END MEMBER EXTRACTING METHOD FOR HYPERSPECTRAL IMAGES, CAPABLE OF EFFECTIVELY EXTRACTING AN END MEMBER FOR A SPECTRAL MIXTURE ANALYSIS OF THE HYPERSPECTRAL IMAGES

摘要

PURPOSE: An end member extracting method for hyperspectral images is provided to perform a spectral mixture analysis per each repeated step while repeatedly increasing the number of end members, thereby generating an occupation proportion graph based on the number of end members in which the sum of the errors of the error images becomes minimized.;CONSTITUTION: An end member extracting method for hyperspectral images is as follows. The data for the hyperspectral images is compressed, and the initial number of the end members is set. The initial end members calculate a volume of a group based on an initial value of an end member set. Elements of the end member set are substituted one by one with respect to all the pixels of the images so that the volume of the group is calculated. If the volume is increased, the corresponding end member element is extracted as a spectral characteristic value of the corresponding pixel. An error image is obtained by applying a linear spectral mixture analysis to the extracted end member. The sum of errors targeting the entire pixels is obtained. The end member search step and the error image analysis steps are repeated while the number of the end members is increased one by one. If the sum of the errors is increased, the steps for repeating are stopped and the end member of the prior process is outputted as a last result.;COPYRIGHT KIPO 2013;[Reference numerals] (AA) Start; (BB) Step 1: Preprocessing; (CC) MNF or PCA transform; (DD) Step 2: Initialization; (EE) Select p pixels randomly as initial end member set{e_1, e_2, ..., e_p} from the p-1 component of transformed data; (FF) Calculate V_max, the maximum volume of simplex formed by the vector elements (e_1, e_2, ..., e_p); (GG) Step 3: Find end members; (HH) For every pixel r, calculate the volume of simplex, V_1 by {r, e_2,...e_p}, V_2 by {e_1, e_2,...e_p},..., V_p by {e_1, e_2, ..., r}); (II) If V_k for k=1;p V_max, set V_max = V_k and let new end member set as {e_1, ..., e_k_1, r, e_k+1, ..., e_p}; (JJ) Repeat step 3; (KK) Step 4: Analyze error image; (LL) Calculate sum of the error image (E_p); (MM) Step 5: Iterate; (NN) Add randomly selected one pixel e_r to the lastly extracted end member set from the first p component of transformed data. {e_1, ..., e_k_1, r, e_k+1, ..., e_p}; (OO) End