IEEE Geoscience and Remote Sensing Magazine - June 2019 - 144

scheme shown in Figure 4. The accurate preprocessing steps,
such as uncalibrated and noisy band removal, bad band
stripe repair, and radiometric and atmospheric corrections,
among others, are usually necessary for enhancing the original spectral representation of the input HS images. In particular, an accurate coregistration process in HS images with
a limited residual misregistration error significantly affects
the performance of CD algorithms. Multitemporal analysis
is conducted to represent and
discover the temporal changes between the HS image pairs
IMAGE STACKING
by considering both the specCOMBINES THE
tral signatures in each single
MULTITEMPORAL IMAGES
time image and the temporal correlation between the
BY EXTENDING THE DATA
multitemporal images. Three
DIMENSIONALITY TOWARD
main types of multitemporal
THE TEMPORAL DIRECTION.
strategies can be used to process HS images relying on 1)
an image-comparison operator, 2) an image-stacking approach, and 3) an independent
image analysis. Afterward, specific CD techniques can be
designed based on the change representation implemented
in the previous step.
Described in greater detail, comparison operators model
temporal variations in multitemporal images according to
different techniques, such as univariate image differencing
[26], change vector analysis (CVA) [27], index-based differencing [19], distance or similarity measures [16], and
so on. A summary of the most widely used comparison operators can be found in [1]. To effectively use these comparison operators in HS images, it is crucial issue to select the
most informative bands [28] or construct specific indexes
based on the relevant bands (e.g., the HS vegetation indices
[29]). Let I 1 and I 2 be a pair of B-dimensional HS images

HS Image of Date t1

HS Image of Date t2

Preprocessing

Preprocessing
Multitemporal Analysis

Image
Comparison

Image
Stacking

Independent
Image Analysis

CD Techniques
CD Map
FIGURE 4. A block scheme of the CD approach in bitemporal HS

images.
144

acquired at times t 1 and t 2, respectively. The corresponding
pixels in two images represent the spectral signatures of the
same area on the ground at the two dates. The difference
image I D can be obtained by comparing I 1 and I 2 with the
vector differencing comparison operator, thus defining the
CVA [16], [27], [30]-[32],
I D = I 2 - I 1,

(1)

where each pixel in I D corresponds to a B-dimensional
spectral change vector (SCV).
Image stacking combines the multitemporal images by
extending the data dimensionality toward the temporal direction. Let I s be the stacked multitemporal images from
two dates I 1 and I 2:
I s =[I 1, I 2],

(2)

where the dimensionality of I s is 2B. Then, proper techniques should be used in the full extended data dimensionality to extract the specific information associated
with the changes.
Independent image analysis classifies each single time
image using supervised classifiers and then compares the
classification maps to discover changes. This is the most
traditional CD method (also termed postclassification comparison [PCC]) [26] in both MS and HS images. It relies
mainly on the availability of comprehensive ground reference data for generating the training samples as well as
the selection of an effective classifier. However, collecting
ground reference data is always an expensive and timeconsuming process, and it becomes even more demanding when the ground references must be collected for multiple acquisition dates and for detailed land-cover classes.
Therefore, the availability of multitemporal ground reference data is typically limited. Moreover, PCC suffers from
the propagation of classification errors between two dates,
which may significantly affect the CD accuracy. To deal
with this issue, multitemporal classification techniques
can be used (e.g., compound detection [33], [34]). However, these approaches require some multitemporal reference data. For these reasons, most articles in the literature
address the CD problem by considering comparison operators (e.g., [28], [35]-[37]) and image-stacking operations
(e.g., [38], [39]) (Figure 5).
Let X ={~ n, X C} be the set of all classes in the considered multitemporal images divided into the unchanged and
changed classes, that is, ~ n and X C, respectively. This definition represents the general case of ACD and binary CD,
which aims to separate the changed targets (foreground)
from the unchanged objects (background). For multiclass
CD, the aim is not only to detect the general change class
X C but also to distinguish different classes. Therefore, X C
can be divided into K different kinds of change, that is,
X C ={~ 1, ~ 2, f, ~ K }. This highlights how the CD problem
can be divided into multiple subproblems, among which
the most important ones are 1) discrimination between
unchanged ~ n and changed samples X C, 2) identification
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE

JUNE 2019
IEEE Geoscience and Remote Sensing Magazine - June 2019

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - June 2019
