Nick Hamm
Thu, 11 Sep 2003 07:02:54 -0700
On Fri, 5 Sep 2003, Paulo Justiniano Ribeiro Jr wrote: > Edzer Pebesma wrote: > > > As a short-term answer: you can obtain block mean values > > by point kriging on a fine grid, and then averaging point kriged > > values within a block. It does not give you block kriging > > standard errors, though. > > -- > > I think we can: > If we generate simulations on the fine grid, for every simulation compute > the mean at specified block (block kriging estimate) and then the variance > of those means you will have the block kriging std. errors This doesn't seem to work..... For another data set (no anisotropy) I have defined a 5 by 5 block and discretized it on a fine grid (400 points). Using gstat I use SGS to simulate 1000 realisations of each point. for each realisation I compute the mean of the 400 points (to give the value of the simulation on the block). I then have 1000 simulations of the mean for the block -- I comupte the meand and variance of these and get: > "Mean: " 8.748847 > "Var : " 0.000704 If I do block kriging straight off I get: > "Mean: " 8.746475 (ie block kriging prediction) > "Var: " 0.001332 (ie block kriging variance) If I do 1000 simulations of the block (ie SGS for the block) I get: > "Mean: " 8.748569 > "Var: " 0.001390 So the mean and variance from the simulations over the block seems to be basically the same as the block kriging prediction and variance. Using Paulo's approach the Mean looks ok, but the Variance seems to be underestimated...... I've checked my code and it looks ok... Nick