R INLA spatial prediction VS universal kriging with geoR

[wenlo...@gmail.com]

Hi,

I did a comparison between R-inla prediction and kriging function in geoR package, found that the R-inla prediction tends to be  not as good as the geoR kriging results in terms of rmse and r2, even with massive mesh locs. And I wonder if I miss something when coding up the inla spatial prediction as below. Any help will be appreciated!

#inla setting snapshot 

 mesh <- inla.mesh.2d(loc,

                     max.edge = c(1, 2),

                     offset = c(0.5, 1.5),

                      cutoff =0.5 )

 A <- inla.spde.make.A(mesh, loc) 

 ...

 result_uni <- inla(form_uni, rep('gaussian', 1),

                       data = inla.stack.data(stack_uni), 

                       control.family = list(hyper.eps),

                       control.predictor = list(A = inla.stack.A(stack_uni), compute=TRUE),

                       control.inla = list(int.strategy = 'eb'),

                       control.compute=list(cpo=TRUE, dic=TRUE,config = TRUE))

...

#pred_mean at mesh locs

  pred_mean <- result_uni$summary.random[[s1]]$mean + result_uni$summary.fixed$mean[1]

#pred_sd at mesh locs

  pred_sd <- result_uni$summary.random[[s1]]$sd

  

#get the prediction at obs locs

  tmp_m <- drop(A %*% pred_mean)

  tmp_sd <- drop(A %*% pred_mean)