In: Wagner W., Szekely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B
2>-
[w4t*] (/) =——
[mUF'"-^)
N,
if x'ij 6 T‘
if x'gT'
(12)
The corresponding log-likelihood is given by:
where j3> 0 tunes the influence of the context and S is the
Kronecker delta function defined as:
d[(Cl,Oj),(C' gh ,C 2 gh j\
|i if (a,a)=(c'.,o
|o if (a,a)^(c^,o
(18)
in C(d' = m)=faj ■ ln[p(x'j I alt)]
¡j=i
3.3 Iterative Labelling
(13)
Solving Eq. (2) is equivalent to maximize the log-likelihood
\rvL(X\X 2 \C ,C 2 ), which can be written as:
Even in this case we assume that convergence is reached if the
relative increase in the log-likelihood is lower than e .
3.1.3 Estimation of p(x\j\aLk)
It is worth noting that Eq. (3) can be re-written as:
K K K
P(ajU)-^MLkk-(/i(x'o)=^Mi-</)k(x'ij)-P(cdn t )-^Wma-<pk(x'j) (14)
\nC(X\X 2 \C x ,C 2 )=-YJU^(X',C)-U^(C',C 2 )-\n(Z) (19)
where ULxa(X',C) = -\n[p(X‘\C)], represents the class-
conditional energy function at date /=1,2, while is
given by (17). Since Z solely depends on the selected type of
neighbourhood, the final problem becomes solving:
Hence, \/k=\,...,K it holds P(ai^)-wUk = wi-P(tu'„,)-u4<*. [C ,C 2 )^^gmm{U^(C\C 2 )+Ulx,(X\C)+Ul,(X 2 ,C 2 )} (20)
Since P(coin)+P(oink)=\, we have: c ’. c2
, Wk-P(Oim)-Wmxk Wk-P(COLx)-wLk
Wink k — =
P(ccU) 1 -P(oim)
(15)
Then, as M', £', W' and WL have been determined, we can
compute p(x'j\coU) substituting (15) into (6), upon it is possi
ble to obtain a reliable estimate P(oinx) for the prior probability
of the class of interest. This can be properly accomplished
throughout the iterative labelling phase presented below.
3.2 Joint Prior Modelling
For modelling the joint prior P(C\C 2 ), we propose an ap
proach based on MRF (Solberg et al., 1996). In particular, we
assume that the couple of labels Cl, Cl associated with pixel
at position (i,j) at times /1 and /2 depends on the couples of
labels associated with pixels belonging to the spatial
neighbourhood Go of (i,j) at the two dates (we always con
sidered first order neighbourhoods).
In other words, the higher the number of its spatial neighbours
experiencing a certain land-cover transition is, the higher the
probability for a given pixel of experiencing the same transition
is. In this hypothesis, according with the MRF theory, it holds
the equivalence:
P(C'j,Cy \ C' g h,dh;(g,h)e 0 i ,) = Z- 1 -exp[-[/ № (a,C^)] (16)
where Z = Z(Qij) is a normalizing constant called partition
function, while C/cmuex is a Gibbs energy function (accounting
for the spatio-temporal context) of the form:
To this aim, we propose a strategy based on the Iterated Condi
tional Modes (ICM) algorithm (Besag, 1996) which allows
maximizing local conditional probabilities sequentially. In par
ticular, at each iteration / we update the estimated prior prob
abilities for the class of interest at each date P(cdnx) and, ac
cordingly, also the class-conditional densities of the unknown
classes p(x\j \ coU). The algorithm works as follows:
Step 1. After estimating ^(x!> \cii M ), VxiyeA", / = 1,2 follow
ing the approach described in the previous paragraphs, set
P(oixxx)=0.5 (no prior knowledge is assumed to be available
about the true prior P(coLx)) and compute the conditional den
sity of the unknown classes p(x'ij\ccL ! .) accordingly;
Step 2. Derive the initial sets of labels C, C 2 by solely mini
mizing the non-contextual terms of Eq. (20), i.e.
{C',C 2 } = argmm{UL(X',C')+UL(X 2 ,C 2 )};
Step 3. On the basis of current C' and C 2 , compute the new
estimated prior probabilities ratioing the number of pixels asso
ciated with the class of interest over the whole number of pix
els, i.e. P(oinx) = \ax\/(I-J), CL = {G i \Gj£C , ,Gj=aL} I i f\
then, update the class-conditional densities for the unknown
classes p(x'j \ccL±) accordingly;
Step 4. Update C ] and C 2 according with Eq. (20);
Step 5. Repeat Step 3 and Step 4 until no changes occur be
tween successive iterations.
At the end of the process, the final targeted change-detection
map C' is defined as:
c'={c;j\’:U,
if Cl = cdnx and Cl = cdxx
otherwise
(21)
4. EXPERIMENTAL RESULTS
U^(Cl,Cl)=- Y, p-S[(Cl,C 2 j),(C gh ,Cl h j\
(g^Sn
(17) In order to assess the effectiveness of the proposed technique,
we carried out several experiments with different combinations