AI-GEOSTATS
Hello Ashton,
I would suggest to do the interpolation of the probability field using a
logistic scale, to avoid getting negative estimates for some categories.
This essentially implies changing your bare zeroes (in the error matrix
below) by a suitable small number (say, the smallest probability/10 or
so), and take a multivariate logistic transform... details in my thesis
(www.tdx.cesca.es/TDX-0123106-122444/index_an.html, or just ask)... we
are working in a paper, but it is still on the air...
Regards
Raimon
En/na Ashton Shortridge ha escrit:
AI-GEOSTATS
Hello all,
I have a land cover dataset with codes 1-7 representing different land cover
categories. This data is not too good, but might be better than nothing.
Let's call this the map.
I have a second dataset - a bunch of point locations at which land cover for
the area has been ground-truthed. This is essentially my reference data.
I can use these things to construct a confusion, or error matrix, like this:
[1,] 0.11111111 0.02222222 0.00000000 0.8666667 0.00000000 0.000 0.00000000
[2,] 0.07777778 0.18888889 0.00000000 0.7222222 0.00000000 0.000 0.01111111
[3,] 0.00000000 0.24166667 0.49166667 0.2500000 0.01666667 0.000 0.00000000
[4,] 0.03333333 0.26666667 0.06666667 0.6333333 0.00000000 0.000 0.00000000
[5,] 0.00000000 0.75000000 0.00000000 0.1250000 0.00000000 0.125 0.00000000
[6,] 0.00000000 0.00000000 0.90000000 0.0000000 0.10000000 0.000 0.00000000
[7,] 0.00000000 0.00000000 0.00000000 0.0000000 0.00000000 0.000 1.00000000
where cell i,j corresponds to the observed probability of observing class j on
the ground, where class i was present in the map. For example, a cell with
class 3 on the map is actually class 3 about 49% of the time. About 24% of
the time it's class 2, and 25% of the time it is class 4. Very rarely (1.7%)
it's actually class 5.
I would like to employ indicator simulation on this data using simple kriging
with locally varying means. I want to generate realizations of reference land
cover, using the map landcover data to improve the prediction by serving as
the mean estimate. This approach is documented in Goovaerts' book and in a
paper by Kyriakidis and Dungan (2001). However, several points are unclear
to me.
First, simple kriging is employed on residuals from the mean. For
multicategorical data of the sort I am investigating here, how would one
calculate the mean at a particular location?
Second and more practically, I've struggled to discover how to implement this
in gstat (R version or standalone), and am wondering if anyone has had
success with another software package.
Thanks in advance for any assistance you can provide.
Ashton
+ To post a message to the list, send it to [email protected]
+ To unsubscribe, send email to majordomo@ jrc.it with no subject and "unsubscribe
ai-geostats" in the message body. DO NOT SEND Subscribe/Unsubscribe requests to the
list
+ As a general service to list users, please remember to post a summary of any
useful responses to your questions.
+ Support to the forum can be found at http://www.ai-geostats.org/
