to extract spectral feature,local texture feature,global texture feature and geometric structure feature,respectively.Then we use probabilistic kernel extreme learning machine based on softmax function to classify each feature separately and design a decision-level fusion classification method to integrate these classification results.Finally,the Markov Random Field(MRF)is employed to explore the spatial relationship among neighborhood pixels to further improve the classification accuracy.Experiments on real hyperspectral images show that this method is able to achieve robust,efficient results in hyperspectral image classification.Note that this method can also achieve good classification results with small training samples.(3)Based on above model and algorithms,we design and realize a hyperspectral remote sensing image classification and processing system based on Matlab GUI framework.The system consists of four modules,including hyperspectral image management and display,feature extraction,hyperspectral image classification,analysis and son on.We provide the system framework,main process design and software core module development and testing.