第十四讲地震波层析成像课堂.pptVIP

. 31 Menke, 1989 . 32 . 33 Least-Squares Solutions Suppose we have a simple set of linear equations T T ( ) ( ) E ? ? ? ? ? ? A X dA X d ? ( AX ? d ) 2 A X = d We can define a simple scalar quantity E Mean square error (or total error) Error function We want to minimize the total error, to do so, find first derivative of function E and set to 0. So, do ? E ? X ? 0 , we should have T 2 ( ) 0 ? ? A A X d A T AX ? A T d This is known as the system of normal equations. X ? ( A T A ) ? 1 A T d So this involves the inversion of the term A T A, This matrix is often called the inner-product matrix, or Toeplitz matrix . The solution is called the least- squares solution, while X=A -1 d is not a least squares solution . 最小二乘法适合于观测数据个数多于未知数个数。 . 34 Left multiply by ( A T A ? ? I ) ? 1 Pre-conditioning for ill-conditioned inverse problem (damping, smoothing, regularization. Purpose: Stabilize, enhance smoothness/simplicity ) Lets use the same definition E ? ? T ? ? AX ? d 2 Define: An objective function J where J ? E ? ? X 2 where m is the damping or regularization parameter. ? J ? X ? ? E ? X ? ? ( ? X 2 ) ? X T 2 ( ) 2 ? ? ? ? A A X d I X I is identity matrix Minimize the above by ? J ? X ? 0 ( A T A ? ? I ) X ? A T d X ? ( A T A ? ? I ) ? 1 A T d (Damped Least Squares solution) 阻尼最小二乘法，适合于观测数据个数少于未知数个数。 . 35 Solve for the inverse problem (1) Standard Least Squares Solution G T G may be singular or ill-conditioned ? singular value decomposition (SVD) (2) Damped Least Squares Solution minimize Ray-based traveltime tomography T 2 1 ( ) ? ? ? ? T mG G I G d . 36 minimize L: Laplacian operator Smooth model Solution: m = (G T G+λ 2 L T L) -1 G T d 3. Smooth model . 37 Combined norm and Laplacian regularization misfit function Crust correction Data misfit Model roughness Model norm . 38 在 Zhao 等 ( 1992, 1994) 方法基础上 , 双差走时定位 ( double difference location; Waldhauser and Ellsw orth, 2000; Wolfe, 2002) 方法被引入到地震波层析成像反演 中 , 即 地震波双差走时层析成像方法 ( Zhang and