[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
basic indicator simulation questions
I have what are probably some fairly basic indicator simulation questions,
but I've checked the list archive as well as Dr. Pebesma's articles and
Dr. Cressie's book on geostatistics and some questions remain- no doubt
due more to my lacking in geostatistical knowledge than anything else.
I have presence/absence data of vegetation alliances in the Mojave Desert
and want to generate vegetation models based on environmental variables
(e.g., elevation, climate, landform...) using generalized linear models
and classification trees. As these models will be used more for prediction
than explanation, I am interested in incorporating spatial dependence as
an additional explanatory variable (even if it just represents some
unmodeled, unobserved spatially dependent environmental variable rather
than an actual biological process resulting in spatial clustering of the
vegetation). Using sample data (coded 1 and 0), I want to generate a
surface of alliance presence using simulation and estimation (indicator
kriging), add this variable to the modeling process, and compare the
I started using GSLIB for this, but as a non-computer programmer I found
it a little unwieldy, and also had problems creating a surface from the
output. I've started using GSTAT through the new version of IDRISI and
have been very pleased with it so far but had a few questions about some
of the results I've been getting:
1. My understanding of conditional simulation is that it uses input sample
data and the resulting simulations follow the mean, variogram and pdf of
the sample data. When I used GSLIB, one of the parameters the user gives
is the global pdf values and I found that the resulting simulation had
this exact proportion of value 1 to value 0; however with GSTAT, I've
found that the resulting proportions have quite a bit of range.
Following a post in the GSTAT archives, Dr. Pebesma said that the simple
kriging mean is approximately equal to the sample mean value. My data has
a sample value of 0.305 and I use that for the simple kriging mean but get
simulations with proportions ranging from 0.24 to 0.37.
I saw in one of the user manuals that pdf violations could be attributed
to insufficient data across all the thresholds and that's possibly my
problem, but I wanted to check and see if I was doing something wrong.
2. Related to the conditional data, my sample data are not fully
"honored". About 10% of the sample points are on a grid cell with the
incorrect value (e.g., a "1" on a "0" grid, or a "0" on a "1" grid).
Also in an archived post I saw that this could be attributed to a nugget
effect in the variogram (which I have: 0.06 NUG(0)), but wanted to see if
something else might be wrong as well.
Sorry for the length of this post and thanks very much for your help.
J e n n i f e r M i l l e r
San Diego State University
San Diego, CA 92182-4493