第四章 主分量分析.ppt

EOF局限性 与区域范围选择有关，这给进行物理解释带来困难 过分强调与分析区域尺度相当的现象，因而使区域性相关结构反映不出来。 4.4 旋转EOF（REOF） REOF 基本思想 如果EOF分析截取了前K个空间型，累计解释场的总方差已达一定要求，例如累计方差贡献率达80%，可否（及怎样）将这K个空间型再做调整，使得调整后的K个空间型累计解释原场总方差的百分率保持不变，而单个空间型尽量反映场的局部相关结构。 这种对K个EOF再做调整的过程在数学上是在它们之间再做线性变换，几何形象是做旋转，所以称为REOF。 EOF分解 分量形式 % [eof,V] = varimax(eof[,sign,tol]) Varimax rotation % % Performs varimax rotation (Kaiser 1958) on the column vectors contained % in eof. We follow Kaisers notation. % % In: % eof: DxL matrix (LD) of column vectors (possibly the loadings from % a factor analysis). % sign: preferred majoritary sign for the vector components (-1: % negative, 1: positive, 0: dont care). Default 0. % tol: minimum relative increase of the objective function to keep on % iterating (default 1e-5). % Out: % eof: DxL matrix with the rotated loadings. % V: training curve for the objective function. % Copyright (c) 1997 by Miguel A. Carreira-Perpinan function [eof,V] = varimax(eof,sign,tol) % Argument defaults if nargin==1 sign=0; end; if nargin=2 tol=1e-5; end; [D,L] = size(eof); h = sqrt(sum(eof.^2))+exp(-700); % Communalities temp=eof./(h*ones(1,L)); V = [sum(sum(temp.^4))-sum(sum(temp.^2).^2)/D]; % Objective function V_old = V*(1-2*tol); while abs(V(length(V))-V_old) tol*V(length(V)) V_old = V(length(V)); for i=1:L-1 for j=i+1:L % Optimal angle to rotate columns i, j x = eof(:,i)./h; y = eof(:,j)./h; u = x.*x - y.*y; v = 2*x.*y; t = atan2( 2*(D*u*v-sum(u)*sum(v)), D*(u*u-v*v)-sum(u)^2+sum(v)^2 )/4; % Anticlockwise rotation of angle t (t+pi is valid, too) temp = [eof(:,i) eof(:,j)]*[cos(t) -sin(t); sin(t) cos(t)]; eof(:,i) = temp(:,1); eof(:,j) = temp(:,2); end end % New value of the objective function h = sqrt(sum(eof.^2))+exp(-700); % Communalities temp=eof./(h*ones(1,L)); V = [V sum(sum(temp.^4))-sum(sum(temp.^2).^2)/D]; % Objective function end % Sign inversion so that each column vector of eof has