Re: applying theory of evidence by Dempster-Shafer in remote sensing image classification

On Mar 7, 8:28 pm, Jonathan Campbell <jg.campbell...@xxxxxxxxx> wrote:
Guessing ---I do not have the paper, and I haven't bothered to revise my
D-S.

Image channel q = band q (confirmed somewhere else in your post).

If we have the values out of band q as xq E {0, 1, ... Lq -1}; (for
8-bit data Lq = 255).

You can estimate the conditional density pq(X = x, C = c) using the
histograms mentioned; c = class E {1, 2, ..., Nc}; Nc land-cover classes
in total.

You can use Bayes' rule to invert the conditionality; but you need to
assign a prior to each class, P(c).

    p^q(C = c, Xq = xq) =
     P(c) . pq(X = x, C = c)/ sum all c P(c) . pq(X = x, C = c)

The denominator (thing on the bottom) is just to normalise the probability..

Then, I guess, you combine the p^qs over each band q.

Of course, if you could estimate a proper multivariate density p(X = x,
C = c), etc. (X now a vector) and use just Bayes', you'd get a better
result.

Best regards,

Jon C.


All right. Thanks. A little bit clearer now.

Larry
.

Loading