Integrating Spatial Spectral Information For Refined Endmember Extraction Of Multispectral Image

Mixed pixel is an unavoidable phenomenon in remote sensing images due to the limitation of spatial resolution and the existing of boundary pixels,regardless the multispectral or hyperspectral images.Mixed pixel problem in multi/hyperspectral images have influenced the classification of image,target detection and hindered the development of quantitative remote sensing.One of the most effective method to solve the problem of mixed pixel is Spectral Mixture Analysis(SMA)or Pixel Unmixing.Models of Pixel unmixing mainly include Linear Spectral Mixing Model and Non Linear Spectral Mixing Model.This research is based on linear spectral mixing model.Endmember Extraction and Abundance Inversion are the core goals of SMA.Endmember extraction is the first and key step,which directly affects the accuracy of pixel unmixing.Due to the fine spectral resolution of hyperspectral image,which provides more specific spectral information to increase the differences between mixed pixel and pure pixels,most endmember extraction algorithms have the chance to develop.However,multispectral images contain limited spectral information,resulting in poor performance of existing endmember extraction methods when applied on multispectral images.The advantages of multispectral images such as high spatial resolution,wide coverage,continuous monitoring ability make them become widely used in resources and environment remote sensing.Therefore,it is of great research meaning and application value to improve or develop the endmember extraction methods which are suitable for multispectral images,especially for medium/high resolution images.Based on this,this research focuses on multispectral remote sensing images and aims at integrating spatial and spectral information for refined endmember extraction.The main content and work of the research list as fellow:(1)A panchromatic band assisted endmember extraction method is proposed.The spatial coefficient of variation(SCV)is constructed by using panchromatic band.Combining with Pixel Purity Index Algorithm,Spatial Pixel Purity Index(SPPI)is constructed through threshold segmentation and pixel weight value using SCV.A panchromatic band assisted endmember extraction method is proposed for multispectral images which contains panchromatic band such as Landsat 8 OLI.(2)An endmember extraction method integrating multiband spatial spectral information is proposed.Based on the spatial and spectral information of different strategies of bands combination,a method to integrate spatial coefficient of variation,interior pixel spectral angle and neighborhood spectral angle to express the purity of a pixel is proposed.At last,an endmember extraction method integrating multiband spatial and spectral information is applied to multispectral images which contains bands of different spatial resolution such as Sentinel 2A.(3)Validation and analysis using simulation and real images shows the refined performance of the methods mentioned above.The simulation data includes simulated Landsat 8 OLI and Sentinel 2A images.The real images are obtained from Landsat 8 OLI and Sentinel 2A.The validation result shows the improved performance of the methods mentioned above.The precision of endmember extraction is improved using the methods proposed in this paper.