Research Of N-FINDR Improved Endmember Extraction Algorithm With Integration Of PSO

The development of hyperspectral remote sensing technology begins in the1980s.With the progress of imaging technology,Hyperspectral remote sensing image plays a huge role in geological exploration, military applications, vegetation detection, ocean remote sensing and other fields with its advantages.But because of the limited spatial resolution of the imaging spectrometer and the earth on the surface of the complex diversity,images within a pixel often contains a variety of different terrain types,Which formed the mixed pixels.And will this pixel to any type of terrain can cause errors,and then can not achieve high precision of remote sensing application.Endmember in hyperspectral data can detailed say a spectral properties of pure pixel under test.The obtained endmember vectors are usually as a prior knowledge of hyperspectral image processing algorithm.So whether the obtained endmember vectors can reveal to study features of spectral properties information is an important premise for further analyzing the hyperspectral data.Among various kinds of endmember extraction algorithm,primary research is N-FINDR algorithm with a good effect of endmember extraction.However the the sample order for the algorithm will cause certain influence on endmember extraction and traditional algorithms of N-FINDR make the performance of the algorithm itself is reduced greatly because of the randomness of the selected initial endmember. At the same time,the algorithm needs a dimension reduction according to the number of endmember during initialization and to traverse all the pixels to replace from time to time causes a large amount of calculation,which will limit the application of the algorithm.The article intents to make better based on the N-FINDR algorithm where was improved according to the correlation between endmember.Firstly,endmember extracted by using the improved N-FINDR algorithm conduct correlation analysis with all pixels in the original pixel geometry,and all pixels which correlation coefficient is greater than a certain threshold conduct Particle Swarm Optimization.Then the pixels closed to the real endmember are deemed to the requested endmember.Secondly,on the measure of PSO,the single volume and endmember variation is to as optimization goal for consideration in order to prevent selecting noise point as the endmember,and to further improve the accuracy of the endmember extraction.Thirdly,after the N-FINDR algorithm,searching all pixels which correlations with each endmember is located within a certain threshold range to ensure to extract the necessary endmember. Finally,create mix-pixel image to verify the improved N-FINDR.Use the spectral curves of real water and gasoline to produce the mixed image of the hyperspectral, and then determining the proportion of each endmember by spectral decomposition with the improved N-FINDR.The results of simulated data and actual test data verified the algorithm proposed by this paper is effective.