| Re: Maximum likelihood question (ENVI) [message #37579] |
Thu, 08 January 2004 23:47 |
Mort Canty
Messages: 134
Registered: March 2003
|
Senior Member |
|
|
Jonathen Greenberg writes:
"Jonathan Greenberg" <greenberg@ucdavis.edu> schrieb im Newsbeitrag
news:BC22E9A9.18624%greenberg@ucdavis.edu...
> Mort:
>
> Did RSI send you ANYTHING? Could you post them to the newsgroup?
Also,
> is there a program to do the Richard's method, or did you code it directly
> from a publication? Thanks!
>
> --j
>
Jonathan:
Right, after some searching in the dead letter department I found it:
--------------------------------------------------
Dear Dr. Canty,
Fist of all our apologies for the delay.
From Research System we now received the requested background information on
rule images regarding the Maximum Likelihood supervised classification.
It is important to keep in mind that the "probabilities" calculated by ENVI
3.5 and prior versions are not true probabilities. The values that ENVI
reports as probabilities are normalized discriminant function values.
Therefore, the probabilities reported by ENVI do not behave as true
probabilities would. For example, if for some pixel the discriminant
function values are all small (i.e., the pixel is not similar to any of the
classes), then the probabilities reported by ENVI will still sum to one. If
true probabilities were calculated in this case, then the probability for
each class would be small, and the sum of probabilities would be
significantly smaller than one.
Therefore the ENVI developers had received a request to change ENVI's
Maximum Likelihood classification to use the non-normalized gi(x) values in
the rule images.
This request had been considered for ENVI 3.6.
Now the rule images, one per class, contain a maximum likelihood
discriminant function with a modified Chi
Squared probability distribution. Higher rule image values indicate higher
probabilities. The final classification allocates each pixel to the class
with the highest probability. The negative values should indicate
unclassified pixels that do not fit the parameters established in the
initial classification.
To convert between the rule image's data space and probability, use the Rule
Classifier. For the classification threshold, enter the probability
threshold used in the Maximum Likelihood classification as a percentage
(e.g., 95%). The Rule Classifier will automatically find the corresponding
rule image Chi Squared value.
It is even possible to find the rule image Chi Squared value that equals the
probability threshold, entered as a percentage. Please follow these steps:
1. Display a rule image
2. Select Enhance->Interactive Stretching...
3. In the Interactive Stretching Dialog box, select Histogram_Source->Band
4. You can now move the cursor in the Input Histogram to determine
Chi-Squared value (or DN) which match the cumulative probability in
percentage. The output information is on the lower left corner.
Hope this is helpful!
-----------------------------------------
It wasn't. I want, like you, to access the class membership probabilities
programmatically. I don't know if there exists any software for PLR. I coded
it in IDL from Richards' description in his book "Remote Sensing Digital
Image Analysis".
Mort
|
|
|
|
| Re: Maximum likelihood question (ENVI) [message #37596 is a reply to message #37579] |
Thu, 08 January 2004 11:04  |
Jonathan Greenberg
Messages: 91
Registered: November 2002
|
Member |
|
|
Mort:
Did RSI send you ANYTHING? Could you post them to the newsgroup? Also,
is there a program to do the Richard's method, or did you code it directly
from a publication? Thanks!
--j
On 1/8/04 7:48 AM, in article btju5i$cm0j$1@zam602.zam.kfa-juelich.de, "Mort
Canty" <m.canty@fz-juelich.de> wrote:
> Hi Jonathan,
>
> I posted the self-same question to this newsgroup shortly after ENVI 3.6
> appeared and after not having received a satifactory answer from RSI.
> Unfortunately it got no takers here. Maybe someone will take pity now that
> there are two of us.
>
> Cheers
>
> Mort Canty
>
> PS I use the probability vectors for Richards' post-calssification
> "Probabilistic Label Relaxation" method.
>
> "Jonathan Greenberg" <greenberg@ucdavis.edu> schrieb im Newsbeitrag
> news:gC4Lb.7043$ux1.4730@newssvr27.news.prodigy.com...
>> I'm trying to generate images of true probabilities from maximum
> likelihood
>> rule images. Since the final probability rule images were modified (from
> the
>> RSI website): "In the ENVI 3.6 implementation, the rule images (one per
>> class) contain a maximum likelihood discriminant function with a modified
>> Chi Squared probability distribution.", I am unclear as to how to get BACK
>> to true probabilities. Typing in the % probability into the rule
> classifier
>> doesn't really get me what I need. Could I get the "true" probability for
> a
>> given class (let's say we have two classes, A and B) and calculating:
>>
>> Rule image A/(rule image A + rule image B) ?
>>
>> Or is there some other technique I need to perform to get a true
> probability
>> image? Thanks!
>>
>> I'm planning on using this in combination with Bayesian logic tools, which
>> is why I need, for each pixel and for each class the posterior probability
>> (0 to 1). As I mentioned, the rule classifier is only good if I want to
>> perform a new classification based on the rule images, not if I want to
> see
>> directly what these probabilities are.
>>
>>
>>
>
>
|
|
|
|
| Re: Maximum likelihood question (ENVI) [message #37602 is a reply to message #37596] |
Thu, 08 January 2004 07:48 |
Mort Canty
Messages: 134
Registered: March 2003
|
Senior Member |
|
|
Hi Jonathan,
I posted the self-same question to this newsgroup shortly after ENVI 3.6
appeared and after not having received a satifactory answer from RSI.
Unfortunately it got no takers here. Maybe someone will take pity now that
there are two of us.
Cheers
Mort Canty
PS I use the probability vectors for Richards' post-calssification
"Probabilistic Label Relaxation" method.
"Jonathan Greenberg" <greenberg@ucdavis.edu> schrieb im Newsbeitrag
news:gC4Lb.7043$ux1.4730@newssvr27.news.prodigy.com...
> I'm trying to generate images of true probabilities from maximum
likelihood
> rule images. Since the final probability rule images were modified (from
the
> RSI website): "In the ENVI 3.6 implementation, the rule images (one per
> class) contain a maximum likelihood discriminant function with a modified
> Chi Squared probability distribution.", I am unclear as to how to get BACK
> to true probabilities. Typing in the % probability into the rule
classifier
> doesn't really get me what I need. Could I get the "true" probability for
a
> given class (let's say we have two classes, A and B) and calculating:
>
> Rule image A/(rule image A + rule image B) ?
>
> Or is there some other technique I need to perform to get a true
probability
> image? Thanks!
>
> I'm planning on using this in combination with Bayesian logic tools, which
> is why I need, for each pixel and for each class the posterior probability
> (0 to 1). As I mentioned, the rule classifier is only good if I want to
> perform a new classification based on the rule images, not if I want to
see
> directly what these probabilities are.
>
>
>
|
|
|
|
comp.lang.idl-pvwave archive
Members
Search
Help
Login
Home
