Upload
lamkhuong
View
212
Download
0
Embed Size (px)
Citation preview
Kriging with large data sets using sparse matrix techniques
Kriging
• One method of obtaining an es9mated map of the variable of this map in different parts of the region is kriging.
Assump9on for ordinary kriging
• (1) Z(s) is a random func9on • (2) E(Z(s))=μ for all s in the region • (3)
• The func9on 2γ(h) is called the variogram of the process. If 2γ(h) exists, the process is called intrinsically sta9onary
• If the random process is second order sta9onary; then the variogram has a sill:
• If the random process have a sill:
This gives the covariogram-‐based version of the kriging equa9on:
Advantage and disadvantage of variogram and covariogram kriging equa9on
• Time consuming computa9on of
Need only be done once.
• The 9me required to solve a dense (few nonzero entries) linear n*n system grows at order n^3, and the required memory is of order n^2
Sparse matrix techniques
• Main idea: to exclude observa9ons far away from s0.
• Lower 9me and storage costs
Variogram VS covariogram
• Variogram: (1) Less biased (2)can be defined for some process that are not second order sta9onary
• Covariogram:
• (1) Es9mate the variogram from the data
• (2)compu9ng the covariogram matrix ∑
• Then, we can retain the low bias of variogram es9ma9on along with the computa9onal advantage of sparsity in ∑.
Inves9ga9on of the rela9ve computa9on
• For each n, R, we obtain the sparse matrix Σ and full matrix Γ
Spherical variogram:
R=0.3 R=0.6 R=0.9
density 9/n 25/n 69/n
slope 1.39 1.63 1.67
Geochemical Data
The advantage of sparse techniques do not depend on the use of a regular lagce.
Bailey and Gatrell fit a spherical variogram to the log of the nickel concentra9on, obtaining the covariogram model:
The resul9ng covariance matrix is very sparse, with only 2096 nonzero elements compared with 839056 elements in full matrix. The sparse techniques are 432 9mes as fast!
Discussion
• We do not es9mate the covariogram directly. To obtain sparse matrices, variogram models with a finite range must be used. (Spherical variogram)
• For irregular data, it is possible that a poor choice of ordering could decrease the efficiency.