IEEE Geoscience and Remote Sensing Magazine - September 2015 - 10

Change indexes are the
main input to unsupervised change detection
procedures that generate
X1
maps without the use of
Ancillary
ground reference infort1 Image
Data
mation. Figure 3 summarizes the basic processing
X2
chain for multitemporal
1
Data
Data
Data
information fusion and
t2 Image
Collection
Pre-Processing
Analysis
extraction performed at
N
feature level. Due to their
Change Detection
unsupervised nature this
Map
kind of approaches are
widely employed since
XN
at an operational level
ground reference infortN Image
mation is often not available (e.g., the user is interested to investigate a
FiGuRE 2. Overall block scheme of multitemporal information fusion.
phenomenon occurred
in the past for which no
information
was
collected),
costly
(i.e., it requires in situ
methods presented in the literature. However, alternative
surveys by experts with proper equipment) or imposcategorizations could be considered.
sible (i.e., ground truth is required over a very large or
Methods in the tree should be implemented taking into
arduous area) to be collected. Unsupervised change deaccount the characteristics of the considered kind of data.
tection approaches mainly distinguish between changed
However in general the overall block scheme of multitemand unchanged pixels. Some techniques allow identifyporal fusion is the one shown in Figure 2.
ing different kinds of changes as well. However they do
not give any explicit label to land-cover transitions. In the
2. Fusion/inTEGRATion AT FEATuRE LEvEL:
final map each pixel is associated with one among the
MuLTiTEMpoRAL iMAGE CoMpARison
following classes: no-change (~ n) or change (X c). In the
The aim of fusion at feature level of multitemporal images is to generate new features that highlight multitemcase that land-cover transitions can be distinguished, the
poral information. These features are often referred to
latter class can be further detailed in K kinds of change
as change indices since they are employed to highlight
as X c ={~ c1, ~ c2, ..., ~ cK} [1]. Fusion at feature level comchanges occurred in bi-temporal image pairs (X 1 and X 2
monly assumes that bi-temporal images are accurately
pre-processed in order to mitigate differences that do not
acquired over the same area at different times t 1 and t 2).
depend on real changes occurred on the ground. In other
words pre-processing aims at making multitemporal images as similar as possible to each other. Pre-processing
Remote Sensing Remote Sensing
usually includes radiometric corrections (relative or abImage (Date t1)
Image (Date t2)
solute), geometric corrections (co-registration, ortho-rectification, geo-referencing), de-noising, etc. These steps
Pre-Processing
Pre-Processing
have to be conducted in different ways depending on the
kind of considered data (either active SAR or passive optical). A literature survey on the pre-processing phase is out
Feature Level Fusion
of the aim of this paper however a large amount of material can be found in the literature.
Analysis of the Change
Several mathematical operators can be applied to
Index
extract a change index. The choice of the specific mathematical operator gives rise to different kinds of techniques [1],[3]-[7]. Change indices by bi-temporal image
Change-Detection
Map (M)
fusion highlight information associated with changes in
the spectral signature or the backscattering coefficient depending on whether optical or SAR images are considered,
FiGuRE 3. Block scheme of a standard change detection approach
respectively. Given the technological differences in the
based on fusion at feature level.
10

ieee Geoscience and remote sensing magazine

september 2015

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - September 2015